sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 1 | |
| 2 | /*---------------------------------------------------------------*/ |
sewardj | 752f906 | 2010-05-03 21:38:49 +0000 | [diff] [blame] | 3 | /*--- begin host_amd64_isel.c ---*/ |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 4 | /*---------------------------------------------------------------*/ |
| 5 | |
| 6 | /* |
sewardj | 752f906 | 2010-05-03 21:38:49 +0000 | [diff] [blame] | 7 | This file is part of Valgrind, a dynamic binary instrumentation |
| 8 | framework. |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 9 | |
sewardj | 89ae847 | 2013-10-18 14:12:58 +0000 | [diff] [blame^] | 10 | Copyright (C) 2004-2013 OpenWorks LLP |
sewardj | 752f906 | 2010-05-03 21:38:49 +0000 | [diff] [blame] | 11 | info@open-works.net |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 12 | |
sewardj | 752f906 | 2010-05-03 21:38:49 +0000 | [diff] [blame] | 13 | This program is free software; you can redistribute it and/or |
| 14 | modify it under the terms of the GNU General Public License as |
| 15 | published by the Free Software Foundation; either version 2 of the |
| 16 | License, or (at your option) any later version. |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 17 | |
sewardj | 752f906 | 2010-05-03 21:38:49 +0000 | [diff] [blame] | 18 | This program is distributed in the hope that it will be useful, but |
| 19 | WITHOUT ANY WARRANTY; without even the implied warranty of |
| 20 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 21 | General Public License for more details. |
| 22 | |
| 23 | You should have received a copy of the GNU General Public License |
| 24 | along with this program; if not, write to the Free Software |
| 25 | Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA |
sewardj | 7bd6ffe | 2005-08-03 16:07:36 +0000 | [diff] [blame] | 26 | 02110-1301, USA. |
| 27 | |
sewardj | 752f906 | 2010-05-03 21:38:49 +0000 | [diff] [blame] | 28 | The GNU General Public License is contained in the file COPYING. |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 29 | |
| 30 | Neither the names of the U.S. Department of Energy nor the |
| 31 | University of California nor the names of its contributors may be |
| 32 | used to endorse or promote products derived from this software |
| 33 | without prior written permission. |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 34 | */ |
| 35 | |
| 36 | #include "libvex_basictypes.h" |
| 37 | #include "libvex_ir.h" |
| 38 | #include "libvex.h" |
| 39 | |
sewardj | cef7d3e | 2009-07-02 12:21:59 +0000 | [diff] [blame] | 40 | #include "ir_match.h" |
| 41 | #include "main_util.h" |
| 42 | #include "main_globals.h" |
| 43 | #include "host_generic_regs.h" |
| 44 | #include "host_generic_simd64.h" |
sewardj | 69d98e3 | 2010-06-18 08:17:41 +0000 | [diff] [blame] | 45 | #include "host_generic_simd128.h" |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 46 | #include "host_generic_simd256.h" |
| 47 | #include "host_generic_maddf.h" |
sewardj | cef7d3e | 2009-07-02 12:21:59 +0000 | [diff] [blame] | 48 | #include "host_amd64_defs.h" |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 49 | |
| 50 | |
| 51 | /*---------------------------------------------------------*/ |
| 52 | /*--- x87/SSE control word stuff ---*/ |
| 53 | /*---------------------------------------------------------*/ |
| 54 | |
| 55 | /* Vex-generated code expects to run with the FPU set as follows: all |
| 56 | exceptions masked, round-to-nearest, precision = 53 bits. This |
| 57 | corresponds to a FPU control word value of 0x027F. |
| 58 | |
| 59 | Similarly the SSE control word (%mxcsr) should be 0x1F80. |
| 60 | |
| 61 | %fpucw and %mxcsr should have these values on entry to |
| 62 | Vex-generated code, and should those values should be |
| 63 | unchanged at exit. |
| 64 | */ |
| 65 | |
| 66 | #define DEFAULT_FPUCW 0x027F |
| 67 | |
| 68 | #define DEFAULT_MXCSR 0x1F80 |
| 69 | |
| 70 | /* debugging only, do not use */ |
| 71 | /* define DEFAULT_FPUCW 0x037F */ |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 72 | |
| 73 | |
| 74 | /*---------------------------------------------------------*/ |
| 75 | /*--- misc helpers ---*/ |
| 76 | /*---------------------------------------------------------*/ |
| 77 | |
| 78 | /* These are duplicated in guest-amd64/toIR.c */ |
| 79 | static IRExpr* unop ( IROp op, IRExpr* a ) |
| 80 | { |
| 81 | return IRExpr_Unop(op, a); |
| 82 | } |
| 83 | |
| 84 | static IRExpr* binop ( IROp op, IRExpr* a1, IRExpr* a2 ) |
| 85 | { |
| 86 | return IRExpr_Binop(op, a1, a2); |
| 87 | } |
| 88 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 89 | static IRExpr* bind ( Int binder ) |
| 90 | { |
| 91 | return IRExpr_Binder(binder); |
| 92 | } |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 93 | |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 94 | static Bool isZeroU8 ( IRExpr* e ) |
| 95 | { |
| 96 | return e->tag == Iex_Const |
| 97 | && e->Iex.Const.con->tag == Ico_U8 |
| 98 | && e->Iex.Const.con->Ico.U8 == 0; |
| 99 | } |
| 100 | |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 101 | |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 102 | /*---------------------------------------------------------*/ |
| 103 | /*--- ISelEnv ---*/ |
| 104 | /*---------------------------------------------------------*/ |
| 105 | |
| 106 | /* This carries around: |
| 107 | |
| 108 | - A mapping from IRTemp to IRType, giving the type of any IRTemp we |
| 109 | might encounter. This is computed before insn selection starts, |
| 110 | and does not change. |
| 111 | |
| 112 | - A mapping from IRTemp to HReg. This tells the insn selector |
| 113 | which virtual register is associated with each IRTemp |
| 114 | temporary. This is computed before insn selection starts, and |
| 115 | does not change. We expect this mapping to map precisely the |
| 116 | same set of IRTemps as the type mapping does. |
| 117 | |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 118 | - vregmap holds the primary register for the IRTemp. |
| 119 | - vregmapHI is only used for 128-bit integer-typed |
| 120 | IRTemps. It holds the identity of a second |
| 121 | 64-bit virtual HReg, which holds the high half |
| 122 | of the value. |
| 123 | |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 124 | - The host subarchitecture we are selecting insns for. |
| 125 | This is set at the start and does not change. |
| 126 | |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 127 | - The code array, that is, the insns selected so far. |
| 128 | |
| 129 | - A counter, for generating new virtual registers. |
| 130 | |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 131 | - A Bool for indicating whether we may generate chain-me |
| 132 | instructions for control flow transfers, or whether we must use |
| 133 | XAssisted. |
| 134 | |
| 135 | - The maximum guest address of any guest insn in this block. |
| 136 | Actually, the address of the highest-addressed byte from any insn |
| 137 | in this block. Is set at the start and does not change. This is |
| 138 | used for detecting jumps which are definitely forward-edges from |
| 139 | this block, and therefore can be made (chained) to the fast entry |
| 140 | point of the destination, thereby avoiding the destination's |
| 141 | event check. |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 142 | |
| 143 | Note, this is all host-independent. (JRS 20050201: well, kinda |
| 144 | ... not completely. Compare with ISelEnv for X86.) |
| 145 | */ |
| 146 | |
| 147 | typedef |
| 148 | struct { |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 149 | /* Constant -- are set at the start and do not change. */ |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 150 | IRTypeEnv* type_env; |
| 151 | |
| 152 | HReg* vregmap; |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 153 | HReg* vregmapHI; |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 154 | Int n_vregmap; |
| 155 | |
sewardj | 5117ce1 | 2006-01-27 21:20:15 +0000 | [diff] [blame] | 156 | UInt hwcaps; |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 157 | |
| 158 | Bool chainingAllowed; |
| 159 | Addr64 max_ga; |
| 160 | |
| 161 | /* These are modified as we go along. */ |
| 162 | HInstrArray* code; |
| 163 | Int vreg_ctr; |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 164 | } |
| 165 | ISelEnv; |
| 166 | |
| 167 | |
| 168 | static HReg lookupIRTemp ( ISelEnv* env, IRTemp tmp ) |
| 169 | { |
| 170 | vassert(tmp >= 0); |
| 171 | vassert(tmp < env->n_vregmap); |
| 172 | return env->vregmap[tmp]; |
| 173 | } |
| 174 | |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 175 | static void lookupIRTempPair ( HReg* vrHI, HReg* vrLO, |
| 176 | ISelEnv* env, IRTemp tmp ) |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 177 | { |
| 178 | vassert(tmp >= 0); |
| 179 | vassert(tmp < env->n_vregmap); |
florian | 79efdc6 | 2013-02-11 00:47:35 +0000 | [diff] [blame] | 180 | vassert(! hregIsInvalid(env->vregmapHI[tmp])); |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 181 | *vrLO = env->vregmap[tmp]; |
| 182 | *vrHI = env->vregmapHI[tmp]; |
| 183 | } |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 184 | |
| 185 | static void addInstr ( ISelEnv* env, AMD64Instr* instr ) |
| 186 | { |
| 187 | addHInstr(env->code, instr); |
| 188 | if (vex_traceflags & VEX_TRACE_VCODE) { |
sewardj | f355f6b | 2006-08-16 00:23:21 +0000 | [diff] [blame] | 189 | ppAMD64Instr(instr, True); |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 190 | vex_printf("\n"); |
| 191 | } |
| 192 | } |
| 193 | |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 194 | static HReg newVRegI ( ISelEnv* env ) |
| 195 | { |
| 196 | HReg reg = mkHReg(env->vreg_ctr, HRcInt64, True/*virtual reg*/); |
| 197 | env->vreg_ctr++; |
| 198 | return reg; |
| 199 | } |
| 200 | |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 201 | static HReg newVRegV ( ISelEnv* env ) |
| 202 | { |
| 203 | HReg reg = mkHReg(env->vreg_ctr, HRcVec128, True/*virtual reg*/); |
| 204 | env->vreg_ctr++; |
| 205 | return reg; |
| 206 | } |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 207 | |
| 208 | |
| 209 | /*---------------------------------------------------------*/ |
| 210 | /*--- ISEL: Forward declarations ---*/ |
| 211 | /*---------------------------------------------------------*/ |
| 212 | |
| 213 | /* These are organised as iselXXX and iselXXX_wrk pairs. The |
| 214 | iselXXX_wrk do the real work, but are not to be called directly. |
| 215 | For each XXX, iselXXX calls its iselXXX_wrk counterpart, then |
| 216 | checks that all returned registers are virtual. You should not |
| 217 | call the _wrk version directly. |
| 218 | */ |
| 219 | static AMD64RMI* iselIntExpr_RMI_wrk ( ISelEnv* env, IRExpr* e ); |
| 220 | static AMD64RMI* iselIntExpr_RMI ( ISelEnv* env, IRExpr* e ); |
| 221 | |
| 222 | static AMD64RI* iselIntExpr_RI_wrk ( ISelEnv* env, IRExpr* e ); |
| 223 | static AMD64RI* iselIntExpr_RI ( ISelEnv* env, IRExpr* e ); |
| 224 | |
| 225 | static AMD64RM* iselIntExpr_RM_wrk ( ISelEnv* env, IRExpr* e ); |
| 226 | static AMD64RM* iselIntExpr_RM ( ISelEnv* env, IRExpr* e ); |
| 227 | |
| 228 | static HReg iselIntExpr_R_wrk ( ISelEnv* env, IRExpr* e ); |
| 229 | static HReg iselIntExpr_R ( ISelEnv* env, IRExpr* e ); |
| 230 | |
| 231 | static AMD64AMode* iselIntExpr_AMode_wrk ( ISelEnv* env, IRExpr* e ); |
| 232 | static AMD64AMode* iselIntExpr_AMode ( ISelEnv* env, IRExpr* e ); |
| 233 | |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 234 | static void iselInt128Expr_wrk ( /*OUT*/HReg* rHi, HReg* rLo, |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 235 | ISelEnv* env, IRExpr* e ); |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 236 | static void iselInt128Expr ( /*OUT*/HReg* rHi, HReg* rLo, |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 237 | ISelEnv* env, IRExpr* e ); |
| 238 | |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 239 | static AMD64CondCode iselCondCode_wrk ( ISelEnv* env, IRExpr* e ); |
| 240 | static AMD64CondCode iselCondCode ( ISelEnv* env, IRExpr* e ); |
| 241 | |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 242 | static HReg iselDblExpr_wrk ( ISelEnv* env, IRExpr* e ); |
| 243 | static HReg iselDblExpr ( ISelEnv* env, IRExpr* e ); |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 244 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 245 | static HReg iselFltExpr_wrk ( ISelEnv* env, IRExpr* e ); |
| 246 | static HReg iselFltExpr ( ISelEnv* env, IRExpr* e ); |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 247 | |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 248 | static HReg iselVecExpr_wrk ( ISelEnv* env, IRExpr* e ); |
| 249 | static HReg iselVecExpr ( ISelEnv* env, IRExpr* e ); |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 250 | |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 251 | static void iselDVecExpr_wrk ( /*OUT*/HReg* rHi, HReg* rLo, |
| 252 | ISelEnv* env, IRExpr* e ); |
| 253 | static void iselDVecExpr ( /*OUT*/HReg* rHi, HReg* rLo, |
| 254 | ISelEnv* env, IRExpr* e ); |
| 255 | |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 256 | |
| 257 | /*---------------------------------------------------------*/ |
| 258 | /*--- ISEL: Misc helpers ---*/ |
| 259 | /*---------------------------------------------------------*/ |
| 260 | |
| 261 | static Bool sane_AMode ( AMD64AMode* am ) |
| 262 | { |
| 263 | switch (am->tag) { |
| 264 | case Aam_IR: |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 265 | return |
| 266 | toBool( hregClass(am->Aam.IR.reg) == HRcInt64 |
| 267 | && (hregIsVirtual(am->Aam.IR.reg) |
florian | 79efdc6 | 2013-02-11 00:47:35 +0000 | [diff] [blame] | 268 | || sameHReg(am->Aam.IR.reg, hregAMD64_RBP())) ); |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 269 | case Aam_IRRS: |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 270 | return |
| 271 | toBool( hregClass(am->Aam.IRRS.base) == HRcInt64 |
| 272 | && hregIsVirtual(am->Aam.IRRS.base) |
| 273 | && hregClass(am->Aam.IRRS.index) == HRcInt64 |
| 274 | && hregIsVirtual(am->Aam.IRRS.index) ); |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 275 | default: |
| 276 | vpanic("sane_AMode: unknown amd64 amode tag"); |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | |
| 281 | /* Can the lower 32 bits be signedly widened to produce the whole |
| 282 | 64-bit value? In other words, are the top 33 bits either all 0 or |
| 283 | all 1 ? */ |
| 284 | static Bool fitsIn32Bits ( ULong x ) |
| 285 | { |
| 286 | Long y0 = (Long)x; |
| 287 | Long y1 = y0; |
| 288 | y1 <<= 32; |
| 289 | y1 >>=/*s*/ 32; |
| 290 | return toBool(x == y1); |
| 291 | } |
| 292 | |
sewardj | eb17e49 | 2007-08-25 23:07:44 +0000 | [diff] [blame] | 293 | /* Is this a 64-bit zero expression? */ |
| 294 | |
| 295 | static Bool isZeroU64 ( IRExpr* e ) |
| 296 | { |
| 297 | return e->tag == Iex_Const |
| 298 | && e->Iex.Const.con->tag == Ico_U64 |
| 299 | && e->Iex.Const.con->Ico.U64 == 0ULL; |
| 300 | } |
| 301 | |
| 302 | static Bool isZeroU32 ( IRExpr* e ) |
| 303 | { |
| 304 | return e->tag == Iex_Const |
| 305 | && e->Iex.Const.con->tag == Ico_U32 |
| 306 | && e->Iex.Const.con->Ico.U32 == 0; |
| 307 | } |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 308 | |
| 309 | /* Make a int reg-reg move. */ |
| 310 | |
| 311 | static AMD64Instr* mk_iMOVsd_RR ( HReg src, HReg dst ) |
| 312 | { |
| 313 | vassert(hregClass(src) == HRcInt64); |
| 314 | vassert(hregClass(dst) == HRcInt64); |
| 315 | return AMD64Instr_Alu64R(Aalu_MOV, AMD64RMI_Reg(src), dst); |
| 316 | } |
| 317 | |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 318 | /* Make a vector (128 bit) reg-reg move. */ |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 319 | |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 320 | static AMD64Instr* mk_vMOVsd_RR ( HReg src, HReg dst ) |
| 321 | { |
| 322 | vassert(hregClass(src) == HRcVec128); |
| 323 | vassert(hregClass(dst) == HRcVec128); |
| 324 | return AMD64Instr_SseReRg(Asse_MOV, src, dst); |
| 325 | } |
| 326 | |
| 327 | /* Advance/retreat %rsp by n. */ |
| 328 | |
| 329 | static void add_to_rsp ( ISelEnv* env, Int n ) |
| 330 | { |
| 331 | vassert(n > 0 && n < 256 && (n%8) == 0); |
| 332 | addInstr(env, |
| 333 | AMD64Instr_Alu64R(Aalu_ADD, AMD64RMI_Imm(n), |
| 334 | hregAMD64_RSP())); |
| 335 | } |
| 336 | |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 337 | static void sub_from_rsp ( ISelEnv* env, Int n ) |
| 338 | { |
| 339 | vassert(n > 0 && n < 256 && (n%8) == 0); |
| 340 | addInstr(env, |
| 341 | AMD64Instr_Alu64R(Aalu_SUB, AMD64RMI_Imm(n), |
| 342 | hregAMD64_RSP())); |
| 343 | } |
| 344 | |
de | d403e79 | 2010-04-02 14:15:58 +0000 | [diff] [blame] | 345 | /* Push 64-bit constants on the stack. */ |
| 346 | static void push_uimm64( ISelEnv* env, ULong uimm64 ) |
| 347 | { |
| 348 | /* If uimm64 can be expressed as the sign extension of its |
| 349 | lower 32 bits, we can do it the easy way. */ |
| 350 | Long simm64 = (Long)uimm64; |
| 351 | if ( simm64 == ((simm64 << 32) >> 32) ) { |
| 352 | addInstr( env, AMD64Instr_Push(AMD64RMI_Imm( (UInt)uimm64 )) ); |
| 353 | } else { |
| 354 | HReg tmp = newVRegI(env); |
| 355 | addInstr( env, AMD64Instr_Imm64(uimm64, tmp) ); |
| 356 | addInstr( env, AMD64Instr_Push(AMD64RMI_Reg(tmp)) ); |
| 357 | } |
| 358 | } |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 359 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 360 | |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 361 | /* Used only in doHelperCall. If possible, produce a single |
| 362 | instruction which computes 'e' into 'dst'. If not possible, return |
| 363 | NULL. */ |
| 364 | |
| 365 | static AMD64Instr* iselIntExpr_single_instruction ( ISelEnv* env, |
| 366 | HReg dst, |
| 367 | IRExpr* e ) |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 368 | { |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 369 | /* Per comments in doHelperCall below, appearance of |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 370 | Iex_VECRET implies ill-formed IR. */ |
| 371 | vassert(e->tag != Iex_VECRET); |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 372 | |
| 373 | /* In this case we give out a copy of the BaseBlock pointer. */ |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 374 | if (UNLIKELY(e->tag == Iex_BBPTR)) { |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 375 | return mk_iMOVsd_RR( hregAMD64_RBP(), dst ); |
| 376 | } |
| 377 | |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 378 | vassert(typeOfIRExpr(env->type_env, e) == Ity_I64); |
| 379 | |
| 380 | if (e->tag == Iex_Const) { |
| 381 | vassert(e->Iex.Const.con->tag == Ico_U64); |
| 382 | if (fitsIn32Bits(e->Iex.Const.con->Ico.U64)) { |
| 383 | return AMD64Instr_Alu64R( |
| 384 | Aalu_MOV, |
| 385 | AMD64RMI_Imm(toUInt(e->Iex.Const.con->Ico.U64)), |
| 386 | dst |
| 387 | ); |
| 388 | } else { |
| 389 | return AMD64Instr_Imm64(e->Iex.Const.con->Ico.U64, dst); |
| 390 | } |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 391 | } |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 392 | |
| 393 | if (e->tag == Iex_RdTmp) { |
| 394 | HReg src = lookupIRTemp(env, e->Iex.RdTmp.tmp); |
| 395 | return mk_iMOVsd_RR(src, dst); |
| 396 | } |
| 397 | |
| 398 | if (e->tag == Iex_Get) { |
| 399 | vassert(e->Iex.Get.ty == Ity_I64); |
| 400 | return AMD64Instr_Alu64R( |
| 401 | Aalu_MOV, |
| 402 | AMD64RMI_Mem( |
| 403 | AMD64AMode_IR(e->Iex.Get.offset, |
| 404 | hregAMD64_RBP())), |
| 405 | dst); |
| 406 | } |
| 407 | |
| 408 | if (e->tag == Iex_Unop |
| 409 | && e->Iex.Unop.op == Iop_32Uto64 |
| 410 | && e->Iex.Unop.arg->tag == Iex_RdTmp) { |
| 411 | HReg src = lookupIRTemp(env, e->Iex.Unop.arg->Iex.RdTmp.tmp); |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 412 | return AMD64Instr_MovxLQ(False, src, dst); |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 413 | } |
| 414 | |
| 415 | if (0) { ppIRExpr(e); vex_printf("\n"); } |
| 416 | |
| 417 | return NULL; |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 418 | } |
| 419 | |
| 420 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 421 | /* Do a complete function call. |guard| is a Ity_Bit expression |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 422 | indicating whether or not the call happens. If guard==NULL, the |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 423 | call is unconditional. |retloc| is set to indicate where the |
| 424 | return value is after the call. The caller (of this fn) must |
| 425 | generate code to add |stackAdjustAfterCall| to the stack pointer |
| 426 | after the call is done. */ |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 427 | |
| 428 | static |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 429 | void doHelperCall ( /*OUT*/UInt* stackAdjustAfterCall, |
| 430 | /*OUT*/RetLoc* retloc, |
| 431 | ISelEnv* env, |
| 432 | IRExpr* guard, |
| 433 | IRCallee* cee, IRType retTy, IRExpr** args ) |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 434 | { |
| 435 | AMD64CondCode cc; |
| 436 | HReg argregs[6]; |
| 437 | HReg tmpregs[6]; |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 438 | AMD64Instr* fastinstrs[6]; |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 439 | UInt n_args, i; |
| 440 | |
| 441 | /* Set default returns. We'll update them later if needed. */ |
| 442 | *stackAdjustAfterCall = 0; |
| 443 | *retloc = mk_RetLoc_INVALID(); |
| 444 | |
| 445 | /* These are used for cross-checking that IR-level constraints on |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 446 | the use of IRExpr_VECRET() and IRExpr_BBPTR() are observed. */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 447 | UInt nVECRETs = 0; |
| 448 | UInt nBBPTRs = 0; |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 449 | |
| 450 | /* Marshal args for a call and do the call. |
| 451 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 452 | This function only deals with a tiny set of possibilities, which |
| 453 | cover all helpers in practice. The restrictions are that only |
| 454 | arguments in registers are supported, hence only 6x64 integer |
| 455 | bits in total can be passed. In fact the only supported arg |
| 456 | type is I64. |
| 457 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 458 | The return type can be I{64,32,16,8} or V{128,256}. In the |
| 459 | latter two cases, it is expected that |args| will contain the |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 460 | special node IRExpr_VECRET(), in which case this routine |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 461 | generates code to allocate space on the stack for the vector |
| 462 | return value. Since we are not passing any scalars on the |
| 463 | stack, it is enough to preallocate the return space before |
| 464 | marshalling any arguments, in this case. |
| 465 | |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 466 | |args| may also contain IRExpr_BBPTR(), in which case the |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 467 | value in %rbp is passed as the corresponding argument. |
| 468 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 469 | Generating code which is both efficient and correct when |
| 470 | parameters are to be passed in registers is difficult, for the |
| 471 | reasons elaborated in detail in comments attached to |
| 472 | doHelperCall() in priv/host-x86/isel.c. Here, we use a variant |
| 473 | of the method described in those comments. |
| 474 | |
| 475 | The problem is split into two cases: the fast scheme and the |
| 476 | slow scheme. In the fast scheme, arguments are computed |
| 477 | directly into the target (real) registers. This is only safe |
| 478 | when we can be sure that computation of each argument will not |
| 479 | trash any real registers set by computation of any other |
| 480 | argument. |
| 481 | |
| 482 | In the slow scheme, all args are first computed into vregs, and |
| 483 | once they are all done, they are moved to the relevant real |
| 484 | regs. This always gives correct code, but it also gives a bunch |
| 485 | of vreg-to-rreg moves which are usually redundant but are hard |
| 486 | for the register allocator to get rid of. |
| 487 | |
| 488 | To decide which scheme to use, all argument expressions are |
| 489 | first examined. If they are all so simple that it is clear they |
| 490 | will be evaluated without use of any fixed registers, use the |
| 491 | fast scheme, else use the slow scheme. Note also that only |
| 492 | unconditional calls may use the fast scheme, since having to |
| 493 | compute a condition expression could itself trash real |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 494 | registers. Note that for simplicity, in the case where |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 495 | IRExpr_VECRET() is present, we use the slow scheme. This is |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 496 | motivated by the desire to avoid any possible complexity |
| 497 | w.r.t. nested calls. |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 498 | |
| 499 | Note this requires being able to examine an expression and |
| 500 | determine whether or not evaluation of it might use a fixed |
| 501 | register. That requires knowledge of how the rest of this insn |
| 502 | selector works. Currently just the following 3 are regarded as |
| 503 | safe -- hopefully they cover the majority of arguments in |
| 504 | practice: IRExpr_Tmp IRExpr_Const IRExpr_Get. |
| 505 | */ |
| 506 | |
| 507 | /* Note that the cee->regparms field is meaningless on AMD64 host |
| 508 | (since there is only one calling convention) and so we always |
| 509 | ignore it. */ |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 510 | n_args = 0; |
| 511 | for (i = 0; args[i]; i++) |
| 512 | n_args++; |
| 513 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 514 | if (n_args > 6) |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 515 | vpanic("doHelperCall(AMD64): cannot currently handle > 6 args"); |
| 516 | |
| 517 | argregs[0] = hregAMD64_RDI(); |
| 518 | argregs[1] = hregAMD64_RSI(); |
| 519 | argregs[2] = hregAMD64_RDX(); |
| 520 | argregs[3] = hregAMD64_RCX(); |
| 521 | argregs[4] = hregAMD64_R8(); |
| 522 | argregs[5] = hregAMD64_R9(); |
| 523 | |
| 524 | tmpregs[0] = tmpregs[1] = tmpregs[2] = |
| 525 | tmpregs[3] = tmpregs[4] = tmpregs[5] = INVALID_HREG; |
| 526 | |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 527 | fastinstrs[0] = fastinstrs[1] = fastinstrs[2] = |
| 528 | fastinstrs[3] = fastinstrs[4] = fastinstrs[5] = NULL; |
| 529 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 530 | /* First decide which scheme (slow or fast) is to be used. First |
| 531 | assume the fast scheme, and select slow if any contraindications |
| 532 | (wow) appear. */ |
| 533 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 534 | /* We'll need space on the stack for the return value. Avoid |
| 535 | possible complications with nested calls by using the slow |
| 536 | scheme. */ |
| 537 | if (retTy == Ity_V128 || retTy == Ity_V256) |
| 538 | goto slowscheme; |
| 539 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 540 | if (guard) { |
| 541 | if (guard->tag == Iex_Const |
| 542 | && guard->Iex.Const.con->tag == Ico_U1 |
| 543 | && guard->Iex.Const.con->Ico.U1 == True) { |
| 544 | /* unconditional */ |
| 545 | } else { |
| 546 | /* Not manifestly unconditional -- be conservative. */ |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 547 | goto slowscheme; |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 548 | } |
| 549 | } |
| 550 | |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 551 | /* Ok, let's try for the fast scheme. If it doesn't pan out, we'll |
| 552 | use the slow scheme. Because this is tentative, we can't call |
| 553 | addInstr (that is, commit to) any instructions until we're |
| 554 | handled all the arguments. So park the resulting instructions |
| 555 | in a buffer and emit that if we're successful. */ |
| 556 | |
| 557 | /* FAST SCHEME */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 558 | /* In this loop, we process args that can be computed into the |
| 559 | destination (real) register with a single instruction, without |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 560 | using any fixed regs. That also includes IRExpr_BBPTR(), but |
| 561 | not IRExpr_VECRET(). Indeed, if the IR is well-formed, we can |
| 562 | never see IRExpr_VECRET() at this point, since the return-type |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 563 | check above should ensure all those cases use the slow scheme |
| 564 | instead. */ |
| 565 | vassert(n_args >= 0 && n_args <= 6); |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 566 | for (i = 0; i < n_args; i++) { |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 567 | IRExpr* arg = args[i]; |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 568 | if (LIKELY(!is_IRExpr_VECRET_or_BBPTR(arg))) { |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 569 | vassert(typeOfIRExpr(env->type_env, args[i]) == Ity_I64); |
| 570 | } |
| 571 | fastinstrs[i] |
| 572 | = iselIntExpr_single_instruction( env, argregs[i], args[i] ); |
| 573 | if (fastinstrs[i] == NULL) |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 574 | goto slowscheme; |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 575 | } |
| 576 | |
| 577 | /* Looks like we're in luck. Emit the accumulated instructions and |
| 578 | move on to doing the call itself. */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 579 | for (i = 0; i < n_args; i++) |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 580 | addInstr(env, fastinstrs[i]); |
| 581 | |
| 582 | /* Fast scheme only applies for unconditional calls. Hence: */ |
| 583 | cc = Acc_ALWAYS; |
| 584 | |
| 585 | goto handle_call; |
| 586 | |
| 587 | |
| 588 | /* SLOW SCHEME; move via temporaries */ |
| 589 | slowscheme: |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 590 | {} |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 591 | # if 0 /* debug only */ |
| 592 | if (n_args > 0) {for (i = 0; args[i]; i++) { |
| 593 | ppIRExpr(args[i]); vex_printf(" "); } |
| 594 | vex_printf("\n");} |
| 595 | # endif |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 596 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 597 | /* If we have a vector return type, allocate a place for it on the |
| 598 | stack and record its address. */ |
| 599 | HReg r_vecRetAddr = INVALID_HREG; |
| 600 | if (retTy == Ity_V128) { |
| 601 | r_vecRetAddr = newVRegI(env); |
| 602 | sub_from_rsp(env, 16); |
| 603 | addInstr(env, mk_iMOVsd_RR( hregAMD64_RSP(), r_vecRetAddr )); |
| 604 | } |
| 605 | else if (retTy == Ity_V256) { |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 606 | r_vecRetAddr = newVRegI(env); |
| 607 | sub_from_rsp(env, 32); |
| 608 | addInstr(env, mk_iMOVsd_RR( hregAMD64_RSP(), r_vecRetAddr )); |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 609 | } |
| 610 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 611 | vassert(n_args >= 0 && n_args <= 6); |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 612 | for (i = 0; i < n_args; i++) { |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 613 | IRExpr* arg = args[i]; |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 614 | if (UNLIKELY(arg->tag == Iex_BBPTR)) { |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 615 | tmpregs[i] = newVRegI(env); |
| 616 | addInstr(env, mk_iMOVsd_RR( hregAMD64_RBP(), tmpregs[i])); |
| 617 | nBBPTRs++; |
| 618 | } |
florian | 9041956 | 2013-08-15 20:54:52 +0000 | [diff] [blame] | 619 | else if (UNLIKELY(arg->tag == Iex_VECRET)) { |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 620 | /* We stashed the address of the return slot earlier, so just |
| 621 | retrieve it now. */ |
| 622 | vassert(!hregIsInvalid(r_vecRetAddr)); |
| 623 | tmpregs[i] = r_vecRetAddr; |
| 624 | nVECRETs++; |
| 625 | } |
| 626 | else { |
| 627 | vassert(typeOfIRExpr(env->type_env, args[i]) == Ity_I64); |
| 628 | tmpregs[i] = iselIntExpr_R(env, args[i]); |
| 629 | } |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 630 | } |
| 631 | |
| 632 | /* Now we can compute the condition. We can't do it earlier |
| 633 | because the argument computations could trash the condition |
| 634 | codes. Be a bit clever to handle the common case where the |
| 635 | guard is 1:Bit. */ |
| 636 | cc = Acc_ALWAYS; |
| 637 | if (guard) { |
| 638 | if (guard->tag == Iex_Const |
| 639 | && guard->Iex.Const.con->tag == Ico_U1 |
| 640 | && guard->Iex.Const.con->Ico.U1 == True) { |
| 641 | /* unconditional -- do nothing */ |
| 642 | } else { |
| 643 | cc = iselCondCode( env, guard ); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 644 | } |
| 645 | } |
| 646 | |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 647 | /* Move the args to their final destinations. */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 648 | for (i = 0; i < n_args; i++) { |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 649 | /* None of these insns, including any spill code that might |
| 650 | be generated, may alter the condition codes. */ |
| 651 | addInstr( env, mk_iMOVsd_RR( tmpregs[i], argregs[i] ) ); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 652 | } |
| 653 | |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 654 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 655 | /* Do final checks, set the return values, and generate the call |
| 656 | instruction proper. */ |
sewardj | 4d77a9c | 2007-08-25 23:21:08 +0000 | [diff] [blame] | 657 | handle_call: |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 658 | |
| 659 | if (retTy == Ity_V128 || retTy == Ity_V256) { |
| 660 | vassert(nVECRETs == 1); |
| 661 | } else { |
| 662 | vassert(nVECRETs == 0); |
| 663 | } |
| 664 | |
| 665 | vassert(nBBPTRs == 0 || nBBPTRs == 1); |
| 666 | |
| 667 | vassert(*stackAdjustAfterCall == 0); |
| 668 | vassert(is_RetLoc_INVALID(*retloc)); |
| 669 | switch (retTy) { |
| 670 | case Ity_INVALID: |
| 671 | /* Function doesn't return a value. */ |
| 672 | *retloc = mk_RetLoc_simple(RLPri_None); |
| 673 | break; |
| 674 | case Ity_I64: case Ity_I32: case Ity_I16: case Ity_I8: |
| 675 | *retloc = mk_RetLoc_simple(RLPri_Int); |
| 676 | break; |
| 677 | case Ity_V128: |
| 678 | *retloc = mk_RetLoc_spRel(RLPri_V128SpRel, 0); |
| 679 | *stackAdjustAfterCall = 16; |
| 680 | break; |
| 681 | case Ity_V256: |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 682 | *retloc = mk_RetLoc_spRel(RLPri_V256SpRel, 0); |
| 683 | *stackAdjustAfterCall = 32; |
| 684 | break; |
| 685 | default: |
| 686 | /* IR can denote other possible return types, but we don't |
| 687 | handle those here. */ |
| 688 | vassert(0); |
| 689 | } |
| 690 | |
| 691 | /* Finally, generate the call itself. This needs the *retloc value |
| 692 | set in the switch above, which is why it's at the end. */ |
| 693 | addInstr(env, |
| 694 | AMD64Instr_Call(cc, Ptr_to_ULong(cee->addr), n_args, *retloc)); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 695 | } |
| 696 | |
| 697 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 698 | /* Given a guest-state array descriptor, an index expression and a |
| 699 | bias, generate an AMD64AMode holding the relevant guest state |
| 700 | offset. */ |
| 701 | |
| 702 | static |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 703 | AMD64AMode* genGuestArrayOffset ( ISelEnv* env, IRRegArray* descr, |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 704 | IRExpr* off, Int bias ) |
| 705 | { |
| 706 | HReg tmp, roff; |
| 707 | Int elemSz = sizeofIRType(descr->elemTy); |
| 708 | Int nElems = descr->nElems; |
| 709 | |
| 710 | /* Throw out any cases not generated by an amd64 front end. In |
| 711 | theory there might be a day where we need to handle them -- if |
| 712 | we ever run non-amd64-guest on amd64 host. */ |
| 713 | |
| 714 | if (nElems != 8 || (elemSz != 1 && elemSz != 8)) |
| 715 | vpanic("genGuestArrayOffset(amd64 host)"); |
| 716 | |
| 717 | /* Compute off into a reg, %off. Then return: |
| 718 | |
| 719 | movq %off, %tmp |
| 720 | addq $bias, %tmp (if bias != 0) |
| 721 | andq %tmp, 7 |
| 722 | ... base(%rbp, %tmp, shift) ... |
| 723 | */ |
| 724 | tmp = newVRegI(env); |
| 725 | roff = iselIntExpr_R(env, off); |
| 726 | addInstr(env, mk_iMOVsd_RR(roff, tmp)); |
| 727 | if (bias != 0) { |
| 728 | /* Make sure the bias is sane, in the sense that there are |
| 729 | no significant bits above bit 30 in it. */ |
| 730 | vassert(-10000 < bias && bias < 10000); |
| 731 | addInstr(env, |
| 732 | AMD64Instr_Alu64R(Aalu_ADD, AMD64RMI_Imm(bias), tmp)); |
| 733 | } |
| 734 | addInstr(env, |
| 735 | AMD64Instr_Alu64R(Aalu_AND, AMD64RMI_Imm(7), tmp)); |
| 736 | vassert(elemSz == 1 || elemSz == 8); |
| 737 | return |
| 738 | AMD64AMode_IRRS( descr->base, hregAMD64_RBP(), tmp, |
| 739 | elemSz==8 ? 3 : 0); |
| 740 | } |
| 741 | |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 742 | |
| 743 | /* Set the SSE unit's rounding mode to default (%mxcsr = 0x1F80) */ |
| 744 | static |
| 745 | void set_SSE_rounding_default ( ISelEnv* env ) |
| 746 | { |
| 747 | /* pushq $DEFAULT_MXCSR |
| 748 | ldmxcsr 0(%rsp) |
| 749 | addq $8, %rsp |
| 750 | */ |
| 751 | AMD64AMode* zero_rsp = AMD64AMode_IR(0, hregAMD64_RSP()); |
| 752 | addInstr(env, AMD64Instr_Push(AMD64RMI_Imm(DEFAULT_MXCSR))); |
| 753 | addInstr(env, AMD64Instr_LdMXCSR(zero_rsp)); |
| 754 | add_to_rsp(env, 8); |
| 755 | } |
| 756 | |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 757 | /* Mess with the FPU's rounding mode: set to the default rounding mode |
| 758 | (DEFAULT_FPUCW). */ |
| 759 | static |
| 760 | void set_FPU_rounding_default ( ISelEnv* env ) |
| 761 | { |
| 762 | /* movq $DEFAULT_FPUCW, -8(%rsp) |
| 763 | fldcw -8(%esp) |
| 764 | */ |
| 765 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
| 766 | addInstr(env, AMD64Instr_Alu64M( |
| 767 | Aalu_MOV, AMD64RI_Imm(DEFAULT_FPUCW), m8_rsp)); |
| 768 | addInstr(env, AMD64Instr_A87LdCW(m8_rsp)); |
| 769 | } |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 770 | |
| 771 | |
| 772 | /* Mess with the SSE unit's rounding mode: 'mode' is an I32-typed |
| 773 | expression denoting a value in the range 0 .. 3, indicating a round |
| 774 | mode encoded as per type IRRoundingMode. Set the SSE machinery to |
| 775 | have the same rounding. |
| 776 | */ |
| 777 | static |
| 778 | void set_SSE_rounding_mode ( ISelEnv* env, IRExpr* mode ) |
| 779 | { |
| 780 | /* Note: this sequence only makes sense because DEFAULT_MXCSR has |
| 781 | both rounding bits == 0. If that wasn't the case, we couldn't |
| 782 | create a new rounding field simply by ORing the new value into |
| 783 | place. */ |
| 784 | |
| 785 | /* movq $3, %reg |
| 786 | andq [[mode]], %reg -- shouldn't be needed; paranoia |
| 787 | shlq $13, %reg |
| 788 | orq $DEFAULT_MXCSR, %reg |
| 789 | pushq %reg |
| 790 | ldmxcsr 0(%esp) |
| 791 | addq $8, %rsp |
| 792 | */ |
| 793 | HReg reg = newVRegI(env); |
| 794 | AMD64AMode* zero_rsp = AMD64AMode_IR(0, hregAMD64_RSP()); |
| 795 | addInstr(env, AMD64Instr_Alu64R(Aalu_MOV, AMD64RMI_Imm(3), reg)); |
| 796 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, |
| 797 | iselIntExpr_RMI(env, mode), reg)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 798 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 13, reg)); |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 799 | addInstr(env, AMD64Instr_Alu64R( |
| 800 | Aalu_OR, AMD64RMI_Imm(DEFAULT_MXCSR), reg)); |
| 801 | addInstr(env, AMD64Instr_Push(AMD64RMI_Reg(reg))); |
| 802 | addInstr(env, AMD64Instr_LdMXCSR(zero_rsp)); |
| 803 | add_to_rsp(env, 8); |
| 804 | } |
| 805 | |
| 806 | |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 807 | /* Mess with the FPU's rounding mode: 'mode' is an I32-typed |
| 808 | expression denoting a value in the range 0 .. 3, indicating a round |
| 809 | mode encoded as per type IRRoundingMode. Set the x87 FPU to have |
| 810 | the same rounding. |
| 811 | */ |
| 812 | static |
| 813 | void set_FPU_rounding_mode ( ISelEnv* env, IRExpr* mode ) |
| 814 | { |
| 815 | HReg rrm = iselIntExpr_R(env, mode); |
| 816 | HReg rrm2 = newVRegI(env); |
| 817 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
| 818 | |
| 819 | /* movq %rrm, %rrm2 |
| 820 | andq $3, %rrm2 -- shouldn't be needed; paranoia |
| 821 | shlq $10, %rrm2 |
| 822 | orq $DEFAULT_FPUCW, %rrm2 |
| 823 | movq %rrm2, -8(%rsp) |
| 824 | fldcw -8(%esp) |
| 825 | */ |
| 826 | addInstr(env, mk_iMOVsd_RR(rrm, rrm2)); |
| 827 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, AMD64RMI_Imm(3), rrm2)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 828 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 10, rrm2)); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 829 | addInstr(env, AMD64Instr_Alu64R(Aalu_OR, |
| 830 | AMD64RMI_Imm(DEFAULT_FPUCW), rrm2)); |
| 831 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, |
| 832 | AMD64RI_Reg(rrm2), m8_rsp)); |
| 833 | addInstr(env, AMD64Instr_A87LdCW(m8_rsp)); |
| 834 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 835 | |
| 836 | |
sewardj | ac53044 | 2005-05-11 16:13:37 +0000 | [diff] [blame] | 837 | /* Generate all-zeroes into a new vector register. |
| 838 | */ |
| 839 | static HReg generate_zeroes_V128 ( ISelEnv* env ) |
| 840 | { |
| 841 | HReg dst = newVRegV(env); |
| 842 | addInstr(env, AMD64Instr_SseReRg(Asse_XOR, dst, dst)); |
| 843 | return dst; |
| 844 | } |
| 845 | |
| 846 | /* Generate all-ones into a new vector register. |
| 847 | */ |
| 848 | static HReg generate_ones_V128 ( ISelEnv* env ) |
| 849 | { |
| 850 | HReg dst = newVRegV(env); |
| 851 | addInstr(env, AMD64Instr_SseReRg(Asse_CMPEQ32, dst, dst)); |
| 852 | return dst; |
| 853 | } |
| 854 | |
| 855 | |
sewardj | 0971734 | 2005-05-05 21:34:02 +0000 | [diff] [blame] | 856 | /* Generate !src into a new vector register. Amazing that there isn't |
| 857 | a less crappy way to do this. |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 858 | */ |
| 859 | static HReg do_sse_NotV128 ( ISelEnv* env, HReg src ) |
| 860 | { |
sewardj | ac53044 | 2005-05-11 16:13:37 +0000 | [diff] [blame] | 861 | HReg dst = generate_ones_V128(env); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 862 | addInstr(env, AMD64Instr_SseReRg(Asse_XOR, src, dst)); |
| 863 | return dst; |
| 864 | } |
| 865 | |
| 866 | |
sewardj | acfbd7d | 2010-08-17 22:52:08 +0000 | [diff] [blame] | 867 | /* Expand the given byte into a 64-bit word, by cloning each bit |
| 868 | 8 times. */ |
| 869 | static ULong bitmask8_to_bytemask64 ( UShort w8 ) |
| 870 | { |
| 871 | vassert(w8 == (w8 & 0xFF)); |
| 872 | ULong w64 = 0; |
| 873 | Int i; |
| 874 | for (i = 0; i < 8; i++) { |
| 875 | if (w8 & (1<<i)) |
| 876 | w64 |= (0xFFULL << (8 * i)); |
| 877 | } |
| 878 | return w64; |
| 879 | } |
| 880 | |
| 881 | |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 882 | /*---------------------------------------------------------*/ |
| 883 | /*--- ISEL: Integer expressions (64/32/16/8 bit) ---*/ |
| 884 | /*---------------------------------------------------------*/ |
| 885 | |
| 886 | /* Select insns for an integer-typed expression, and add them to the |
| 887 | code list. Return a reg holding the result. This reg will be a |
| 888 | virtual register. THE RETURNED REG MUST NOT BE MODIFIED. If you |
| 889 | want to modify it, ask for a new vreg, copy it in there, and modify |
| 890 | the copy. The register allocator will do its best to map both |
| 891 | vregs to the same real register, so the copies will often disappear |
| 892 | later in the game. |
| 893 | |
| 894 | This should handle expressions of 64, 32, 16 and 8-bit type. All |
| 895 | results are returned in a 64-bit register. For 32-, 16- and 8-bit |
sewardj | e13074c | 2012-11-08 10:57:08 +0000 | [diff] [blame] | 896 | expressions, the upper 32/48/56 bits are arbitrary, so you should |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 897 | mask or sign extend partial values if necessary. |
| 898 | */ |
| 899 | |
| 900 | static HReg iselIntExpr_R ( ISelEnv* env, IRExpr* e ) |
| 901 | { |
| 902 | HReg r = iselIntExpr_R_wrk(env, e); |
| 903 | /* sanity checks ... */ |
| 904 | # if 0 |
| 905 | vex_printf("\niselIntExpr_R: "); ppIRExpr(e); vex_printf("\n"); |
| 906 | # endif |
| 907 | vassert(hregClass(r) == HRcInt64); |
| 908 | vassert(hregIsVirtual(r)); |
| 909 | return r; |
| 910 | } |
| 911 | |
| 912 | /* DO NOT CALL THIS DIRECTLY ! */ |
| 913 | static HReg iselIntExpr_R_wrk ( ISelEnv* env, IRExpr* e ) |
| 914 | { |
sewardj | e790566 | 2005-05-09 18:15:21 +0000 | [diff] [blame] | 915 | /* Used for unary/binary SIMD64 ops. */ |
| 916 | HWord fn = 0; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 917 | Bool second_is_UInt; |
sewardj | e790566 | 2005-05-09 18:15:21 +0000 | [diff] [blame] | 918 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 919 | MatchInfo mi; |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 920 | DECLARE_PATTERN(p_1Uto8_64to1); |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 921 | DECLARE_PATTERN(p_LDle8_then_8Uto64); |
| 922 | DECLARE_PATTERN(p_LDle16_then_16Uto64); |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 923 | |
| 924 | IRType ty = typeOfIRExpr(env->type_env,e); |
sewardj | 13f12a5 | 2011-05-03 07:51:49 +0000 | [diff] [blame] | 925 | switch (ty) { |
| 926 | case Ity_I64: case Ity_I32: case Ity_I16: case Ity_I8: break; |
| 927 | default: vassert(0); |
| 928 | } |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 929 | |
| 930 | switch (e->tag) { |
| 931 | |
| 932 | /* --------- TEMP --------- */ |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 933 | case Iex_RdTmp: { |
| 934 | return lookupIRTemp(env, e->Iex.RdTmp.tmp); |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 935 | } |
| 936 | |
| 937 | /* --------- LOAD --------- */ |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 938 | case Iex_Load: { |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 939 | HReg dst = newVRegI(env); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 940 | AMD64AMode* amode = iselIntExpr_AMode ( env, e->Iex.Load.addr ); |
| 941 | |
sewardj | e9d8a26 | 2009-07-01 08:06:34 +0000 | [diff] [blame] | 942 | /* We can't handle big-endian loads, nor load-linked. */ |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 943 | if (e->Iex.Load.end != Iend_LE) |
| 944 | goto irreducible; |
| 945 | |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 946 | if (ty == Ity_I64) { |
| 947 | addInstr(env, AMD64Instr_Alu64R(Aalu_MOV, |
| 948 | AMD64RMI_Mem(amode), dst) ); |
| 949 | return dst; |
| 950 | } |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 951 | if (ty == Ity_I32) { |
| 952 | addInstr(env, AMD64Instr_LoadEX(4,False,amode,dst)); |
| 953 | return dst; |
| 954 | } |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 955 | if (ty == Ity_I16) { |
| 956 | addInstr(env, AMD64Instr_LoadEX(2,False,amode,dst)); |
| 957 | return dst; |
| 958 | } |
sewardj | 7f039c4 | 2005-02-04 21:13:55 +0000 | [diff] [blame] | 959 | if (ty == Ity_I8) { |
| 960 | addInstr(env, AMD64Instr_LoadEX(1,False,amode,dst)); |
| 961 | return dst; |
| 962 | } |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 963 | break; |
| 964 | } |
| 965 | |
| 966 | /* --------- BINARY OP --------- */ |
| 967 | case Iex_Binop: { |
| 968 | AMD64AluOp aluOp; |
| 969 | AMD64ShiftOp shOp; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 970 | |
sewardj | eb17e49 | 2007-08-25 23:07:44 +0000 | [diff] [blame] | 971 | /* Pattern: Sub64(0,x) */ |
| 972 | /* and: Sub32(0,x) */ |
| 973 | if ((e->Iex.Binop.op == Iop_Sub64 && isZeroU64(e->Iex.Binop.arg1)) |
| 974 | || (e->Iex.Binop.op == Iop_Sub32 && isZeroU32(e->Iex.Binop.arg1))) { |
| 975 | HReg dst = newVRegI(env); |
| 976 | HReg reg = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 977 | addInstr(env, mk_iMOVsd_RR(reg,dst)); |
| 978 | addInstr(env, AMD64Instr_Unary64(Aun_NEG,dst)); |
| 979 | return dst; |
| 980 | } |
| 981 | |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 982 | /* Is it an addition or logical style op? */ |
| 983 | switch (e->Iex.Binop.op) { |
| 984 | case Iop_Add8: case Iop_Add16: case Iop_Add32: case Iop_Add64: |
| 985 | aluOp = Aalu_ADD; break; |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 986 | case Iop_Sub8: case Iop_Sub16: case Iop_Sub32: case Iop_Sub64: |
| 987 | aluOp = Aalu_SUB; break; |
| 988 | case Iop_And8: case Iop_And16: case Iop_And32: case Iop_And64: |
| 989 | aluOp = Aalu_AND; break; |
sewardj | e169895 | 2005-02-08 15:02:39 +0000 | [diff] [blame] | 990 | case Iop_Or8: case Iop_Or16: case Iop_Or32: case Iop_Or64: |
sewardj | 3119107 | 2005-02-05 18:24:47 +0000 | [diff] [blame] | 991 | aluOp = Aalu_OR; break; |
sewardj | e169895 | 2005-02-08 15:02:39 +0000 | [diff] [blame] | 992 | case Iop_Xor8: case Iop_Xor16: case Iop_Xor32: case Iop_Xor64: |
| 993 | aluOp = Aalu_XOR; break; |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 994 | case Iop_Mul16: case Iop_Mul32: case Iop_Mul64: |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 995 | aluOp = Aalu_MUL; break; |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 996 | default: |
| 997 | aluOp = Aalu_INVALID; break; |
| 998 | } |
| 999 | /* For commutative ops we assume any literal |
| 1000 | values are on the second operand. */ |
| 1001 | if (aluOp != Aalu_INVALID) { |
| 1002 | HReg dst = newVRegI(env); |
| 1003 | HReg reg = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1004 | AMD64RMI* rmi = iselIntExpr_RMI(env, e->Iex.Binop.arg2); |
| 1005 | addInstr(env, mk_iMOVsd_RR(reg,dst)); |
| 1006 | addInstr(env, AMD64Instr_Alu64R(aluOp, rmi, dst)); |
| 1007 | return dst; |
| 1008 | } |
| 1009 | |
| 1010 | /* Perhaps a shift op? */ |
| 1011 | switch (e->Iex.Binop.op) { |
| 1012 | case Iop_Shl64: case Iop_Shl32: case Iop_Shl16: case Iop_Shl8: |
| 1013 | shOp = Ash_SHL; break; |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 1014 | case Iop_Shr64: case Iop_Shr32: case Iop_Shr16: case Iop_Shr8: |
| 1015 | shOp = Ash_SHR; break; |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1016 | case Iop_Sar64: case Iop_Sar32: case Iop_Sar16: case Iop_Sar8: |
| 1017 | shOp = Ash_SAR; break; |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1018 | default: |
| 1019 | shOp = Ash_INVALID; break; |
| 1020 | } |
| 1021 | if (shOp != Ash_INVALID) { |
| 1022 | HReg dst = newVRegI(env); |
| 1023 | |
| 1024 | /* regL = the value to be shifted */ |
| 1025 | HReg regL = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1026 | addInstr(env, mk_iMOVsd_RR(regL,dst)); |
| 1027 | |
| 1028 | /* Do any necessary widening for 32/16/8 bit operands */ |
| 1029 | switch (e->Iex.Binop.op) { |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1030 | case Iop_Shr64: case Iop_Shl64: case Iop_Sar64: |
| 1031 | break; |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1032 | case Iop_Shl32: case Iop_Shl16: case Iop_Shl8: |
sewardj | b095fba | 2005-02-13 14:13:04 +0000 | [diff] [blame] | 1033 | break; |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1034 | case Iop_Shr8: |
| 1035 | addInstr(env, AMD64Instr_Alu64R( |
| 1036 | Aalu_AND, AMD64RMI_Imm(0xFF), dst)); |
| 1037 | break; |
| 1038 | case Iop_Shr16: |
| 1039 | addInstr(env, AMD64Instr_Alu64R( |
| 1040 | Aalu_AND, AMD64RMI_Imm(0xFFFF), dst)); |
| 1041 | break; |
sewardj | b095fba | 2005-02-13 14:13:04 +0000 | [diff] [blame] | 1042 | case Iop_Shr32: |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1043 | addInstr(env, AMD64Instr_MovxLQ(False, dst, dst)); |
sewardj | b095fba | 2005-02-13 14:13:04 +0000 | [diff] [blame] | 1044 | break; |
sewardj | e83d9b2 | 2005-08-13 23:58:34 +0000 | [diff] [blame] | 1045 | case Iop_Sar8: |
| 1046 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 56, dst)); |
| 1047 | addInstr(env, AMD64Instr_Sh64(Ash_SAR, 56, dst)); |
| 1048 | break; |
| 1049 | case Iop_Sar16: |
| 1050 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 48, dst)); |
| 1051 | addInstr(env, AMD64Instr_Sh64(Ash_SAR, 48, dst)); |
| 1052 | break; |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1053 | case Iop_Sar32: |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1054 | addInstr(env, AMD64Instr_MovxLQ(True, dst, dst)); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1055 | break; |
| 1056 | default: |
sewardj | 909c06d | 2005-02-19 22:47:41 +0000 | [diff] [blame] | 1057 | ppIROp(e->Iex.Binop.op); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1058 | vassert(0); |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1059 | } |
| 1060 | |
| 1061 | /* Now consider the shift amount. If it's a literal, we |
| 1062 | can do a much better job than the general case. */ |
| 1063 | if (e->Iex.Binop.arg2->tag == Iex_Const) { |
| 1064 | /* assert that the IR is well-typed */ |
| 1065 | Int nshift; |
| 1066 | vassert(e->Iex.Binop.arg2->Iex.Const.con->tag == Ico_U8); |
| 1067 | nshift = e->Iex.Binop.arg2->Iex.Const.con->Ico.U8; |
| 1068 | vassert(nshift >= 0); |
| 1069 | if (nshift > 0) |
| 1070 | /* Can't allow nshift==0 since that means %cl */ |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1071 | addInstr(env, AMD64Instr_Sh64(shOp, nshift, dst)); |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1072 | } else { |
| 1073 | /* General case; we have to force the amount into %cl. */ |
| 1074 | HReg regR = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 1075 | addInstr(env, mk_iMOVsd_RR(regR,hregAMD64_RCX())); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1076 | addInstr(env, AMD64Instr_Sh64(shOp, 0/* %cl */, dst)); |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1077 | } |
| 1078 | return dst; |
| 1079 | } |
| 1080 | |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1081 | /* Deal with 64-bit SIMD binary ops */ |
| 1082 | second_is_UInt = False; |
| 1083 | switch (e->Iex.Binop.op) { |
| 1084 | case Iop_Add8x8: |
| 1085 | fn = (HWord)h_generic_calc_Add8x8; break; |
| 1086 | case Iop_Add16x4: |
| 1087 | fn = (HWord)h_generic_calc_Add16x4; break; |
| 1088 | case Iop_Add32x2: |
| 1089 | fn = (HWord)h_generic_calc_Add32x2; break; |
sewardj | a7ba8c4 | 2005-05-10 20:08:34 +0000 | [diff] [blame] | 1090 | |
| 1091 | case Iop_Avg8Ux8: |
| 1092 | fn = (HWord)h_generic_calc_Avg8Ux8; break; |
| 1093 | case Iop_Avg16Ux4: |
| 1094 | fn = (HWord)h_generic_calc_Avg16Ux4; break; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1095 | |
| 1096 | case Iop_CmpEQ8x8: |
| 1097 | fn = (HWord)h_generic_calc_CmpEQ8x8; break; |
| 1098 | case Iop_CmpEQ16x4: |
| 1099 | fn = (HWord)h_generic_calc_CmpEQ16x4; break; |
| 1100 | case Iop_CmpEQ32x2: |
| 1101 | fn = (HWord)h_generic_calc_CmpEQ32x2; break; |
| 1102 | |
| 1103 | case Iop_CmpGT8Sx8: |
| 1104 | fn = (HWord)h_generic_calc_CmpGT8Sx8; break; |
| 1105 | case Iop_CmpGT16Sx4: |
| 1106 | fn = (HWord)h_generic_calc_CmpGT16Sx4; break; |
| 1107 | case Iop_CmpGT32Sx2: |
| 1108 | fn = (HWord)h_generic_calc_CmpGT32Sx2; break; |
| 1109 | |
| 1110 | case Iop_InterleaveHI8x8: |
| 1111 | fn = (HWord)h_generic_calc_InterleaveHI8x8; break; |
| 1112 | case Iop_InterleaveLO8x8: |
| 1113 | fn = (HWord)h_generic_calc_InterleaveLO8x8; break; |
| 1114 | case Iop_InterleaveHI16x4: |
| 1115 | fn = (HWord)h_generic_calc_InterleaveHI16x4; break; |
| 1116 | case Iop_InterleaveLO16x4: |
| 1117 | fn = (HWord)h_generic_calc_InterleaveLO16x4; break; |
| 1118 | case Iop_InterleaveHI32x2: |
| 1119 | fn = (HWord)h_generic_calc_InterleaveHI32x2; break; |
| 1120 | case Iop_InterleaveLO32x2: |
| 1121 | fn = (HWord)h_generic_calc_InterleaveLO32x2; break; |
sewardj | d166e28 | 2008-02-06 11:42:45 +0000 | [diff] [blame] | 1122 | case Iop_CatOddLanes16x4: |
| 1123 | fn = (HWord)h_generic_calc_CatOddLanes16x4; break; |
| 1124 | case Iop_CatEvenLanes16x4: |
| 1125 | fn = (HWord)h_generic_calc_CatEvenLanes16x4; break; |
| 1126 | case Iop_Perm8x8: |
| 1127 | fn = (HWord)h_generic_calc_Perm8x8; break; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1128 | |
sewardj | a7ba8c4 | 2005-05-10 20:08:34 +0000 | [diff] [blame] | 1129 | case Iop_Max8Ux8: |
| 1130 | fn = (HWord)h_generic_calc_Max8Ux8; break; |
| 1131 | case Iop_Max16Sx4: |
| 1132 | fn = (HWord)h_generic_calc_Max16Sx4; break; |
| 1133 | case Iop_Min8Ux8: |
| 1134 | fn = (HWord)h_generic_calc_Min8Ux8; break; |
| 1135 | case Iop_Min16Sx4: |
| 1136 | fn = (HWord)h_generic_calc_Min16Sx4; break; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1137 | |
| 1138 | case Iop_Mul16x4: |
| 1139 | fn = (HWord)h_generic_calc_Mul16x4; break; |
sewardj | d166e28 | 2008-02-06 11:42:45 +0000 | [diff] [blame] | 1140 | case Iop_Mul32x2: |
| 1141 | fn = (HWord)h_generic_calc_Mul32x2; break; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1142 | case Iop_MulHi16Sx4: |
| 1143 | fn = (HWord)h_generic_calc_MulHi16Sx4; break; |
sewardj | a7ba8c4 | 2005-05-10 20:08:34 +0000 | [diff] [blame] | 1144 | case Iop_MulHi16Ux4: |
| 1145 | fn = (HWord)h_generic_calc_MulHi16Ux4; break; |
| 1146 | |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1147 | case Iop_QAdd8Sx8: |
| 1148 | fn = (HWord)h_generic_calc_QAdd8Sx8; break; |
| 1149 | case Iop_QAdd16Sx4: |
| 1150 | fn = (HWord)h_generic_calc_QAdd16Sx4; break; |
| 1151 | case Iop_QAdd8Ux8: |
| 1152 | fn = (HWord)h_generic_calc_QAdd8Ux8; break; |
| 1153 | case Iop_QAdd16Ux4: |
| 1154 | fn = (HWord)h_generic_calc_QAdd16Ux4; break; |
| 1155 | |
sewardj | 5f438dd | 2011-06-16 11:36:23 +0000 | [diff] [blame] | 1156 | case Iop_QNarrowBin32Sto16Sx4: |
| 1157 | fn = (HWord)h_generic_calc_QNarrowBin32Sto16Sx4; break; |
| 1158 | case Iop_QNarrowBin16Sto8Sx8: |
| 1159 | fn = (HWord)h_generic_calc_QNarrowBin16Sto8Sx8; break; |
| 1160 | case Iop_QNarrowBin16Sto8Ux8: |
| 1161 | fn = (HWord)h_generic_calc_QNarrowBin16Sto8Ux8; break; |
sewardj | ad2c9ea | 2011-10-22 09:32:16 +0000 | [diff] [blame] | 1162 | case Iop_NarrowBin16to8x8: |
| 1163 | fn = (HWord)h_generic_calc_NarrowBin16to8x8; break; |
| 1164 | case Iop_NarrowBin32to16x4: |
| 1165 | fn = (HWord)h_generic_calc_NarrowBin32to16x4; break; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1166 | |
| 1167 | case Iop_QSub8Sx8: |
| 1168 | fn = (HWord)h_generic_calc_QSub8Sx8; break; |
| 1169 | case Iop_QSub16Sx4: |
| 1170 | fn = (HWord)h_generic_calc_QSub16Sx4; break; |
| 1171 | case Iop_QSub8Ux8: |
| 1172 | fn = (HWord)h_generic_calc_QSub8Ux8; break; |
| 1173 | case Iop_QSub16Ux4: |
| 1174 | fn = (HWord)h_generic_calc_QSub16Ux4; break; |
| 1175 | |
| 1176 | case Iop_Sub8x8: |
| 1177 | fn = (HWord)h_generic_calc_Sub8x8; break; |
| 1178 | case Iop_Sub16x4: |
| 1179 | fn = (HWord)h_generic_calc_Sub16x4; break; |
| 1180 | case Iop_Sub32x2: |
| 1181 | fn = (HWord)h_generic_calc_Sub32x2; break; |
| 1182 | |
| 1183 | case Iop_ShlN32x2: |
| 1184 | fn = (HWord)h_generic_calc_ShlN32x2; |
| 1185 | second_is_UInt = True; |
| 1186 | break; |
| 1187 | case Iop_ShlN16x4: |
| 1188 | fn = (HWord)h_generic_calc_ShlN16x4; |
| 1189 | second_is_UInt = True; |
| 1190 | break; |
sewardj | d166e28 | 2008-02-06 11:42:45 +0000 | [diff] [blame] | 1191 | case Iop_ShlN8x8: |
| 1192 | fn = (HWord)h_generic_calc_ShlN8x8; |
| 1193 | second_is_UInt = True; |
| 1194 | break; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1195 | case Iop_ShrN32x2: |
| 1196 | fn = (HWord)h_generic_calc_ShrN32x2; |
| 1197 | second_is_UInt = True; |
| 1198 | break; |
| 1199 | case Iop_ShrN16x4: |
| 1200 | fn = (HWord)h_generic_calc_ShrN16x4; |
| 1201 | second_is_UInt = True; |
| 1202 | break; |
| 1203 | case Iop_SarN32x2: |
| 1204 | fn = (HWord)h_generic_calc_SarN32x2; |
| 1205 | second_is_UInt = True; |
| 1206 | break; |
| 1207 | case Iop_SarN16x4: |
| 1208 | fn = (HWord)h_generic_calc_SarN16x4; |
| 1209 | second_is_UInt = True; |
| 1210 | break; |
sewardj | 02f79f1 | 2007-09-01 18:59:53 +0000 | [diff] [blame] | 1211 | case Iop_SarN8x8: |
| 1212 | fn = (HWord)h_generic_calc_SarN8x8; |
| 1213 | second_is_UInt = True; |
| 1214 | break; |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1215 | |
| 1216 | default: |
| 1217 | fn = (HWord)0; break; |
| 1218 | } |
| 1219 | if (fn != (HWord)0) { |
| 1220 | /* Note: the following assumes all helpers are of signature |
| 1221 | ULong fn ( ULong, ULong ), and they are |
| 1222 | not marked as regparm functions. |
| 1223 | */ |
| 1224 | HReg dst = newVRegI(env); |
| 1225 | HReg argL = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1226 | HReg argR = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 1227 | if (second_is_UInt) |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1228 | addInstr(env, AMD64Instr_MovxLQ(False, argR, argR)); |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1229 | addInstr(env, mk_iMOVsd_RR(argL, hregAMD64_RDI()) ); |
| 1230 | addInstr(env, mk_iMOVsd_RR(argR, hregAMD64_RSI()) ); |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 1231 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, 2, |
| 1232 | mk_RetLoc_simple(RLPri_Int) )); |
sewardj | 8711f66 | 2005-05-09 17:52:56 +0000 | [diff] [blame] | 1233 | addInstr(env, mk_iMOVsd_RR(hregAMD64_RAX(), dst)); |
| 1234 | return dst; |
| 1235 | } |
| 1236 | |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1237 | /* Handle misc other ops. */ |
| 1238 | |
sewardj | 478646f | 2008-05-01 20:13:04 +0000 | [diff] [blame] | 1239 | if (e->Iex.Binop.op == Iop_Max32U) { |
sewardj | 9cc2bbf | 2011-06-05 17:56:03 +0000 | [diff] [blame] | 1240 | HReg src1 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1241 | HReg dst = newVRegI(env); |
| 1242 | HReg src2 = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 1243 | addInstr(env, mk_iMOVsd_RR(src1, dst)); |
| 1244 | addInstr(env, AMD64Instr_Alu32R(Aalu_CMP, AMD64RMI_Reg(src2), dst)); |
sewardj | 478646f | 2008-05-01 20:13:04 +0000 | [diff] [blame] | 1245 | addInstr(env, AMD64Instr_CMov64(Acc_B, AMD64RM_Reg(src2), dst)); |
| 1246 | return dst; |
| 1247 | } |
| 1248 | |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1249 | if (e->Iex.Binop.op == Iop_DivModS64to32 |
| 1250 | || e->Iex.Binop.op == Iop_DivModU64to32) { |
| 1251 | /* 64 x 32 -> (32(rem),32(div)) division */ |
| 1252 | /* Get the 64-bit operand into edx:eax, and the other into |
| 1253 | any old R/M. */ |
| 1254 | HReg rax = hregAMD64_RAX(); |
| 1255 | HReg rdx = hregAMD64_RDX(); |
| 1256 | HReg dst = newVRegI(env); |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 1257 | Bool syned = toBool(e->Iex.Binop.op == Iop_DivModS64to32); |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1258 | AMD64RM* rmRight = iselIntExpr_RM(env, e->Iex.Binop.arg2); |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1259 | /* Compute the left operand into a reg, and then |
| 1260 | put the top half in edx and the bottom in eax. */ |
| 1261 | HReg left64 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1262 | addInstr(env, mk_iMOVsd_RR(left64, rdx)); |
| 1263 | addInstr(env, mk_iMOVsd_RR(left64, rax)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1264 | addInstr(env, AMD64Instr_Sh64(Ash_SHR, 32, rdx)); |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1265 | addInstr(env, AMD64Instr_Div(syned, 4, rmRight)); |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1266 | addInstr(env, AMD64Instr_MovxLQ(False, rdx, rdx)); |
| 1267 | addInstr(env, AMD64Instr_MovxLQ(False, rax, rax)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1268 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 32, rdx)); |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1269 | addInstr(env, mk_iMOVsd_RR(rax, dst)); |
| 1270 | addInstr(env, AMD64Instr_Alu64R(Aalu_OR, AMD64RMI_Reg(rdx), dst)); |
| 1271 | return dst; |
| 1272 | } |
| 1273 | |
| 1274 | if (e->Iex.Binop.op == Iop_32HLto64) { |
| 1275 | HReg hi32 = newVRegI(env); |
| 1276 | HReg lo32 = newVRegI(env); |
| 1277 | HReg hi32s = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1278 | HReg lo32s = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 1279 | addInstr(env, mk_iMOVsd_RR(hi32s, hi32)); |
| 1280 | addInstr(env, mk_iMOVsd_RR(lo32s, lo32)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1281 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 32, hi32)); |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1282 | addInstr(env, AMD64Instr_MovxLQ(False, lo32, lo32)); |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1283 | addInstr(env, AMD64Instr_Alu64R( |
| 1284 | Aalu_OR, AMD64RMI_Reg(lo32), hi32)); |
| 1285 | return hi32; |
| 1286 | } |
| 1287 | |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1288 | if (e->Iex.Binop.op == Iop_16HLto32) { |
| 1289 | HReg hi16 = newVRegI(env); |
| 1290 | HReg lo16 = newVRegI(env); |
| 1291 | HReg hi16s = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1292 | HReg lo16s = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 1293 | addInstr(env, mk_iMOVsd_RR(hi16s, hi16)); |
| 1294 | addInstr(env, mk_iMOVsd_RR(lo16s, lo16)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1295 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 16, hi16)); |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1296 | addInstr(env, AMD64Instr_Alu64R( |
| 1297 | Aalu_AND, AMD64RMI_Imm(0xFFFF), lo16)); |
| 1298 | addInstr(env, AMD64Instr_Alu64R( |
| 1299 | Aalu_OR, AMD64RMI_Reg(lo16), hi16)); |
| 1300 | return hi16; |
| 1301 | } |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1302 | |
sewardj | a64f8ad | 2005-04-24 00:26:37 +0000 | [diff] [blame] | 1303 | if (e->Iex.Binop.op == Iop_8HLto16) { |
| 1304 | HReg hi8 = newVRegI(env); |
| 1305 | HReg lo8 = newVRegI(env); |
| 1306 | HReg hi8s = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1307 | HReg lo8s = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 1308 | addInstr(env, mk_iMOVsd_RR(hi8s, hi8)); |
| 1309 | addInstr(env, mk_iMOVsd_RR(lo8s, lo8)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1310 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 8, hi8)); |
sewardj | a64f8ad | 2005-04-24 00:26:37 +0000 | [diff] [blame] | 1311 | addInstr(env, AMD64Instr_Alu64R( |
| 1312 | Aalu_AND, AMD64RMI_Imm(0xFF), lo8)); |
| 1313 | addInstr(env, AMD64Instr_Alu64R( |
| 1314 | Aalu_OR, AMD64RMI_Reg(lo8), hi8)); |
| 1315 | return hi8; |
| 1316 | } |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1317 | |
| 1318 | if (e->Iex.Binop.op == Iop_MullS32 |
| 1319 | || e->Iex.Binop.op == Iop_MullS16 |
| 1320 | || e->Iex.Binop.op == Iop_MullS8 |
| 1321 | || e->Iex.Binop.op == Iop_MullU32 |
| 1322 | || e->Iex.Binop.op == Iop_MullU16 |
| 1323 | || e->Iex.Binop.op == Iop_MullU8) { |
| 1324 | HReg a32 = newVRegI(env); |
| 1325 | HReg b32 = newVRegI(env); |
| 1326 | HReg a32s = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1327 | HReg b32s = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 1328 | Int shift = 0; |
| 1329 | AMD64ShiftOp shr_op = Ash_SHR; |
| 1330 | switch (e->Iex.Binop.op) { |
| 1331 | case Iop_MullS32: shr_op = Ash_SAR; shift = 32; break; |
| 1332 | case Iop_MullS16: shr_op = Ash_SAR; shift = 48; break; |
| 1333 | case Iop_MullS8: shr_op = Ash_SAR; shift = 56; break; |
| 1334 | case Iop_MullU32: shr_op = Ash_SHR; shift = 32; break; |
| 1335 | case Iop_MullU16: shr_op = Ash_SHR; shift = 48; break; |
| 1336 | case Iop_MullU8: shr_op = Ash_SHR; shift = 56; break; |
| 1337 | default: vassert(0); |
| 1338 | } |
| 1339 | |
| 1340 | addInstr(env, mk_iMOVsd_RR(a32s, a32)); |
| 1341 | addInstr(env, mk_iMOVsd_RR(b32s, b32)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1342 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, shift, a32)); |
| 1343 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, shift, b32)); |
| 1344 | addInstr(env, AMD64Instr_Sh64(shr_op, shift, a32)); |
| 1345 | addInstr(env, AMD64Instr_Sh64(shr_op, shift, b32)); |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1346 | addInstr(env, AMD64Instr_Alu64R(Aalu_MUL, AMD64RMI_Reg(a32), b32)); |
| 1347 | return b32; |
| 1348 | } |
| 1349 | |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 1350 | if (e->Iex.Binop.op == Iop_CmpF64) { |
| 1351 | HReg fL = iselDblExpr(env, e->Iex.Binop.arg1); |
| 1352 | HReg fR = iselDblExpr(env, e->Iex.Binop.arg2); |
| 1353 | HReg dst = newVRegI(env); |
| 1354 | addInstr(env, AMD64Instr_SseUComIS(8,fL,fR,dst)); |
| 1355 | /* Mask out irrelevant parts of the result so as to conform |
| 1356 | to the CmpF64 definition. */ |
| 1357 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, AMD64RMI_Imm(0x45), dst)); |
| 1358 | return dst; |
| 1359 | } |
| 1360 | |
sewardj | 6c299f3 | 2009-12-31 18:00:12 +0000 | [diff] [blame] | 1361 | if (e->Iex.Binop.op == Iop_F64toI32S |
| 1362 | || e->Iex.Binop.op == Iop_F64toI64S) { |
| 1363 | Int szD = e->Iex.Binop.op==Iop_F64toI32S ? 4 : 8; |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 1364 | HReg rf = iselDblExpr(env, e->Iex.Binop.arg2); |
| 1365 | HReg dst = newVRegI(env); |
| 1366 | set_SSE_rounding_mode( env, e->Iex.Binop.arg1 ); |
sewardj | 37d5257 | 2005-02-25 14:22:12 +0000 | [diff] [blame] | 1367 | addInstr(env, AMD64Instr_SseSF2SI( 8, szD, rf, dst )); |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 1368 | set_SSE_rounding_default(env); |
| 1369 | return dst; |
| 1370 | } |
| 1371 | |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1372 | break; |
| 1373 | } |
| 1374 | |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 1375 | /* --------- UNARY OP --------- */ |
| 1376 | case Iex_Unop: { |
sewardj | 7f039c4 | 2005-02-04 21:13:55 +0000 | [diff] [blame] | 1377 | |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1378 | /* 1Uto8(64to1(expr64)) */ |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1379 | { |
| 1380 | DEFINE_PATTERN( p_1Uto8_64to1, |
| 1381 | unop(Iop_1Uto8, unop(Iop_64to1, bind(0))) ); |
| 1382 | if (matchIRExpr(&mi,p_1Uto8_64to1,e)) { |
| 1383 | IRExpr* expr64 = mi.bindee[0]; |
| 1384 | HReg dst = newVRegI(env); |
| 1385 | HReg src = iselIntExpr_R(env, expr64); |
| 1386 | addInstr(env, mk_iMOVsd_RR(src,dst) ); |
| 1387 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, |
| 1388 | AMD64RMI_Imm(1), dst)); |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 1389 | return dst; |
| 1390 | } |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1391 | } |
| 1392 | |
| 1393 | /* 8Uto64(LDle(expr64)) */ |
| 1394 | { |
| 1395 | DEFINE_PATTERN(p_LDle8_then_8Uto64, |
| 1396 | unop(Iop_8Uto64, |
| 1397 | IRExpr_Load(Iend_LE,Ity_I8,bind(0))) ); |
| 1398 | if (matchIRExpr(&mi,p_LDle8_then_8Uto64,e)) { |
| 1399 | HReg dst = newVRegI(env); |
| 1400 | AMD64AMode* amode = iselIntExpr_AMode ( env, mi.bindee[0] ); |
| 1401 | addInstr(env, AMD64Instr_LoadEX(1,False,amode,dst)); |
| 1402 | return dst; |
| 1403 | } |
| 1404 | } |
| 1405 | |
| 1406 | /* 16Uto64(LDle(expr64)) */ |
| 1407 | { |
| 1408 | DEFINE_PATTERN(p_LDle16_then_16Uto64, |
| 1409 | unop(Iop_16Uto64, |
| 1410 | IRExpr_Load(Iend_LE,Ity_I16,bind(0))) ); |
| 1411 | if (matchIRExpr(&mi,p_LDle16_then_16Uto64,e)) { |
| 1412 | HReg dst = newVRegI(env); |
| 1413 | AMD64AMode* amode = iselIntExpr_AMode ( env, mi.bindee[0] ); |
| 1414 | addInstr(env, AMD64Instr_LoadEX(2,False,amode,dst)); |
| 1415 | return dst; |
| 1416 | } |
| 1417 | } |
| 1418 | |
sewardj | 9cc2bbf | 2011-06-05 17:56:03 +0000 | [diff] [blame] | 1419 | /* 32Uto64( Add32/Sub32/And32/Or32/Xor32(expr32, expr32) ) |
| 1420 | Use 32 bit arithmetic and let the default zero-extend rule |
| 1421 | do the 32Uto64 for free. */ |
| 1422 | if (e->Iex.Unop.op == Iop_32Uto64 && e->Iex.Unop.arg->tag == Iex_Binop) { |
| 1423 | IROp opi = e->Iex.Unop.arg->Iex.Binop.op; /* inner op */ |
| 1424 | IRExpr* argL = e->Iex.Unop.arg->Iex.Binop.arg1; |
| 1425 | IRExpr* argR = e->Iex.Unop.arg->Iex.Binop.arg2; |
| 1426 | AMD64AluOp aluOp = Aalu_INVALID; |
| 1427 | switch (opi) { |
| 1428 | case Iop_Add32: aluOp = Aalu_ADD; break; |
| 1429 | case Iop_Sub32: aluOp = Aalu_SUB; break; |
| 1430 | case Iop_And32: aluOp = Aalu_AND; break; |
| 1431 | case Iop_Or32: aluOp = Aalu_OR; break; |
| 1432 | case Iop_Xor32: aluOp = Aalu_XOR; break; |
| 1433 | default: break; |
| 1434 | } |
| 1435 | if (aluOp != Aalu_INVALID) { |
| 1436 | /* For commutative ops we assume any literal values are on |
| 1437 | the second operand. */ |
| 1438 | HReg dst = newVRegI(env); |
| 1439 | HReg reg = iselIntExpr_R(env, argL); |
| 1440 | AMD64RMI* rmi = iselIntExpr_RMI(env, argR); |
| 1441 | addInstr(env, mk_iMOVsd_RR(reg,dst)); |
| 1442 | addInstr(env, AMD64Instr_Alu32R(aluOp, rmi, dst)); |
| 1443 | return dst; |
| 1444 | } |
| 1445 | /* just fall through to normal handling for Iop_32Uto64 */ |
| 1446 | } |
| 1447 | |
| 1448 | /* Fallback cases */ |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1449 | switch (e->Iex.Unop.op) { |
| 1450 | case Iop_32Uto64: |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1451 | case Iop_32Sto64: { |
| 1452 | HReg dst = newVRegI(env); |
| 1453 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1454 | addInstr(env, AMD64Instr_MovxLQ(e->Iex.Unop.op == Iop_32Sto64, |
| 1455 | src, dst) ); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1456 | return dst; |
| 1457 | } |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 1458 | case Iop_128HIto64: { |
| 1459 | HReg rHi, rLo; |
| 1460 | iselInt128Expr(&rHi,&rLo, env, e->Iex.Unop.arg); |
| 1461 | return rHi; /* and abandon rLo */ |
| 1462 | } |
| 1463 | case Iop_128to64: { |
| 1464 | HReg rHi, rLo; |
| 1465 | iselInt128Expr(&rHi,&rLo, env, e->Iex.Unop.arg); |
| 1466 | return rLo; /* and abandon rHi */ |
| 1467 | } |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1468 | case Iop_8Uto16: |
sewardj | ec93f98 | 2005-06-21 13:51:18 +0000 | [diff] [blame] | 1469 | case Iop_8Uto32: |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1470 | case Iop_8Uto64: |
| 1471 | case Iop_16Uto64: |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1472 | case Iop_16Uto32: { |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1473 | HReg dst = newVRegI(env); |
| 1474 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
sewardj | 65b17c6 | 2005-05-02 15:52:44 +0000 | [diff] [blame] | 1475 | Bool srcIs16 = toBool( e->Iex.Unop.op==Iop_16Uto32 |
| 1476 | || e->Iex.Unop.op==Iop_16Uto64 ); |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1477 | UInt mask = srcIs16 ? 0xFFFF : 0xFF; |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1478 | addInstr(env, mk_iMOVsd_RR(src,dst) ); |
| 1479 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, |
| 1480 | AMD64RMI_Imm(mask), dst)); |
| 1481 | return dst; |
| 1482 | } |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1483 | case Iop_8Sto16: |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1484 | case Iop_8Sto64: |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1485 | case Iop_8Sto32: |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1486 | case Iop_16Sto32: |
| 1487 | case Iop_16Sto64: { |
| 1488 | HReg dst = newVRegI(env); |
| 1489 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
sewardj | 65b17c6 | 2005-05-02 15:52:44 +0000 | [diff] [blame] | 1490 | Bool srcIs16 = toBool( e->Iex.Unop.op==Iop_16Sto32 |
| 1491 | || e->Iex.Unop.op==Iop_16Sto64 ); |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1492 | UInt amt = srcIs16 ? 48 : 56; |
sewardj | 486074e | 2005-02-08 20:10:04 +0000 | [diff] [blame] | 1493 | addInstr(env, mk_iMOVsd_RR(src,dst) ); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1494 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, amt, dst)); |
| 1495 | addInstr(env, AMD64Instr_Sh64(Ash_SAR, amt, dst)); |
sewardj | 486074e | 2005-02-08 20:10:04 +0000 | [diff] [blame] | 1496 | return dst; |
| 1497 | } |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1498 | case Iop_Not8: |
| 1499 | case Iop_Not16: |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1500 | case Iop_Not32: |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 1501 | case Iop_Not64: { |
| 1502 | HReg dst = newVRegI(env); |
| 1503 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1504 | addInstr(env, mk_iMOVsd_RR(src,dst) ); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1505 | addInstr(env, AMD64Instr_Unary64(Aun_NOT,dst)); |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 1506 | return dst; |
| 1507 | } |
de | 5a70f5c | 2010-04-01 23:08:59 +0000 | [diff] [blame] | 1508 | case Iop_16HIto8: |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1509 | case Iop_32HIto16: |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1510 | case Iop_64HIto32: { |
| 1511 | HReg dst = newVRegI(env); |
| 1512 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1513 | Int shift = 0; |
| 1514 | switch (e->Iex.Unop.op) { |
sewardj | 9ba870d | 2010-04-02 11:29:23 +0000 | [diff] [blame] | 1515 | case Iop_16HIto8: shift = 8; break; |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1516 | case Iop_32HIto16: shift = 16; break; |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1517 | case Iop_64HIto32: shift = 32; break; |
| 1518 | default: vassert(0); |
| 1519 | } |
| 1520 | addInstr(env, mk_iMOVsd_RR(src,dst) ); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1521 | addInstr(env, AMD64Instr_Sh64(Ash_SHR, shift, dst)); |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1522 | return dst; |
| 1523 | } |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1524 | case Iop_1Uto64: |
sewardj | 0af46ab | 2005-04-26 01:52:29 +0000 | [diff] [blame] | 1525 | case Iop_1Uto32: |
sewardj | f53b735 | 2005-04-06 20:01:56 +0000 | [diff] [blame] | 1526 | case Iop_1Uto8: { |
| 1527 | HReg dst = newVRegI(env); |
| 1528 | AMD64CondCode cond = iselCondCode(env, e->Iex.Unop.arg); |
| 1529 | addInstr(env, AMD64Instr_Set64(cond,dst)); |
| 1530 | return dst; |
| 1531 | } |
sewardj | a64f8ad | 2005-04-24 00:26:37 +0000 | [diff] [blame] | 1532 | case Iop_1Sto8: |
sewardj | 478fe70 | 2005-04-23 01:15:47 +0000 | [diff] [blame] | 1533 | case Iop_1Sto16: |
| 1534 | case Iop_1Sto32: |
sewardj | 42322b5 | 2005-04-20 22:57:11 +0000 | [diff] [blame] | 1535 | case Iop_1Sto64: { |
| 1536 | /* could do better than this, but for now ... */ |
| 1537 | HReg dst = newVRegI(env); |
| 1538 | AMD64CondCode cond = iselCondCode(env, e->Iex.Unop.arg); |
| 1539 | addInstr(env, AMD64Instr_Set64(cond,dst)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1540 | addInstr(env, AMD64Instr_Sh64(Ash_SHL, 63, dst)); |
| 1541 | addInstr(env, AMD64Instr_Sh64(Ash_SAR, 63, dst)); |
sewardj | 42322b5 | 2005-04-20 22:57:11 +0000 | [diff] [blame] | 1542 | return dst; |
| 1543 | } |
sewardj | f53b735 | 2005-04-06 20:01:56 +0000 | [diff] [blame] | 1544 | case Iop_Ctz64: { |
| 1545 | /* Count trailing zeroes, implemented by amd64 'bsfq' */ |
| 1546 | HReg dst = newVRegI(env); |
| 1547 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1548 | addInstr(env, AMD64Instr_Bsfr64(True,src,dst)); |
| 1549 | return dst; |
| 1550 | } |
sewardj | 537cab0 | 2005-04-07 02:03:52 +0000 | [diff] [blame] | 1551 | case Iop_Clz64: { |
| 1552 | /* Count leading zeroes. Do 'bsrq' to establish the index |
| 1553 | of the highest set bit, and subtract that value from |
| 1554 | 63. */ |
| 1555 | HReg tmp = newVRegI(env); |
| 1556 | HReg dst = newVRegI(env); |
| 1557 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1558 | addInstr(env, AMD64Instr_Bsfr64(False,src,tmp)); |
| 1559 | addInstr(env, AMD64Instr_Alu64R(Aalu_MOV, |
| 1560 | AMD64RMI_Imm(63), dst)); |
| 1561 | addInstr(env, AMD64Instr_Alu64R(Aalu_SUB, |
| 1562 | AMD64RMI_Reg(tmp), dst)); |
| 1563 | return dst; |
| 1564 | } |
sewardj | eb17e49 | 2007-08-25 23:07:44 +0000 | [diff] [blame] | 1565 | |
| 1566 | case Iop_CmpwNEZ64: { |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1567 | HReg dst = newVRegI(env); |
sewardj | eb17e49 | 2007-08-25 23:07:44 +0000 | [diff] [blame] | 1568 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1569 | addInstr(env, mk_iMOVsd_RR(src,dst)); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 1570 | addInstr(env, AMD64Instr_Unary64(Aun_NEG,dst)); |
sewardj | eb17e49 | 2007-08-25 23:07:44 +0000 | [diff] [blame] | 1571 | addInstr(env, AMD64Instr_Alu64R(Aalu_OR, |
| 1572 | AMD64RMI_Reg(src), dst)); |
| 1573 | addInstr(env, AMD64Instr_Sh64(Ash_SAR, 63, dst)); |
| 1574 | return dst; |
| 1575 | } |
| 1576 | |
| 1577 | case Iop_CmpwNEZ32: { |
| 1578 | HReg src = newVRegI(env); |
| 1579 | HReg dst = newVRegI(env); |
| 1580 | HReg pre = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1581 | addInstr(env, mk_iMOVsd_RR(pre,src)); |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1582 | addInstr(env, AMD64Instr_MovxLQ(False, src, src)); |
sewardj | eb17e49 | 2007-08-25 23:07:44 +0000 | [diff] [blame] | 1583 | addInstr(env, mk_iMOVsd_RR(src,dst)); |
| 1584 | addInstr(env, AMD64Instr_Unary64(Aun_NEG,dst)); |
| 1585 | addInstr(env, AMD64Instr_Alu64R(Aalu_OR, |
| 1586 | AMD64RMI_Reg(src), dst)); |
| 1587 | addInstr(env, AMD64Instr_Sh64(Ash_SAR, 63, dst)); |
| 1588 | return dst; |
| 1589 | } |
| 1590 | |
| 1591 | case Iop_Left8: |
| 1592 | case Iop_Left16: |
| 1593 | case Iop_Left32: |
| 1594 | case Iop_Left64: { |
| 1595 | HReg dst = newVRegI(env); |
| 1596 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1597 | addInstr(env, mk_iMOVsd_RR(src, dst)); |
| 1598 | addInstr(env, AMD64Instr_Unary64(Aun_NEG, dst)); |
| 1599 | addInstr(env, AMD64Instr_Alu64R(Aalu_OR, AMD64RMI_Reg(src), dst)); |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1600 | return dst; |
| 1601 | } |
sewardj | 537cab0 | 2005-04-07 02:03:52 +0000 | [diff] [blame] | 1602 | |
sewardj | 478fe70 | 2005-04-23 01:15:47 +0000 | [diff] [blame] | 1603 | case Iop_V128to32: { |
| 1604 | HReg dst = newVRegI(env); |
| 1605 | HReg vec = iselVecExpr(env, e->Iex.Unop.arg); |
| 1606 | AMD64AMode* rsp_m16 = AMD64AMode_IR(-16, hregAMD64_RSP()); |
| 1607 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, vec, rsp_m16)); |
| 1608 | addInstr(env, AMD64Instr_LoadEX(4, False/*z-widen*/, rsp_m16, dst)); |
| 1609 | return dst; |
| 1610 | } |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 1611 | |
| 1612 | /* V128{HI}to64 */ |
| 1613 | case Iop_V128HIto64: |
| 1614 | case Iop_V128to64: { |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 1615 | HReg dst = newVRegI(env); |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 1616 | Int off = e->Iex.Unop.op==Iop_V128HIto64 ? -8 : -16; |
| 1617 | HReg rsp = hregAMD64_RSP(); |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 1618 | HReg vec = iselVecExpr(env, e->Iex.Unop.arg); |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 1619 | AMD64AMode* m16_rsp = AMD64AMode_IR(-16, rsp); |
| 1620 | AMD64AMode* off_rsp = AMD64AMode_IR(off, rsp); |
| 1621 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, |
| 1622 | 16, vec, m16_rsp)); |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 1623 | addInstr(env, AMD64Instr_Alu64R( Aalu_MOV, |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 1624 | AMD64RMI_Mem(off_rsp), dst )); |
| 1625 | return dst; |
| 1626 | } |
| 1627 | |
| 1628 | case Iop_V256to64_0: case Iop_V256to64_1: |
| 1629 | case Iop_V256to64_2: case Iop_V256to64_3: { |
| 1630 | HReg vHi, vLo, vec; |
| 1631 | iselDVecExpr(&vHi, &vLo, env, e->Iex.Unop.arg); |
| 1632 | /* Do the first part of the selection by deciding which of |
| 1633 | the 128 bit registers do look at, and second part using |
| 1634 | the same scheme as for V128{HI}to64 above. */ |
| 1635 | Int off = 0; |
| 1636 | switch (e->Iex.Unop.op) { |
| 1637 | case Iop_V256to64_0: vec = vLo; off = -16; break; |
| 1638 | case Iop_V256to64_1: vec = vLo; off = -8; break; |
| 1639 | case Iop_V256to64_2: vec = vHi; off = -16; break; |
| 1640 | case Iop_V256to64_3: vec = vHi; off = -8; break; |
| 1641 | default: vassert(0); |
| 1642 | } |
| 1643 | HReg dst = newVRegI(env); |
| 1644 | HReg rsp = hregAMD64_RSP(); |
| 1645 | AMD64AMode* m16_rsp = AMD64AMode_IR(-16, rsp); |
| 1646 | AMD64AMode* off_rsp = AMD64AMode_IR(off, rsp); |
| 1647 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, |
| 1648 | 16, vec, m16_rsp)); |
| 1649 | addInstr(env, AMD64Instr_Alu64R( Aalu_MOV, |
| 1650 | AMD64RMI_Mem(off_rsp), dst )); |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 1651 | return dst; |
| 1652 | } |
| 1653 | |
sewardj | 924215b | 2005-03-26 21:50:31 +0000 | [diff] [blame] | 1654 | /* ReinterpF64asI64(e) */ |
| 1655 | /* Given an IEEE754 double, produce an I64 with the same bit |
| 1656 | pattern. */ |
| 1657 | case Iop_ReinterpF64asI64: { |
| 1658 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
| 1659 | HReg dst = newVRegI(env); |
| 1660 | HReg src = iselDblExpr(env, e->Iex.Unop.arg); |
| 1661 | /* paranoia */ |
| 1662 | set_SSE_rounding_default(env); |
| 1663 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 8, src, m8_rsp)); |
| 1664 | addInstr(env, AMD64Instr_Alu64R( |
| 1665 | Aalu_MOV, AMD64RMI_Mem(m8_rsp), dst)); |
| 1666 | return dst; |
| 1667 | } |
| 1668 | |
sewardj | 7950111 | 2008-07-29 09:48:26 +0000 | [diff] [blame] | 1669 | /* ReinterpF32asI32(e) */ |
| 1670 | /* Given an IEEE754 single, produce an I64 with the same bit |
| 1671 | pattern in the lower half. */ |
| 1672 | case Iop_ReinterpF32asI32: { |
| 1673 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
| 1674 | HReg dst = newVRegI(env); |
| 1675 | HReg src = iselFltExpr(env, e->Iex.Unop.arg); |
| 1676 | /* paranoia */ |
| 1677 | set_SSE_rounding_default(env); |
| 1678 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 4, src, m8_rsp)); |
| 1679 | addInstr(env, AMD64Instr_LoadEX(4, False/*unsigned*/, m8_rsp, dst )); |
| 1680 | return dst; |
| 1681 | } |
| 1682 | |
sewardj | 85520e4 | 2005-02-19 15:22:38 +0000 | [diff] [blame] | 1683 | case Iop_16to8: |
sewardj | a6b93d1 | 2005-02-17 09:28:28 +0000 | [diff] [blame] | 1684 | case Iop_32to8: |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1685 | case Iop_64to8: |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 1686 | case Iop_32to16: |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 1687 | case Iop_64to16: |
sewardj | 486074e | 2005-02-08 20:10:04 +0000 | [diff] [blame] | 1688 | case Iop_64to32: |
| 1689 | /* These are no-ops. */ |
| 1690 | return iselIntExpr_R(env, e->Iex.Unop.arg); |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 1691 | |
sewardj | e13074c | 2012-11-08 10:57:08 +0000 | [diff] [blame] | 1692 | case Iop_GetMSBs8x8: { |
| 1693 | /* Note: the following assumes the helper is of |
| 1694 | signature |
| 1695 | UInt fn ( ULong ), and is not a regparm fn. |
| 1696 | */ |
| 1697 | HReg dst = newVRegI(env); |
| 1698 | HReg arg = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1699 | fn = (HWord)h_generic_calc_GetMSBs8x8; |
| 1700 | addInstr(env, mk_iMOVsd_RR(arg, hregAMD64_RDI()) ); |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 1701 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 1702 | 1, mk_RetLoc_simple(RLPri_Int) )); |
sewardj | e13074c | 2012-11-08 10:57:08 +0000 | [diff] [blame] | 1703 | /* MovxLQ is not exactly the right thing here. We just |
| 1704 | need to get the bottom 8 bits of RAX into dst, and zero |
| 1705 | out everything else. Assuming that the helper returns |
| 1706 | a UInt with the top 24 bits zeroed out, it'll do, |
| 1707 | though. */ |
| 1708 | addInstr(env, AMD64Instr_MovxLQ(False, hregAMD64_RAX(), dst)); |
| 1709 | return dst; |
| 1710 | } |
| 1711 | |
sewardj | 78a2059 | 2012-12-13 18:29:56 +0000 | [diff] [blame] | 1712 | case Iop_GetMSBs8x16: { |
| 1713 | /* Note: the following assumes the helper is of signature |
| 1714 | UInt fn ( ULong w64hi, ULong w64Lo ), |
| 1715 | and is not a regparm fn. */ |
| 1716 | HReg dst = newVRegI(env); |
| 1717 | HReg vec = iselVecExpr(env, e->Iex.Unop.arg); |
| 1718 | HReg rsp = hregAMD64_RSP(); |
| 1719 | fn = (HWord)h_generic_calc_GetMSBs8x16; |
| 1720 | AMD64AMode* m8_rsp = AMD64AMode_IR( -8, rsp); |
| 1721 | AMD64AMode* m16_rsp = AMD64AMode_IR(-16, rsp); |
| 1722 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, |
| 1723 | 16, vec, m16_rsp)); |
| 1724 | /* hi 64 bits into RDI -- the first arg */ |
| 1725 | addInstr(env, AMD64Instr_Alu64R( Aalu_MOV, |
| 1726 | AMD64RMI_Mem(m8_rsp), |
| 1727 | hregAMD64_RDI() )); /* 1st arg */ |
| 1728 | /* lo 64 bits into RSI -- the 2nd arg */ |
| 1729 | addInstr(env, AMD64Instr_Alu64R( Aalu_MOV, |
| 1730 | AMD64RMI_Mem(m16_rsp), |
| 1731 | hregAMD64_RSI() )); /* 2nd arg */ |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 1732 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 1733 | 2, mk_RetLoc_simple(RLPri_Int) )); |
sewardj | 78a2059 | 2012-12-13 18:29:56 +0000 | [diff] [blame] | 1734 | /* MovxLQ is not exactly the right thing here. We just |
sewardj | 9213c61 | 2012-12-19 08:39:11 +0000 | [diff] [blame] | 1735 | need to get the bottom 16 bits of RAX into dst, and zero |
sewardj | 78a2059 | 2012-12-13 18:29:56 +0000 | [diff] [blame] | 1736 | out everything else. Assuming that the helper returns |
sewardj | 9213c61 | 2012-12-19 08:39:11 +0000 | [diff] [blame] | 1737 | a UInt with the top 16 bits zeroed out, it'll do, |
sewardj | 78a2059 | 2012-12-13 18:29:56 +0000 | [diff] [blame] | 1738 | though. */ |
| 1739 | addInstr(env, AMD64Instr_MovxLQ(False, hregAMD64_RAX(), dst)); |
| 1740 | return dst; |
| 1741 | } |
| 1742 | |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 1743 | default: |
| 1744 | break; |
| 1745 | } |
sewardj | e790566 | 2005-05-09 18:15:21 +0000 | [diff] [blame] | 1746 | |
| 1747 | /* Deal with unary 64-bit SIMD ops. */ |
| 1748 | switch (e->Iex.Unop.op) { |
| 1749 | case Iop_CmpNEZ32x2: |
| 1750 | fn = (HWord)h_generic_calc_CmpNEZ32x2; break; |
| 1751 | case Iop_CmpNEZ16x4: |
| 1752 | fn = (HWord)h_generic_calc_CmpNEZ16x4; break; |
| 1753 | case Iop_CmpNEZ8x8: |
| 1754 | fn = (HWord)h_generic_calc_CmpNEZ8x8; break; |
| 1755 | default: |
| 1756 | fn = (HWord)0; break; |
| 1757 | } |
| 1758 | if (fn != (HWord)0) { |
| 1759 | /* Note: the following assumes all helpers are of |
| 1760 | signature |
| 1761 | ULong fn ( ULong ), and they are |
| 1762 | not marked as regparm functions. |
| 1763 | */ |
| 1764 | HReg dst = newVRegI(env); |
| 1765 | HReg arg = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 1766 | addInstr(env, mk_iMOVsd_RR(arg, hregAMD64_RDI()) ); |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 1767 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, 1, |
| 1768 | mk_RetLoc_simple(RLPri_Int) )); |
sewardj | e790566 | 2005-05-09 18:15:21 +0000 | [diff] [blame] | 1769 | addInstr(env, mk_iMOVsd_RR(hregAMD64_RAX(), dst)); |
| 1770 | return dst; |
| 1771 | } |
| 1772 | |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 1773 | break; |
| 1774 | } |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1775 | |
| 1776 | /* --------- GET --------- */ |
| 1777 | case Iex_Get: { |
| 1778 | if (ty == Ity_I64) { |
| 1779 | HReg dst = newVRegI(env); |
| 1780 | addInstr(env, AMD64Instr_Alu64R( |
| 1781 | Aalu_MOV, |
| 1782 | AMD64RMI_Mem( |
| 1783 | AMD64AMode_IR(e->Iex.Get.offset, |
| 1784 | hregAMD64_RBP())), |
| 1785 | dst)); |
| 1786 | return dst; |
| 1787 | } |
| 1788 | if (ty == Ity_I8 || ty == Ity_I16 || ty == Ity_I32) { |
| 1789 | HReg dst = newVRegI(env); |
| 1790 | addInstr(env, AMD64Instr_LoadEX( |
sewardj | 1e49935 | 2005-03-23 03:02:50 +0000 | [diff] [blame] | 1791 | toUChar(ty==Ity_I8 ? 1 : (ty==Ity_I16 ? 2 : 4)), |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1792 | False, |
| 1793 | AMD64AMode_IR(e->Iex.Get.offset,hregAMD64_RBP()), |
| 1794 | dst)); |
| 1795 | return dst; |
| 1796 | } |
| 1797 | break; |
| 1798 | } |
| 1799 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 1800 | case Iex_GetI: { |
| 1801 | AMD64AMode* am |
| 1802 | = genGuestArrayOffset( |
| 1803 | env, e->Iex.GetI.descr, |
| 1804 | e->Iex.GetI.ix, e->Iex.GetI.bias ); |
| 1805 | HReg dst = newVRegI(env); |
| 1806 | if (ty == Ity_I8) { |
| 1807 | addInstr(env, AMD64Instr_LoadEX( 1, False, am, dst )); |
| 1808 | return dst; |
| 1809 | } |
sewardj | 1e015d8 | 2005-04-23 23:41:46 +0000 | [diff] [blame] | 1810 | if (ty == Ity_I64) { |
| 1811 | addInstr(env, AMD64Instr_Alu64R( Aalu_MOV, AMD64RMI_Mem(am), dst )); |
| 1812 | return dst; |
| 1813 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 1814 | break; |
| 1815 | } |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1816 | |
| 1817 | /* --------- CCALL --------- */ |
| 1818 | case Iex_CCall: { |
| 1819 | HReg dst = newVRegI(env); |
sewardj | 7f039c4 | 2005-02-04 21:13:55 +0000 | [diff] [blame] | 1820 | vassert(ty == e->Iex.CCall.retty); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1821 | |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 1822 | /* be very restrictive for now. Only 64-bit ints allowed for |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 1823 | args, and 64 or 32 bits for return type. */ |
sewardj | e8aaa87 | 2005-07-07 13:12:04 +0000 | [diff] [blame] | 1824 | if (e->Iex.CCall.retty != Ity_I64 && e->Iex.CCall.retty != Ity_I32) |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1825 | goto irreducible; |
| 1826 | |
sewardj | 7f039c4 | 2005-02-04 21:13:55 +0000 | [diff] [blame] | 1827 | /* Marshal args, do the call. */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 1828 | UInt addToSp = 0; |
| 1829 | RetLoc rloc = mk_RetLoc_INVALID(); |
| 1830 | doHelperCall( &addToSp, &rloc, env, NULL/*guard*/, |
| 1831 | e->Iex.CCall.cee, e->Iex.CCall.retty, e->Iex.CCall.args ); |
| 1832 | vassert(is_sane_RetLoc(rloc)); |
| 1833 | vassert(rloc.pri == RLPri_Int); |
| 1834 | vassert(addToSp == 0); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1835 | |
sewardj | e8aaa87 | 2005-07-07 13:12:04 +0000 | [diff] [blame] | 1836 | /* Move to dst, and zero out the top 32 bits if the result type is |
| 1837 | Ity_I32. Probably overkill, but still .. */ |
| 1838 | if (e->Iex.CCall.retty == Ity_I64) |
| 1839 | addInstr(env, mk_iMOVsd_RR(hregAMD64_RAX(), dst)); |
| 1840 | else |
sewardj | ca257bc | 2010-09-08 08:34:52 +0000 | [diff] [blame] | 1841 | addInstr(env, AMD64Instr_MovxLQ(False, hregAMD64_RAX(), dst)); |
sewardj | e8aaa87 | 2005-07-07 13:12:04 +0000 | [diff] [blame] | 1842 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1843 | return dst; |
| 1844 | } |
| 1845 | |
sewardj | 7f039c4 | 2005-02-04 21:13:55 +0000 | [diff] [blame] | 1846 | /* --------- LITERAL --------- */ |
| 1847 | /* 64/32/16/8-bit literals */ |
| 1848 | case Iex_Const: |
| 1849 | if (ty == Ity_I64) { |
| 1850 | HReg r = newVRegI(env); |
| 1851 | addInstr(env, AMD64Instr_Imm64(e->Iex.Const.con->Ico.U64, r)); |
| 1852 | return r; |
| 1853 | } else { |
| 1854 | AMD64RMI* rmi = iselIntExpr_RMI ( env, e ); |
| 1855 | HReg r = newVRegI(env); |
| 1856 | addInstr(env, AMD64Instr_Alu64R(Aalu_MOV, rmi, r)); |
| 1857 | return r; |
| 1858 | } |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1859 | |
| 1860 | /* --------- MULTIPLEX --------- */ |
florian | 99dd03e | 2013-01-29 03:56:06 +0000 | [diff] [blame] | 1861 | case Iex_ITE: { // VFD |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 1862 | if ((ty == Ity_I64 || ty == Ity_I32 || ty == Ity_I16 || ty == Ity_I8) |
florian | 99dd03e | 2013-01-29 03:56:06 +0000 | [diff] [blame] | 1863 | && typeOfIRExpr(env->type_env,e->Iex.ITE.cond) == Ity_I1) { |
| 1864 | HReg r1 = iselIntExpr_R(env, e->Iex.ITE.iftrue); |
| 1865 | AMD64RM* r0 = iselIntExpr_RM(env, e->Iex.ITE.iffalse); |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 1866 | HReg dst = newVRegI(env); |
florian | 99dd03e | 2013-01-29 03:56:06 +0000 | [diff] [blame] | 1867 | addInstr(env, mk_iMOVsd_RR(r1,dst)); |
| 1868 | AMD64CondCode cc = iselCondCode(env, e->Iex.ITE.cond); |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 1869 | addInstr(env, AMD64Instr_CMov64(cc ^ 1, r0, dst)); |
| 1870 | return dst; |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1871 | } |
| 1872 | break; |
| 1873 | } |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1874 | |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 1875 | /* --------- TERNARY OP --------- */ |
| 1876 | case Iex_Triop: { |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 1877 | IRTriop *triop = e->Iex.Triop.details; |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 1878 | /* C3210 flags following FPU partial remainder (fprem), both |
| 1879 | IEEE compliant (PREM1) and non-IEEE compliant (PREM). */ |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 1880 | if (triop->op == Iop_PRemC3210F64 |
| 1881 | || triop->op == Iop_PRem1C3210F64) { |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 1882 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 1883 | HReg arg1 = iselDblExpr(env, triop->arg2); |
| 1884 | HReg arg2 = iselDblExpr(env, triop->arg3); |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 1885 | HReg dst = newVRegI(env); |
| 1886 | addInstr(env, AMD64Instr_A87Free(2)); |
| 1887 | |
| 1888 | /* one arg -> top of x87 stack */ |
| 1889 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 8, arg2, m8_rsp)); |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 1890 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, True/*push*/, 8)); |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 1891 | |
| 1892 | /* other arg -> top of x87 stack */ |
| 1893 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 8, arg1, m8_rsp)); |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 1894 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, True/*push*/, 8)); |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 1895 | |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 1896 | switch (triop->op) { |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 1897 | case Iop_PRemC3210F64: |
| 1898 | addInstr(env, AMD64Instr_A87FpOp(Afp_PREM)); |
| 1899 | break; |
sewardj | 4970e4e | 2008-10-11 10:07:55 +0000 | [diff] [blame] | 1900 | case Iop_PRem1C3210F64: |
| 1901 | addInstr(env, AMD64Instr_A87FpOp(Afp_PREM1)); |
| 1902 | break; |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 1903 | default: |
| 1904 | vassert(0); |
| 1905 | } |
| 1906 | /* Ignore the result, and instead make off with the FPU's |
| 1907 | C3210 flags (in the status word). */ |
| 1908 | addInstr(env, AMD64Instr_A87StSW(m8_rsp)); |
| 1909 | addInstr(env, AMD64Instr_Alu64R(Aalu_MOV,AMD64RMI_Mem(m8_rsp),dst)); |
| 1910 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND,AMD64RMI_Imm(0x4700),dst)); |
| 1911 | return dst; |
| 1912 | } |
| 1913 | break; |
| 1914 | } |
| 1915 | |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1916 | default: |
| 1917 | break; |
| 1918 | } /* switch (e->tag) */ |
| 1919 | |
| 1920 | /* We get here if no pattern matched. */ |
| 1921 | irreducible: |
| 1922 | ppIRExpr(e); |
| 1923 | vpanic("iselIntExpr_R(amd64): cannot reduce tree"); |
| 1924 | } |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 1925 | |
| 1926 | |
| 1927 | /*---------------------------------------------------------*/ |
| 1928 | /*--- ISEL: Integer expression auxiliaries ---*/ |
| 1929 | /*---------------------------------------------------------*/ |
| 1930 | |
| 1931 | /* --------------------- AMODEs --------------------- */ |
| 1932 | |
| 1933 | /* Return an AMode which computes the value of the specified |
| 1934 | expression, possibly also adding insns to the code list as a |
| 1935 | result. The expression may only be a 32-bit one. |
| 1936 | */ |
| 1937 | |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1938 | static AMD64AMode* iselIntExpr_AMode ( ISelEnv* env, IRExpr* e ) |
| 1939 | { |
| 1940 | AMD64AMode* am = iselIntExpr_AMode_wrk(env, e); |
| 1941 | vassert(sane_AMode(am)); |
| 1942 | return am; |
| 1943 | } |
| 1944 | |
| 1945 | /* DO NOT CALL THIS DIRECTLY ! */ |
| 1946 | static AMD64AMode* iselIntExpr_AMode_wrk ( ISelEnv* env, IRExpr* e ) |
| 1947 | { |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1948 | MatchInfo mi; |
| 1949 | DECLARE_PATTERN(p_complex); |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1950 | IRType ty = typeOfIRExpr(env->type_env,e); |
| 1951 | vassert(ty == Ity_I64); |
| 1952 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1953 | /* Add64( Add64(expr1, Shl64(expr2, imm8)), simm32 ) */ |
| 1954 | /* bind0 bind1 bind2 bind3 */ |
| 1955 | DEFINE_PATTERN(p_complex, |
| 1956 | binop( Iop_Add64, |
| 1957 | binop( Iop_Add64, |
| 1958 | bind(0), |
| 1959 | binop(Iop_Shl64, bind(1), bind(2)) |
| 1960 | ), |
| 1961 | bind(3) |
| 1962 | ) |
| 1963 | ); |
| 1964 | if (matchIRExpr(&mi, p_complex, e)) { |
| 1965 | IRExpr* expr1 = mi.bindee[0]; |
| 1966 | IRExpr* expr2 = mi.bindee[1]; |
| 1967 | IRExpr* imm8 = mi.bindee[2]; |
| 1968 | IRExpr* simm32 = mi.bindee[3]; |
| 1969 | if (imm8->tag == Iex_Const |
| 1970 | && imm8->Iex.Const.con->tag == Ico_U8 |
| 1971 | && imm8->Iex.Const.con->Ico.U8 < 4 |
| 1972 | /* imm8 is OK, now check simm32 */ |
| 1973 | && simm32->tag == Iex_Const |
| 1974 | && simm32->Iex.Const.con->tag == Ico_U64 |
| 1975 | && fitsIn32Bits(simm32->Iex.Const.con->Ico.U64)) { |
| 1976 | UInt shift = imm8->Iex.Const.con->Ico.U8; |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 1977 | UInt offset = toUInt(simm32->Iex.Const.con->Ico.U64); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 1978 | HReg r1 = iselIntExpr_R(env, expr1); |
| 1979 | HReg r2 = iselIntExpr_R(env, expr2); |
| 1980 | vassert(shift == 0 || shift == 1 || shift == 2 || shift == 3); |
| 1981 | return AMD64AMode_IRRS(offset, r1, r2, shift); |
| 1982 | } |
| 1983 | } |
| 1984 | |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 1985 | /* Add64(expr1, Shl64(expr2, imm)) */ |
| 1986 | if (e->tag == Iex_Binop |
| 1987 | && e->Iex.Binop.op == Iop_Add64 |
| 1988 | && e->Iex.Binop.arg2->tag == Iex_Binop |
| 1989 | && e->Iex.Binop.arg2->Iex.Binop.op == Iop_Shl64 |
| 1990 | && e->Iex.Binop.arg2->Iex.Binop.arg2->tag == Iex_Const |
| 1991 | && e->Iex.Binop.arg2->Iex.Binop.arg2->Iex.Const.con->tag == Ico_U8) { |
| 1992 | UInt shift = e->Iex.Binop.arg2->Iex.Binop.arg2->Iex.Const.con->Ico.U8; |
| 1993 | if (shift == 1 || shift == 2 || shift == 3) { |
| 1994 | HReg r1 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 1995 | HReg r2 = iselIntExpr_R(env, e->Iex.Binop.arg2->Iex.Binop.arg1 ); |
| 1996 | return AMD64AMode_IRRS(0, r1, r2, shift); |
| 1997 | } |
| 1998 | } |
| 1999 | |
| 2000 | /* Add64(expr,i) */ |
| 2001 | if (e->tag == Iex_Binop |
| 2002 | && e->Iex.Binop.op == Iop_Add64 |
| 2003 | && e->Iex.Binop.arg2->tag == Iex_Const |
| 2004 | && e->Iex.Binop.arg2->Iex.Const.con->tag == Ico_U64 |
| 2005 | && fitsIn32Bits(e->Iex.Binop.arg2->Iex.Const.con->Ico.U64)) { |
| 2006 | HReg r1 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 2007 | return AMD64AMode_IR( |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 2008 | toUInt(e->Iex.Binop.arg2->Iex.Const.con->Ico.U64), |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 2009 | r1 |
| 2010 | ); |
| 2011 | } |
| 2012 | |
| 2013 | /* Doesn't match anything in particular. Generate it into |
| 2014 | a register and use that. */ |
| 2015 | { |
| 2016 | HReg r1 = iselIntExpr_R(env, e); |
| 2017 | return AMD64AMode_IR(0, r1); |
| 2018 | } |
| 2019 | } |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 2020 | |
| 2021 | |
| 2022 | /* --------------------- RMIs --------------------- */ |
| 2023 | |
| 2024 | /* Similarly, calculate an expression into an X86RMI operand. As with |
| 2025 | iselIntExpr_R, the expression can have type 32, 16 or 8 bits. */ |
| 2026 | |
| 2027 | static AMD64RMI* iselIntExpr_RMI ( ISelEnv* env, IRExpr* e ) |
| 2028 | { |
| 2029 | AMD64RMI* rmi = iselIntExpr_RMI_wrk(env, e); |
| 2030 | /* sanity checks ... */ |
| 2031 | switch (rmi->tag) { |
| 2032 | case Armi_Imm: |
| 2033 | return rmi; |
| 2034 | case Armi_Reg: |
| 2035 | vassert(hregClass(rmi->Armi.Reg.reg) == HRcInt64); |
| 2036 | vassert(hregIsVirtual(rmi->Armi.Reg.reg)); |
| 2037 | return rmi; |
| 2038 | case Armi_Mem: |
| 2039 | vassert(sane_AMode(rmi->Armi.Mem.am)); |
| 2040 | return rmi; |
| 2041 | default: |
| 2042 | vpanic("iselIntExpr_RMI: unknown amd64 RMI tag"); |
| 2043 | } |
| 2044 | } |
| 2045 | |
| 2046 | /* DO NOT CALL THIS DIRECTLY ! */ |
| 2047 | static AMD64RMI* iselIntExpr_RMI_wrk ( ISelEnv* env, IRExpr* e ) |
| 2048 | { |
| 2049 | IRType ty = typeOfIRExpr(env->type_env,e); |
| 2050 | vassert(ty == Ity_I64 || ty == Ity_I32 |
| 2051 | || ty == Ity_I16 || ty == Ity_I8); |
| 2052 | |
| 2053 | /* special case: immediate 64/32/16/8 */ |
| 2054 | if (e->tag == Iex_Const) { |
| 2055 | switch (e->Iex.Const.con->tag) { |
| 2056 | case Ico_U64: |
| 2057 | if (fitsIn32Bits(e->Iex.Const.con->Ico.U64)) { |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 2058 | return AMD64RMI_Imm(toUInt(e->Iex.Const.con->Ico.U64)); |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 2059 | } |
| 2060 | break; |
| 2061 | case Ico_U32: |
| 2062 | return AMD64RMI_Imm(e->Iex.Const.con->Ico.U32); break; |
| 2063 | case Ico_U16: |
| 2064 | return AMD64RMI_Imm(0xFFFF & e->Iex.Const.con->Ico.U16); break; |
| 2065 | case Ico_U8: |
| 2066 | return AMD64RMI_Imm(0xFF & e->Iex.Const.con->Ico.U8); break; |
| 2067 | default: |
| 2068 | vpanic("iselIntExpr_RMI.Iex_Const(amd64)"); |
| 2069 | } |
| 2070 | } |
| 2071 | |
| 2072 | /* special case: 64-bit GET */ |
| 2073 | if (e->tag == Iex_Get && ty == Ity_I64) { |
| 2074 | return AMD64RMI_Mem(AMD64AMode_IR(e->Iex.Get.offset, |
| 2075 | hregAMD64_RBP())); |
| 2076 | } |
| 2077 | |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 2078 | /* special case: 64-bit load from memory */ |
sewardj | e9d8a26 | 2009-07-01 08:06:34 +0000 | [diff] [blame] | 2079 | if (e->tag == Iex_Load && ty == Ity_I64 |
sewardj | e768e92 | 2009-11-26 17:17:37 +0000 | [diff] [blame] | 2080 | && e->Iex.Load.end == Iend_LE) { |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 2081 | AMD64AMode* am = iselIntExpr_AMode(env, e->Iex.Load.addr); |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 2082 | return AMD64RMI_Mem(am); |
| 2083 | } |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 2084 | |
| 2085 | /* default case: calculate into a register and return that */ |
sewardj | 8258a8c | 2005-02-02 03:11:24 +0000 | [diff] [blame] | 2086 | { |
| 2087 | HReg r = iselIntExpr_R ( env, e ); |
| 2088 | return AMD64RMI_Reg(r); |
| 2089 | } |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 2090 | } |
| 2091 | |
| 2092 | |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 2093 | /* --------------------- RIs --------------------- */ |
| 2094 | |
| 2095 | /* Calculate an expression into an AMD64RI operand. As with |
| 2096 | iselIntExpr_R, the expression can have type 64, 32, 16 or 8 |
| 2097 | bits. */ |
| 2098 | |
| 2099 | static AMD64RI* iselIntExpr_RI ( ISelEnv* env, IRExpr* e ) |
| 2100 | { |
| 2101 | AMD64RI* ri = iselIntExpr_RI_wrk(env, e); |
| 2102 | /* sanity checks ... */ |
| 2103 | switch (ri->tag) { |
| 2104 | case Ari_Imm: |
| 2105 | return ri; |
sewardj | 80d6e6d | 2008-05-28 09:40:29 +0000 | [diff] [blame] | 2106 | case Ari_Reg: |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 2107 | vassert(hregClass(ri->Ari.Reg.reg) == HRcInt64); |
| 2108 | vassert(hregIsVirtual(ri->Ari.Reg.reg)); |
| 2109 | return ri; |
| 2110 | default: |
| 2111 | vpanic("iselIntExpr_RI: unknown amd64 RI tag"); |
| 2112 | } |
| 2113 | } |
| 2114 | |
| 2115 | /* DO NOT CALL THIS DIRECTLY ! */ |
| 2116 | static AMD64RI* iselIntExpr_RI_wrk ( ISelEnv* env, IRExpr* e ) |
| 2117 | { |
| 2118 | IRType ty = typeOfIRExpr(env->type_env,e); |
| 2119 | vassert(ty == Ity_I64 || ty == Ity_I32 |
| 2120 | || ty == Ity_I16 || ty == Ity_I8); |
| 2121 | |
| 2122 | /* special case: immediate */ |
| 2123 | if (e->tag == Iex_Const) { |
| 2124 | switch (e->Iex.Const.con->tag) { |
| 2125 | case Ico_U64: |
| 2126 | if (fitsIn32Bits(e->Iex.Const.con->Ico.U64)) { |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 2127 | return AMD64RI_Imm(toUInt(e->Iex.Const.con->Ico.U64)); |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 2128 | } |
| 2129 | break; |
| 2130 | case Ico_U32: |
| 2131 | return AMD64RI_Imm(e->Iex.Const.con->Ico.U32); |
| 2132 | case Ico_U16: |
| 2133 | return AMD64RI_Imm(0xFFFF & e->Iex.Const.con->Ico.U16); |
| 2134 | case Ico_U8: |
| 2135 | return AMD64RI_Imm(0xFF & e->Iex.Const.con->Ico.U8); |
| 2136 | default: |
| 2137 | vpanic("iselIntExpr_RMI.Iex_Const(amd64)"); |
| 2138 | } |
| 2139 | } |
| 2140 | |
| 2141 | /* default case: calculate into a register and return that */ |
| 2142 | { |
| 2143 | HReg r = iselIntExpr_R ( env, e ); |
| 2144 | return AMD64RI_Reg(r); |
| 2145 | } |
| 2146 | } |
| 2147 | |
| 2148 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 2149 | /* --------------------- RMs --------------------- */ |
| 2150 | |
| 2151 | /* Similarly, calculate an expression into an AMD64RM operand. As |
| 2152 | with iselIntExpr_R, the expression can have type 64, 32, 16 or 8 |
| 2153 | bits. */ |
| 2154 | |
| 2155 | static AMD64RM* iselIntExpr_RM ( ISelEnv* env, IRExpr* e ) |
| 2156 | { |
| 2157 | AMD64RM* rm = iselIntExpr_RM_wrk(env, e); |
| 2158 | /* sanity checks ... */ |
| 2159 | switch (rm->tag) { |
| 2160 | case Arm_Reg: |
| 2161 | vassert(hregClass(rm->Arm.Reg.reg) == HRcInt64); |
| 2162 | vassert(hregIsVirtual(rm->Arm.Reg.reg)); |
| 2163 | return rm; |
| 2164 | case Arm_Mem: |
| 2165 | vassert(sane_AMode(rm->Arm.Mem.am)); |
| 2166 | return rm; |
| 2167 | default: |
| 2168 | vpanic("iselIntExpr_RM: unknown amd64 RM tag"); |
| 2169 | } |
| 2170 | } |
| 2171 | |
| 2172 | /* DO NOT CALL THIS DIRECTLY ! */ |
| 2173 | static AMD64RM* iselIntExpr_RM_wrk ( ISelEnv* env, IRExpr* e ) |
| 2174 | { |
| 2175 | IRType ty = typeOfIRExpr(env->type_env,e); |
| 2176 | vassert(ty == Ity_I64 || ty == Ity_I32 || ty == Ity_I16 || ty == Ity_I8); |
| 2177 | |
| 2178 | /* special case: 64-bit GET */ |
| 2179 | if (e->tag == Iex_Get && ty == Ity_I64) { |
| 2180 | return AMD64RM_Mem(AMD64AMode_IR(e->Iex.Get.offset, |
| 2181 | hregAMD64_RBP())); |
| 2182 | } |
| 2183 | |
| 2184 | /* special case: load from memory */ |
| 2185 | |
| 2186 | /* default case: calculate into a register and return that */ |
| 2187 | { |
| 2188 | HReg r = iselIntExpr_R ( env, e ); |
| 2189 | return AMD64RM_Reg(r); |
| 2190 | } |
| 2191 | } |
| 2192 | |
| 2193 | |
| 2194 | /* --------------------- CONDCODE --------------------- */ |
| 2195 | |
| 2196 | /* Generate code to evaluated a bit-typed expression, returning the |
| 2197 | condition code which would correspond when the expression would |
| 2198 | notionally have returned 1. */ |
| 2199 | |
| 2200 | static AMD64CondCode iselCondCode ( ISelEnv* env, IRExpr* e ) |
| 2201 | { |
| 2202 | /* Uh, there's nothing we can sanity check here, unfortunately. */ |
| 2203 | return iselCondCode_wrk(env,e); |
| 2204 | } |
| 2205 | |
| 2206 | /* DO NOT CALL THIS DIRECTLY ! */ |
| 2207 | static AMD64CondCode iselCondCode_wrk ( ISelEnv* env, IRExpr* e ) |
| 2208 | { |
sewardj | f8c37f7 | 2005-02-07 18:55:29 +0000 | [diff] [blame] | 2209 | MatchInfo mi; |
sewardj | 0af46ab | 2005-04-26 01:52:29 +0000 | [diff] [blame] | 2210 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 2211 | vassert(e); |
| 2212 | vassert(typeOfIRExpr(env->type_env,e) == Ity_I1); |
| 2213 | |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2214 | /* var */ |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 2215 | if (e->tag == Iex_RdTmp) { |
| 2216 | HReg r64 = lookupIRTemp(env, e->Iex.RdTmp.tmp); |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2217 | HReg dst = newVRegI(env); |
| 2218 | addInstr(env, mk_iMOVsd_RR(r64,dst)); |
| 2219 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND,AMD64RMI_Imm(1),dst)); |
| 2220 | return Acc_NZ; |
| 2221 | } |
| 2222 | |
sewardj | 109e935 | 2005-07-19 08:42:56 +0000 | [diff] [blame] | 2223 | /* Constant 1:Bit */ |
| 2224 | if (e->tag == Iex_Const) { |
| 2225 | HReg r; |
| 2226 | vassert(e->Iex.Const.con->tag == Ico_U1); |
| 2227 | vassert(e->Iex.Const.con->Ico.U1 == True |
| 2228 | || e->Iex.Const.con->Ico.U1 == False); |
| 2229 | r = newVRegI(env); |
| 2230 | addInstr(env, AMD64Instr_Alu64R(Aalu_MOV,AMD64RMI_Imm(0),r)); |
| 2231 | addInstr(env, AMD64Instr_Alu64R(Aalu_XOR,AMD64RMI_Reg(r),r)); |
| 2232 | return e->Iex.Const.con->Ico.U1 ? Acc_Z : Acc_NZ; |
| 2233 | } |
sewardj | 486074e | 2005-02-08 20:10:04 +0000 | [diff] [blame] | 2234 | |
| 2235 | /* Not1(...) */ |
| 2236 | if (e->tag == Iex_Unop && e->Iex.Unop.op == Iop_Not1) { |
| 2237 | /* Generate code for the arg, and negate the test condition */ |
| 2238 | return 1 ^ iselCondCode(env, e->Iex.Unop.arg); |
| 2239 | } |
| 2240 | |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2241 | /* --- patterns rooted at: 64to1 --- */ |
| 2242 | |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2243 | /* 64to1 */ |
| 2244 | if (e->tag == Iex_Unop && e->Iex.Unop.op == Iop_64to1) { |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 2245 | HReg reg = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 2246 | addInstr(env, AMD64Instr_Test64(1,reg)); |
sewardj | f8c37f7 | 2005-02-07 18:55:29 +0000 | [diff] [blame] | 2247 | return Acc_NZ; |
| 2248 | } |
| 2249 | |
florian | c862f28 | 2012-07-19 17:22:33 +0000 | [diff] [blame] | 2250 | /* --- patterns rooted at: 32to1 --- */ |
| 2251 | |
| 2252 | /* 32to1 */ |
| 2253 | if (e->tag == Iex_Unop && e->Iex.Unop.op == Iop_32to1) { |
| 2254 | HReg reg = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 2255 | addInstr(env, AMD64Instr_Test64(1,reg)); |
| 2256 | return Acc_NZ; |
| 2257 | } |
| 2258 | |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2259 | /* --- patterns rooted at: CmpNEZ8 --- */ |
| 2260 | |
| 2261 | /* CmpNEZ8(x) */ |
| 2262 | if (e->tag == Iex_Unop |
| 2263 | && e->Iex.Unop.op == Iop_CmpNEZ8) { |
| 2264 | HReg r = iselIntExpr_R(env, e->Iex.Unop.arg); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 2265 | addInstr(env, AMD64Instr_Test64(0xFF,r)); |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2266 | return Acc_NZ; |
| 2267 | } |
| 2268 | |
sewardj | 86ec28b | 2005-04-27 13:39:35 +0000 | [diff] [blame] | 2269 | /* --- patterns rooted at: CmpNEZ16 --- */ |
| 2270 | |
| 2271 | /* CmpNEZ16(x) */ |
| 2272 | if (e->tag == Iex_Unop |
| 2273 | && e->Iex.Unop.op == Iop_CmpNEZ16) { |
| 2274 | HReg r = iselIntExpr_R(env, e->Iex.Unop.arg); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 2275 | addInstr(env, AMD64Instr_Test64(0xFFFF,r)); |
sewardj | 86ec28b | 2005-04-27 13:39:35 +0000 | [diff] [blame] | 2276 | return Acc_NZ; |
| 2277 | } |
| 2278 | |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2279 | /* --- patterns rooted at: CmpNEZ32 --- */ |
| 2280 | |
| 2281 | /* CmpNEZ32(x) */ |
| 2282 | if (e->tag == Iex_Unop |
| 2283 | && e->Iex.Unop.op == Iop_CmpNEZ32) { |
| 2284 | HReg r1 = iselIntExpr_R(env, e->Iex.Unop.arg); |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2285 | AMD64RMI* rmi2 = AMD64RMI_Imm(0); |
sewardj | 9cc2bbf | 2011-06-05 17:56:03 +0000 | [diff] [blame] | 2286 | addInstr(env, AMD64Instr_Alu32R(Aalu_CMP,rmi2,r1)); |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2287 | return Acc_NZ; |
| 2288 | } |
| 2289 | |
| 2290 | /* --- patterns rooted at: CmpNEZ64 --- */ |
| 2291 | |
sewardj | 0bc78ab | 2005-05-11 22:47:32 +0000 | [diff] [blame] | 2292 | /* CmpNEZ64(Or64(x,y)) */ |
| 2293 | { |
| 2294 | DECLARE_PATTERN(p_CmpNEZ64_Or64); |
| 2295 | DEFINE_PATTERN(p_CmpNEZ64_Or64, |
| 2296 | unop(Iop_CmpNEZ64, binop(Iop_Or64, bind(0), bind(1)))); |
| 2297 | if (matchIRExpr(&mi, p_CmpNEZ64_Or64, e)) { |
| 2298 | HReg r0 = iselIntExpr_R(env, mi.bindee[0]); |
| 2299 | AMD64RMI* rmi1 = iselIntExpr_RMI(env, mi.bindee[1]); |
| 2300 | HReg tmp = newVRegI(env); |
| 2301 | addInstr(env, mk_iMOVsd_RR(r0, tmp)); |
| 2302 | addInstr(env, AMD64Instr_Alu64R(Aalu_OR,rmi1,tmp)); |
| 2303 | return Acc_NZ; |
| 2304 | } |
| 2305 | } |
| 2306 | |
sewardj | 176ad2f | 2005-04-27 11:55:08 +0000 | [diff] [blame] | 2307 | /* CmpNEZ64(x) */ |
| 2308 | if (e->tag == Iex_Unop |
| 2309 | && e->Iex.Unop.op == Iop_CmpNEZ64) { |
| 2310 | HReg r1 = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 2311 | AMD64RMI* rmi2 = AMD64RMI_Imm(0); |
| 2312 | addInstr(env, AMD64Instr_Alu64R(Aalu_CMP,rmi2,r1)); |
| 2313 | return Acc_NZ; |
| 2314 | } |
| 2315 | |
| 2316 | /* --- patterns rooted at: Cmp{EQ,NE}{8,16,32} --- */ |
| 2317 | |
sewardj | 42322b5 | 2005-04-20 22:57:11 +0000 | [diff] [blame] | 2318 | /* CmpEQ8 / CmpNE8 */ |
| 2319 | if (e->tag == Iex_Binop |
| 2320 | && (e->Iex.Binop.op == Iop_CmpEQ8 |
sewardj | 1fb8c92 | 2009-07-12 12:56:53 +0000 | [diff] [blame] | 2321 | || e->Iex.Binop.op == Iop_CmpNE8 |
| 2322 | || e->Iex.Binop.op == Iop_CasCmpEQ8 |
| 2323 | || e->Iex.Binop.op == Iop_CasCmpNE8)) { |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 2324 | if (isZeroU8(e->Iex.Binop.arg2)) { |
| 2325 | HReg r1 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 2326 | addInstr(env, AMD64Instr_Test64(0xFF,r1)); |
| 2327 | switch (e->Iex.Binop.op) { |
| 2328 | case Iop_CmpEQ8: case Iop_CasCmpEQ8: return Acc_Z; |
| 2329 | case Iop_CmpNE8: case Iop_CasCmpNE8: return Acc_NZ; |
| 2330 | default: vpanic("iselCondCode(amd64): CmpXX8(expr,0:I8)"); |
| 2331 | } |
| 2332 | } else { |
| 2333 | HReg r1 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 2334 | AMD64RMI* rmi2 = iselIntExpr_RMI(env, e->Iex.Binop.arg2); |
| 2335 | HReg r = newVRegI(env); |
| 2336 | addInstr(env, mk_iMOVsd_RR(r1,r)); |
| 2337 | addInstr(env, AMD64Instr_Alu64R(Aalu_XOR,rmi2,r)); |
| 2338 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND,AMD64RMI_Imm(0xFF),r)); |
| 2339 | switch (e->Iex.Binop.op) { |
| 2340 | case Iop_CmpEQ8: case Iop_CasCmpEQ8: return Acc_Z; |
| 2341 | case Iop_CmpNE8: case Iop_CasCmpNE8: return Acc_NZ; |
| 2342 | default: vpanic("iselCondCode(amd64): CmpXX8(expr,expr)"); |
| 2343 | } |
sewardj | 42322b5 | 2005-04-20 22:57:11 +0000 | [diff] [blame] | 2344 | } |
| 2345 | } |
| 2346 | |
sewardj | 0af46ab | 2005-04-26 01:52:29 +0000 | [diff] [blame] | 2347 | /* CmpEQ16 / CmpNE16 */ |
| 2348 | if (e->tag == Iex_Binop |
| 2349 | && (e->Iex.Binop.op == Iop_CmpEQ16 |
sewardj | 1fb8c92 | 2009-07-12 12:56:53 +0000 | [diff] [blame] | 2350 | || e->Iex.Binop.op == Iop_CmpNE16 |
| 2351 | || e->Iex.Binop.op == Iop_CasCmpEQ16 |
| 2352 | || e->Iex.Binop.op == Iop_CasCmpNE16)) { |
sewardj | 0af46ab | 2005-04-26 01:52:29 +0000 | [diff] [blame] | 2353 | HReg r1 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 2354 | AMD64RMI* rmi2 = iselIntExpr_RMI(env, e->Iex.Binop.arg2); |
| 2355 | HReg r = newVRegI(env); |
| 2356 | addInstr(env, mk_iMOVsd_RR(r1,r)); |
| 2357 | addInstr(env, AMD64Instr_Alu64R(Aalu_XOR,rmi2,r)); |
| 2358 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND,AMD64RMI_Imm(0xFFFF),r)); |
| 2359 | switch (e->Iex.Binop.op) { |
sewardj | 1fb8c92 | 2009-07-12 12:56:53 +0000 | [diff] [blame] | 2360 | case Iop_CmpEQ16: case Iop_CasCmpEQ16: return Acc_Z; |
| 2361 | case Iop_CmpNE16: case Iop_CasCmpNE16: return Acc_NZ; |
sewardj | 0af46ab | 2005-04-26 01:52:29 +0000 | [diff] [blame] | 2362 | default: vpanic("iselCondCode(amd64): CmpXX16"); |
| 2363 | } |
| 2364 | } |
| 2365 | |
sewardj | 50d89bf | 2011-01-10 15:10:48 +0000 | [diff] [blame] | 2366 | /* CmpNE64(ccall, 64-bit constant) (--smc-check=all optimisation). |
| 2367 | Saves a "movq %rax, %tmp" compared to the default route. */ |
| 2368 | if (e->tag == Iex_Binop |
| 2369 | && e->Iex.Binop.op == Iop_CmpNE64 |
| 2370 | && e->Iex.Binop.arg1->tag == Iex_CCall |
| 2371 | && e->Iex.Binop.arg2->tag == Iex_Const) { |
| 2372 | IRExpr* cal = e->Iex.Binop.arg1; |
| 2373 | IRExpr* con = e->Iex.Binop.arg2; |
| 2374 | HReg tmp = newVRegI(env); |
| 2375 | /* clone & partial-eval of generic Iex_CCall and Iex_Const cases */ |
| 2376 | vassert(cal->Iex.CCall.retty == Ity_I64); /* else ill-typed IR */ |
| 2377 | vassert(con->Iex.Const.con->tag == Ico_U64); |
| 2378 | /* Marshal args, do the call. */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 2379 | UInt addToSp = 0; |
| 2380 | RetLoc rloc = mk_RetLoc_INVALID(); |
| 2381 | doHelperCall( &addToSp, &rloc, env, NULL/*guard*/, |
| 2382 | cal->Iex.CCall.cee, |
| 2383 | cal->Iex.CCall.retty, cal->Iex.CCall.args ); |
| 2384 | vassert(is_sane_RetLoc(rloc)); |
| 2385 | vassert(rloc.pri == RLPri_Int); |
| 2386 | vassert(addToSp == 0); |
| 2387 | /* */ |
sewardj | 50d89bf | 2011-01-10 15:10:48 +0000 | [diff] [blame] | 2388 | addInstr(env, AMD64Instr_Imm64(con->Iex.Const.con->Ico.U64, tmp)); |
| 2389 | addInstr(env, AMD64Instr_Alu64R(Aalu_CMP, |
| 2390 | AMD64RMI_Reg(hregAMD64_RAX()), tmp)); |
| 2391 | return Acc_NZ; |
| 2392 | } |
| 2393 | |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 2394 | /* Cmp*64*(x,y) */ |
| 2395 | if (e->tag == Iex_Binop |
| 2396 | && (e->Iex.Binop.op == Iop_CmpEQ64 |
| 2397 | || e->Iex.Binop.op == Iop_CmpNE64 |
sewardj | 0af46ab | 2005-04-26 01:52:29 +0000 | [diff] [blame] | 2398 | || e->Iex.Binop.op == Iop_CmpLT64S |
| 2399 | || e->Iex.Binop.op == Iop_CmpLT64U |
| 2400 | || e->Iex.Binop.op == Iop_CmpLE64S |
sewardj | a9e4a80 | 2005-12-26 19:33:55 +0000 | [diff] [blame] | 2401 | || e->Iex.Binop.op == Iop_CmpLE64U |
sewardj | 1fb8c92 | 2009-07-12 12:56:53 +0000 | [diff] [blame] | 2402 | || e->Iex.Binop.op == Iop_CasCmpEQ64 |
sewardj | e13074c | 2012-11-08 10:57:08 +0000 | [diff] [blame] | 2403 | || e->Iex.Binop.op == Iop_CasCmpNE64 |
| 2404 | || e->Iex.Binop.op == Iop_ExpCmpNE64)) { |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 2405 | HReg r1 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 2406 | AMD64RMI* rmi2 = iselIntExpr_RMI(env, e->Iex.Binop.arg2); |
| 2407 | addInstr(env, AMD64Instr_Alu64R(Aalu_CMP,rmi2,r1)); |
| 2408 | switch (e->Iex.Binop.op) { |
sewardj | 1fb8c92 | 2009-07-12 12:56:53 +0000 | [diff] [blame] | 2409 | case Iop_CmpEQ64: case Iop_CasCmpEQ64: return Acc_Z; |
sewardj | e13074c | 2012-11-08 10:57:08 +0000 | [diff] [blame] | 2410 | case Iop_CmpNE64: |
| 2411 | case Iop_CasCmpNE64: case Iop_ExpCmpNE64: return Acc_NZ; |
sewardj | 0af46ab | 2005-04-26 01:52:29 +0000 | [diff] [blame] | 2412 | case Iop_CmpLT64S: return Acc_L; |
| 2413 | case Iop_CmpLT64U: return Acc_B; |
| 2414 | case Iop_CmpLE64S: return Acc_LE; |
sewardj | a9e4a80 | 2005-12-26 19:33:55 +0000 | [diff] [blame] | 2415 | case Iop_CmpLE64U: return Acc_BE; |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 2416 | default: vpanic("iselCondCode(amd64): CmpXX64"); |
| 2417 | } |
| 2418 | } |
| 2419 | |
sewardj | 9cc2bbf | 2011-06-05 17:56:03 +0000 | [diff] [blame] | 2420 | /* Cmp*32*(x,y) */ |
| 2421 | if (e->tag == Iex_Binop |
| 2422 | && (e->Iex.Binop.op == Iop_CmpEQ32 |
| 2423 | || e->Iex.Binop.op == Iop_CmpNE32 |
| 2424 | || e->Iex.Binop.op == Iop_CmpLT32S |
| 2425 | || e->Iex.Binop.op == Iop_CmpLT32U |
| 2426 | || e->Iex.Binop.op == Iop_CmpLE32S |
| 2427 | || e->Iex.Binop.op == Iop_CmpLE32U |
| 2428 | || e->Iex.Binop.op == Iop_CasCmpEQ32 |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 2429 | || e->Iex.Binop.op == Iop_CasCmpNE32 |
| 2430 | || e->Iex.Binop.op == Iop_ExpCmpNE32)) { |
sewardj | 9cc2bbf | 2011-06-05 17:56:03 +0000 | [diff] [blame] | 2431 | HReg r1 = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 2432 | AMD64RMI* rmi2 = iselIntExpr_RMI(env, e->Iex.Binop.arg2); |
| 2433 | addInstr(env, AMD64Instr_Alu32R(Aalu_CMP,rmi2,r1)); |
| 2434 | switch (e->Iex.Binop.op) { |
| 2435 | case Iop_CmpEQ32: case Iop_CasCmpEQ32: return Acc_Z; |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 2436 | case Iop_CmpNE32: |
| 2437 | case Iop_CasCmpNE32: case Iop_ExpCmpNE32: return Acc_NZ; |
sewardj | 9cc2bbf | 2011-06-05 17:56:03 +0000 | [diff] [blame] | 2438 | case Iop_CmpLT32S: return Acc_L; |
| 2439 | case Iop_CmpLT32U: return Acc_B; |
| 2440 | case Iop_CmpLE32S: return Acc_LE; |
| 2441 | case Iop_CmpLE32U: return Acc_BE; |
| 2442 | default: vpanic("iselCondCode(amd64): CmpXX32"); |
| 2443 | } |
| 2444 | } |
| 2445 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 2446 | ppIRExpr(e); |
| 2447 | vpanic("iselCondCode(amd64)"); |
| 2448 | } |
| 2449 | |
| 2450 | |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2451 | /*---------------------------------------------------------*/ |
| 2452 | /*--- ISEL: Integer expressions (128 bit) ---*/ |
| 2453 | /*---------------------------------------------------------*/ |
| 2454 | |
| 2455 | /* Compute a 128-bit value into a register pair, which is returned as |
| 2456 | the first two parameters. As with iselIntExpr_R, these may be |
| 2457 | either real or virtual regs; in any case they must not be changed |
| 2458 | by subsequent code emitted by the caller. */ |
| 2459 | |
| 2460 | static void iselInt128Expr ( HReg* rHi, HReg* rLo, |
| 2461 | ISelEnv* env, IRExpr* e ) |
| 2462 | { |
| 2463 | iselInt128Expr_wrk(rHi, rLo, env, e); |
| 2464 | # if 0 |
| 2465 | vex_printf("\n"); ppIRExpr(e); vex_printf("\n"); |
| 2466 | # endif |
| 2467 | vassert(hregClass(*rHi) == HRcInt64); |
| 2468 | vassert(hregIsVirtual(*rHi)); |
| 2469 | vassert(hregClass(*rLo) == HRcInt64); |
| 2470 | vassert(hregIsVirtual(*rLo)); |
| 2471 | } |
| 2472 | |
| 2473 | /* DO NOT CALL THIS DIRECTLY ! */ |
| 2474 | static void iselInt128Expr_wrk ( HReg* rHi, HReg* rLo, |
| 2475 | ISelEnv* env, IRExpr* e ) |
| 2476 | { |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2477 | vassert(e); |
| 2478 | vassert(typeOfIRExpr(env->type_env,e) == Ity_I128); |
| 2479 | |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2480 | /* read 128-bit IRTemp */ |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 2481 | if (e->tag == Iex_RdTmp) { |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 2482 | lookupIRTempPair( rHi, rLo, env, e->Iex.RdTmp.tmp); |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2483 | return; |
| 2484 | } |
| 2485 | |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2486 | /* --------- BINARY ops --------- */ |
| 2487 | if (e->tag == Iex_Binop) { |
| 2488 | switch (e->Iex.Binop.op) { |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 2489 | /* 64 x 64 -> 128 multiply */ |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2490 | case Iop_MullU64: |
| 2491 | case Iop_MullS64: { |
| 2492 | /* get one operand into %rax, and the other into a R/M. |
| 2493 | Need to make an educated guess about which is better in |
| 2494 | which. */ |
| 2495 | HReg tLo = newVRegI(env); |
| 2496 | HReg tHi = newVRegI(env); |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 2497 | Bool syned = toBool(e->Iex.Binop.op == Iop_MullS64); |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2498 | AMD64RM* rmLeft = iselIntExpr_RM(env, e->Iex.Binop.arg1); |
| 2499 | HReg rRight = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 2500 | addInstr(env, mk_iMOVsd_RR(rRight, hregAMD64_RAX())); |
sewardj | 501a339 | 2005-05-11 15:37:50 +0000 | [diff] [blame] | 2501 | addInstr(env, AMD64Instr_MulL(syned, rmLeft)); |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2502 | /* Result is now in RDX:RAX. Tell the caller. */ |
| 2503 | addInstr(env, mk_iMOVsd_RR(hregAMD64_RDX(), tHi)); |
| 2504 | addInstr(env, mk_iMOVsd_RR(hregAMD64_RAX(), tLo)); |
| 2505 | *rHi = tHi; |
| 2506 | *rLo = tLo; |
| 2507 | return; |
| 2508 | } |
sewardj | 7de0d3c | 2005-02-13 02:26:41 +0000 | [diff] [blame] | 2509 | |
sewardj | a6b93d1 | 2005-02-17 09:28:28 +0000 | [diff] [blame] | 2510 | /* 128 x 64 -> (64(rem),64(div)) division */ |
| 2511 | case Iop_DivModU128to64: |
| 2512 | case Iop_DivModS128to64: { |
| 2513 | /* Get the 128-bit operand into rdx:rax, and the other into |
| 2514 | any old R/M. */ |
| 2515 | HReg sHi, sLo; |
| 2516 | HReg tLo = newVRegI(env); |
| 2517 | HReg tHi = newVRegI(env); |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 2518 | Bool syned = toBool(e->Iex.Binop.op == Iop_DivModS128to64); |
sewardj | a6b93d1 | 2005-02-17 09:28:28 +0000 | [diff] [blame] | 2519 | AMD64RM* rmRight = iselIntExpr_RM(env, e->Iex.Binop.arg2); |
| 2520 | iselInt128Expr(&sHi,&sLo, env, e->Iex.Binop.arg1); |
| 2521 | addInstr(env, mk_iMOVsd_RR(sHi, hregAMD64_RDX())); |
| 2522 | addInstr(env, mk_iMOVsd_RR(sLo, hregAMD64_RAX())); |
| 2523 | addInstr(env, AMD64Instr_Div(syned, 8, rmRight)); |
| 2524 | addInstr(env, mk_iMOVsd_RR(hregAMD64_RDX(), tHi)); |
| 2525 | addInstr(env, mk_iMOVsd_RR(hregAMD64_RAX(), tLo)); |
| 2526 | *rHi = tHi; |
| 2527 | *rLo = tLo; |
| 2528 | return; |
| 2529 | } |
| 2530 | |
| 2531 | /* 64HLto128(e1,e2) */ |
| 2532 | case Iop_64HLto128: |
| 2533 | *rHi = iselIntExpr_R(env, e->Iex.Binop.arg1); |
| 2534 | *rLo = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 2535 | return; |
| 2536 | |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2537 | default: |
| 2538 | break; |
| 2539 | } |
| 2540 | } /* if (e->tag == Iex_Binop) */ |
| 2541 | |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 2542 | ppIRExpr(e); |
| 2543 | vpanic("iselInt128Expr"); |
| 2544 | } |
| 2545 | |
| 2546 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2547 | /*---------------------------------------------------------*/ |
| 2548 | /*--- ISEL: Floating point expressions (32 bit) ---*/ |
| 2549 | /*---------------------------------------------------------*/ |
| 2550 | |
| 2551 | /* Nothing interesting here; really just wrappers for |
| 2552 | 64-bit stuff. */ |
| 2553 | |
| 2554 | static HReg iselFltExpr ( ISelEnv* env, IRExpr* e ) |
| 2555 | { |
| 2556 | HReg r = iselFltExpr_wrk( env, e ); |
| 2557 | # if 0 |
| 2558 | vex_printf("\n"); ppIRExpr(e); vex_printf("\n"); |
| 2559 | # endif |
| 2560 | vassert(hregClass(r) == HRcVec128); |
| 2561 | vassert(hregIsVirtual(r)); |
| 2562 | return r; |
| 2563 | } |
| 2564 | |
| 2565 | /* DO NOT CALL THIS DIRECTLY */ |
| 2566 | static HReg iselFltExpr_wrk ( ISelEnv* env, IRExpr* e ) |
| 2567 | { |
| 2568 | IRType ty = typeOfIRExpr(env->type_env,e); |
| 2569 | vassert(ty == Ity_F32); |
| 2570 | |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 2571 | if (e->tag == Iex_RdTmp) { |
| 2572 | return lookupIRTemp(env, e->Iex.RdTmp.tmp); |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 2573 | } |
| 2574 | |
sewardj | e768e92 | 2009-11-26 17:17:37 +0000 | [diff] [blame] | 2575 | if (e->tag == Iex_Load && e->Iex.Load.end == Iend_LE) { |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 2576 | AMD64AMode* am; |
| 2577 | HReg res = newVRegV(env); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 2578 | vassert(e->Iex.Load.ty == Ity_F32); |
| 2579 | am = iselIntExpr_AMode(env, e->Iex.Load.addr); |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 2580 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 4, res, am)); |
| 2581 | return res; |
| 2582 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2583 | |
| 2584 | if (e->tag == Iex_Binop |
| 2585 | && e->Iex.Binop.op == Iop_F64toF32) { |
| 2586 | /* Although the result is still held in a standard SSE register, |
| 2587 | we need to round it to reflect the loss of accuracy/range |
| 2588 | entailed in casting it to a 32-bit float. */ |
| 2589 | HReg dst = newVRegV(env); |
| 2590 | HReg src = iselDblExpr(env, e->Iex.Binop.arg2); |
| 2591 | set_SSE_rounding_mode( env, e->Iex.Binop.arg1 ); |
| 2592 | addInstr(env, AMD64Instr_SseSDSS(True/*D->S*/,src,dst)); |
| 2593 | set_SSE_rounding_default( env ); |
| 2594 | return dst; |
| 2595 | } |
| 2596 | |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 2597 | if (e->tag == Iex_Get) { |
| 2598 | AMD64AMode* am = AMD64AMode_IR( e->Iex.Get.offset, |
| 2599 | hregAMD64_RBP() ); |
| 2600 | HReg res = newVRegV(env); |
| 2601 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 4, res, am )); |
| 2602 | return res; |
| 2603 | } |
| 2604 | |
sewardj | 5992bd0 | 2005-05-11 02:13:42 +0000 | [diff] [blame] | 2605 | if (e->tag == Iex_Unop |
| 2606 | && e->Iex.Unop.op == Iop_ReinterpI32asF32) { |
| 2607 | /* Given an I32, produce an IEEE754 float with the same bit |
| 2608 | pattern. */ |
| 2609 | HReg dst = newVRegV(env); |
| 2610 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 2611 | AMD64AMode* m4_rsp = AMD64AMode_IR(-4, hregAMD64_RSP()); |
| 2612 | addInstr(env, AMD64Instr_Store(4, src, m4_rsp)); |
| 2613 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 4, dst, m4_rsp )); |
| 2614 | return dst; |
| 2615 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2616 | |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 2617 | if (e->tag == Iex_Binop && e->Iex.Binop.op == Iop_RoundF32toInt) { |
| 2618 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
| 2619 | HReg arg = iselFltExpr(env, e->Iex.Binop.arg2); |
| 2620 | HReg dst = newVRegV(env); |
| 2621 | |
| 2622 | /* rf now holds the value to be rounded. The first thing to do |
| 2623 | is set the FPU's rounding mode accordingly. */ |
| 2624 | |
| 2625 | /* Set host x87 rounding mode */ |
| 2626 | set_FPU_rounding_mode( env, e->Iex.Binop.arg1 ); |
| 2627 | |
| 2628 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 4, arg, m8_rsp)); |
| 2629 | addInstr(env, AMD64Instr_A87Free(1)); |
| 2630 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, True/*push*/, 4)); |
| 2631 | addInstr(env, AMD64Instr_A87FpOp(Afp_ROUND)); |
| 2632 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, False/*pop*/, 4)); |
| 2633 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 4, dst, m8_rsp)); |
| 2634 | |
| 2635 | /* Restore default x87 rounding. */ |
| 2636 | set_FPU_rounding_default( env ); |
| 2637 | |
| 2638 | return dst; |
| 2639 | } |
| 2640 | |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 2641 | if (e->tag == Iex_Unop && e->Iex.Unop.op == Iop_NegF32) { |
| 2642 | /* Sigh ... very rough code. Could do much better. */ |
| 2643 | /* Get the 128-bit literal 00---0 10---0 into a register |
| 2644 | and xor it with the value to be negated. */ |
| 2645 | HReg r1 = newVRegI(env); |
| 2646 | HReg dst = newVRegV(env); |
| 2647 | HReg tmp = newVRegV(env); |
| 2648 | HReg src = iselFltExpr(env, e->Iex.Unop.arg); |
| 2649 | AMD64AMode* rsp0 = AMD64AMode_IR(0, hregAMD64_RSP()); |
| 2650 | addInstr(env, mk_vMOVsd_RR(src,tmp)); |
| 2651 | addInstr(env, AMD64Instr_Push(AMD64RMI_Imm(0))); |
| 2652 | addInstr(env, AMD64Instr_Imm64( 1ULL<<31, r1 )); |
| 2653 | addInstr(env, AMD64Instr_Push(AMD64RMI_Reg(r1))); |
| 2654 | addInstr(env, AMD64Instr_SseLdSt(True, 16, dst, rsp0)); |
| 2655 | addInstr(env, AMD64Instr_SseReRg(Asse_XOR, tmp, dst)); |
| 2656 | add_to_rsp(env, 16); |
| 2657 | return dst; |
| 2658 | } |
| 2659 | |
| 2660 | if (e->tag == Iex_Qop && e->Iex.Qop.details->op == Iop_MAddF32) { |
| 2661 | IRQop *qop = e->Iex.Qop.details; |
| 2662 | HReg dst = newVRegV(env); |
| 2663 | HReg argX = iselFltExpr(env, qop->arg2); |
| 2664 | HReg argY = iselFltExpr(env, qop->arg3); |
| 2665 | HReg argZ = iselFltExpr(env, qop->arg4); |
| 2666 | /* XXXROUNDINGFIXME */ |
| 2667 | /* set roundingmode here */ |
| 2668 | /* subq $16, %rsp -- make a space*/ |
| 2669 | sub_from_rsp(env, 16); |
| 2670 | /* Prepare 4 arg regs: |
| 2671 | leaq 0(%rsp), %rdi |
| 2672 | leaq 4(%rsp), %rsi |
| 2673 | leaq 8(%rsp), %rdx |
| 2674 | leaq 12(%rsp), %rcx |
| 2675 | */ |
| 2676 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(0, hregAMD64_RSP()), |
| 2677 | hregAMD64_RDI())); |
| 2678 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(4, hregAMD64_RSP()), |
| 2679 | hregAMD64_RSI())); |
| 2680 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(8, hregAMD64_RSP()), |
| 2681 | hregAMD64_RDX())); |
| 2682 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(12, hregAMD64_RSP()), |
| 2683 | hregAMD64_RCX())); |
| 2684 | /* Store the three args, at (%rsi), (%rdx) and (%rcx): |
| 2685 | movss %argX, 0(%rsi) |
| 2686 | movss %argY, 0(%rdx) |
| 2687 | movss %argZ, 0(%rcx) |
| 2688 | */ |
| 2689 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 4, argX, |
| 2690 | AMD64AMode_IR(0, hregAMD64_RSI()))); |
| 2691 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 4, argY, |
| 2692 | AMD64AMode_IR(0, hregAMD64_RDX()))); |
| 2693 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 4, argZ, |
| 2694 | AMD64AMode_IR(0, hregAMD64_RCX()))); |
| 2695 | /* call the helper */ |
| 2696 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, |
| 2697 | (ULong)(HWord)h_generic_calc_MAddF32, |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 2698 | 4, mk_RetLoc_simple(RLPri_None) )); |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 2699 | /* fetch the result from memory, using %r_argp, which the |
| 2700 | register allocator will keep alive across the call. */ |
| 2701 | addInstr(env, AMD64Instr_SseLdSt(True/*isLoad*/, 4, dst, |
| 2702 | AMD64AMode_IR(0, hregAMD64_RSP()))); |
| 2703 | /* and finally, clear the space */ |
| 2704 | add_to_rsp(env, 16); |
| 2705 | return dst; |
| 2706 | } |
| 2707 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2708 | ppIRExpr(e); |
| 2709 | vpanic("iselFltExpr_wrk"); |
| 2710 | } |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 2711 | |
| 2712 | |
| 2713 | /*---------------------------------------------------------*/ |
| 2714 | /*--- ISEL: Floating point expressions (64 bit) ---*/ |
| 2715 | /*---------------------------------------------------------*/ |
| 2716 | |
| 2717 | /* Compute a 64-bit floating point value into the lower half of an xmm |
| 2718 | register, the identity of which is returned. As with |
| 2719 | iselIntExpr_R, the returned reg will be virtual, and it must not be |
| 2720 | changed by subsequent code emitted by the caller. |
| 2721 | */ |
| 2722 | |
| 2723 | /* IEEE 754 formats. From http://www.freesoft.org/CIE/RFC/1832/32.htm: |
| 2724 | |
| 2725 | Type S (1 bit) E (11 bits) F (52 bits) |
| 2726 | ---- --------- ----------- ----------- |
| 2727 | signalling NaN u 2047 (max) .0uuuuu---u |
| 2728 | (with at least |
| 2729 | one 1 bit) |
| 2730 | quiet NaN u 2047 (max) .1uuuuu---u |
| 2731 | |
| 2732 | negative infinity 1 2047 (max) .000000---0 |
| 2733 | |
| 2734 | positive infinity 0 2047 (max) .000000---0 |
| 2735 | |
| 2736 | negative zero 1 0 .000000---0 |
| 2737 | |
| 2738 | positive zero 0 0 .000000---0 |
| 2739 | */ |
| 2740 | |
| 2741 | static HReg iselDblExpr ( ISelEnv* env, IRExpr* e ) |
| 2742 | { |
| 2743 | HReg r = iselDblExpr_wrk( env, e ); |
| 2744 | # if 0 |
| 2745 | vex_printf("\n"); ppIRExpr(e); vex_printf("\n"); |
| 2746 | # endif |
| 2747 | vassert(hregClass(r) == HRcVec128); |
| 2748 | vassert(hregIsVirtual(r)); |
| 2749 | return r; |
| 2750 | } |
| 2751 | |
| 2752 | /* DO NOT CALL THIS DIRECTLY */ |
| 2753 | static HReg iselDblExpr_wrk ( ISelEnv* env, IRExpr* e ) |
| 2754 | { |
| 2755 | IRType ty = typeOfIRExpr(env->type_env,e); |
| 2756 | vassert(e); |
| 2757 | vassert(ty == Ity_F64); |
| 2758 | |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 2759 | if (e->tag == Iex_RdTmp) { |
| 2760 | return lookupIRTemp(env, e->Iex.RdTmp.tmp); |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 2761 | } |
| 2762 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2763 | if (e->tag == Iex_Const) { |
| 2764 | union { ULong u64; Double f64; } u; |
| 2765 | HReg res = newVRegV(env); |
| 2766 | HReg tmp = newVRegI(env); |
| 2767 | vassert(sizeof(u) == 8); |
| 2768 | vassert(sizeof(u.u64) == 8); |
| 2769 | vassert(sizeof(u.f64) == 8); |
| 2770 | |
| 2771 | if (e->Iex.Const.con->tag == Ico_F64) { |
| 2772 | u.f64 = e->Iex.Const.con->Ico.F64; |
| 2773 | } |
| 2774 | else if (e->Iex.Const.con->tag == Ico_F64i) { |
| 2775 | u.u64 = e->Iex.Const.con->Ico.F64i; |
| 2776 | } |
| 2777 | else |
| 2778 | vpanic("iselDblExpr(amd64): const"); |
| 2779 | |
| 2780 | addInstr(env, AMD64Instr_Imm64(u.u64, tmp)); |
| 2781 | addInstr(env, AMD64Instr_Push(AMD64RMI_Reg(tmp))); |
| 2782 | addInstr(env, AMD64Instr_SseLdSt( |
| 2783 | True/*load*/, 8, res, |
| 2784 | AMD64AMode_IR(0, hregAMD64_RSP()) |
| 2785 | )); |
| 2786 | add_to_rsp(env, 8); |
| 2787 | return res; |
| 2788 | } |
sewardj | 9da1697 | 2005-02-21 13:58:26 +0000 | [diff] [blame] | 2789 | |
sewardj | e768e92 | 2009-11-26 17:17:37 +0000 | [diff] [blame] | 2790 | if (e->tag == Iex_Load && e->Iex.Load.end == Iend_LE) { |
sewardj | 9da1697 | 2005-02-21 13:58:26 +0000 | [diff] [blame] | 2791 | AMD64AMode* am; |
| 2792 | HReg res = newVRegV(env); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 2793 | vassert(e->Iex.Load.ty == Ity_F64); |
| 2794 | am = iselIntExpr_AMode(env, e->Iex.Load.addr); |
sewardj | 9da1697 | 2005-02-21 13:58:26 +0000 | [diff] [blame] | 2795 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 8, res, am )); |
| 2796 | return res; |
| 2797 | } |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 2798 | |
| 2799 | if (e->tag == Iex_Get) { |
| 2800 | AMD64AMode* am = AMD64AMode_IR( e->Iex.Get.offset, |
| 2801 | hregAMD64_RBP() ); |
| 2802 | HReg res = newVRegV(env); |
| 2803 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 8, res, am )); |
| 2804 | return res; |
| 2805 | } |
| 2806 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2807 | if (e->tag == Iex_GetI) { |
| 2808 | AMD64AMode* am |
| 2809 | = genGuestArrayOffset( |
| 2810 | env, e->Iex.GetI.descr, |
| 2811 | e->Iex.GetI.ix, e->Iex.GetI.bias ); |
| 2812 | HReg res = newVRegV(env); |
| 2813 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 8, res, am )); |
| 2814 | return res; |
| 2815 | } |
| 2816 | |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 2817 | if (e->tag == Iex_Triop) { |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 2818 | IRTriop *triop = e->Iex.Triop.details; |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 2819 | AMD64SseOp op = Asse_INVALID; |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 2820 | switch (triop->op) { |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 2821 | case Iop_AddF64: op = Asse_ADDF; break; |
| 2822 | case Iop_SubF64: op = Asse_SUBF; break; |
| 2823 | case Iop_MulF64: op = Asse_MULF; break; |
| 2824 | case Iop_DivF64: op = Asse_DIVF; break; |
| 2825 | default: break; |
| 2826 | } |
| 2827 | if (op != Asse_INVALID) { |
| 2828 | HReg dst = newVRegV(env); |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 2829 | HReg argL = iselDblExpr(env, triop->arg2); |
| 2830 | HReg argR = iselDblExpr(env, triop->arg3); |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 2831 | addInstr(env, mk_vMOVsd_RR(argL, dst)); |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 2832 | /* XXXROUNDINGFIXME */ |
| 2833 | /* set roundingmode here */ |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 2834 | addInstr(env, AMD64Instr_Sse64FLo(op, argR, dst)); |
| 2835 | return dst; |
| 2836 | } |
| 2837 | } |
| 2838 | |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 2839 | if (e->tag == Iex_Qop && e->Iex.Qop.details->op == Iop_MAddF64) { |
| 2840 | IRQop *qop = e->Iex.Qop.details; |
| 2841 | HReg dst = newVRegV(env); |
| 2842 | HReg argX = iselDblExpr(env, qop->arg2); |
| 2843 | HReg argY = iselDblExpr(env, qop->arg3); |
| 2844 | HReg argZ = iselDblExpr(env, qop->arg4); |
| 2845 | /* XXXROUNDINGFIXME */ |
| 2846 | /* set roundingmode here */ |
| 2847 | /* subq $32, %rsp -- make a space*/ |
| 2848 | sub_from_rsp(env, 32); |
| 2849 | /* Prepare 4 arg regs: |
| 2850 | leaq 0(%rsp), %rdi |
| 2851 | leaq 8(%rsp), %rsi |
| 2852 | leaq 16(%rsp), %rdx |
| 2853 | leaq 24(%rsp), %rcx |
| 2854 | */ |
| 2855 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(0, hregAMD64_RSP()), |
| 2856 | hregAMD64_RDI())); |
| 2857 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(8, hregAMD64_RSP()), |
| 2858 | hregAMD64_RSI())); |
| 2859 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(16, hregAMD64_RSP()), |
| 2860 | hregAMD64_RDX())); |
| 2861 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(24, hregAMD64_RSP()), |
| 2862 | hregAMD64_RCX())); |
| 2863 | /* Store the three args, at (%rsi), (%rdx) and (%rcx): |
| 2864 | movsd %argX, 0(%rsi) |
| 2865 | movsd %argY, 0(%rdx) |
| 2866 | movsd %argZ, 0(%rcx) |
| 2867 | */ |
| 2868 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 8, argX, |
| 2869 | AMD64AMode_IR(0, hregAMD64_RSI()))); |
| 2870 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 8, argY, |
| 2871 | AMD64AMode_IR(0, hregAMD64_RDX()))); |
| 2872 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 8, argZ, |
| 2873 | AMD64AMode_IR(0, hregAMD64_RCX()))); |
| 2874 | /* call the helper */ |
| 2875 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, |
| 2876 | (ULong)(HWord)h_generic_calc_MAddF64, |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 2877 | 4, mk_RetLoc_simple(RLPri_None) )); |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 2878 | /* fetch the result from memory, using %r_argp, which the |
| 2879 | register allocator will keep alive across the call. */ |
| 2880 | addInstr(env, AMD64Instr_SseLdSt(True/*isLoad*/, 8, dst, |
| 2881 | AMD64AMode_IR(0, hregAMD64_RSP()))); |
| 2882 | /* and finally, clear the space */ |
| 2883 | add_to_rsp(env, 32); |
| 2884 | return dst; |
| 2885 | } |
| 2886 | |
sewardj | b183b85 | 2006-02-03 16:08:03 +0000 | [diff] [blame] | 2887 | if (e->tag == Iex_Binop && e->Iex.Binop.op == Iop_RoundF64toInt) { |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2888 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
| 2889 | HReg arg = iselDblExpr(env, e->Iex.Binop.arg2); |
| 2890 | HReg dst = newVRegV(env); |
| 2891 | |
| 2892 | /* rf now holds the value to be rounded. The first thing to do |
| 2893 | is set the FPU's rounding mode accordingly. */ |
| 2894 | |
| 2895 | /* Set host x87 rounding mode */ |
| 2896 | set_FPU_rounding_mode( env, e->Iex.Binop.arg1 ); |
| 2897 | |
| 2898 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 8, arg, m8_rsp)); |
| 2899 | addInstr(env, AMD64Instr_A87Free(1)); |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 2900 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, True/*push*/, 8)); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2901 | addInstr(env, AMD64Instr_A87FpOp(Afp_ROUND)); |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 2902 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, False/*pop*/, 8)); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2903 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 8, dst, m8_rsp)); |
| 2904 | |
| 2905 | /* Restore default x87 rounding. */ |
| 2906 | set_FPU_rounding_default( env ); |
| 2907 | |
| 2908 | return dst; |
| 2909 | } |
| 2910 | |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 2911 | IRTriop *triop = e->Iex.Triop.details; |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 2912 | if (e->tag == Iex_Triop |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 2913 | && (triop->op == Iop_ScaleF64 |
| 2914 | || triop->op == Iop_AtanF64 |
| 2915 | || triop->op == Iop_Yl2xF64 |
| 2916 | || triop->op == Iop_Yl2xp1F64 |
| 2917 | || triop->op == Iop_PRemF64 |
| 2918 | || triop->op == Iop_PRem1F64) |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2919 | ) { |
| 2920 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 2921 | HReg arg1 = iselDblExpr(env, triop->arg2); |
| 2922 | HReg arg2 = iselDblExpr(env, triop->arg3); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2923 | HReg dst = newVRegV(env); |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 2924 | Bool arg2first = toBool(triop->op == Iop_ScaleF64 |
| 2925 | || triop->op == Iop_PRemF64 |
| 2926 | || triop->op == Iop_PRem1F64); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2927 | addInstr(env, AMD64Instr_A87Free(2)); |
| 2928 | |
| 2929 | /* one arg -> top of x87 stack */ |
| 2930 | addInstr(env, AMD64Instr_SseLdSt( |
| 2931 | False/*store*/, 8, arg2first ? arg2 : arg1, m8_rsp)); |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 2932 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, True/*push*/, 8)); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2933 | |
| 2934 | /* other arg -> top of x87 stack */ |
| 2935 | addInstr(env, AMD64Instr_SseLdSt( |
| 2936 | False/*store*/, 8, arg2first ? arg1 : arg2, m8_rsp)); |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 2937 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, True/*push*/, 8)); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2938 | |
| 2939 | /* do it */ |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 2940 | /* XXXROUNDINGFIXME */ |
| 2941 | /* set roundingmode here */ |
florian | 420bfa9 | 2012-06-02 20:29:22 +0000 | [diff] [blame] | 2942 | switch (triop->op) { |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2943 | case Iop_ScaleF64: |
| 2944 | addInstr(env, AMD64Instr_A87FpOp(Afp_SCALE)); |
| 2945 | break; |
| 2946 | case Iop_AtanF64: |
| 2947 | addInstr(env, AMD64Instr_A87FpOp(Afp_ATAN)); |
| 2948 | break; |
| 2949 | case Iop_Yl2xF64: |
| 2950 | addInstr(env, AMD64Instr_A87FpOp(Afp_YL2X)); |
| 2951 | break; |
sewardj | 5e20537 | 2005-05-09 02:57:08 +0000 | [diff] [blame] | 2952 | case Iop_Yl2xp1F64: |
| 2953 | addInstr(env, AMD64Instr_A87FpOp(Afp_YL2XP1)); |
| 2954 | break; |
sewardj | f4c803b | 2006-09-11 11:07:34 +0000 | [diff] [blame] | 2955 | case Iop_PRemF64: |
| 2956 | addInstr(env, AMD64Instr_A87FpOp(Afp_PREM)); |
| 2957 | break; |
sewardj | 4970e4e | 2008-10-11 10:07:55 +0000 | [diff] [blame] | 2958 | case Iop_PRem1F64: |
| 2959 | addInstr(env, AMD64Instr_A87FpOp(Afp_PREM1)); |
| 2960 | break; |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2961 | default: |
| 2962 | vassert(0); |
| 2963 | } |
| 2964 | |
| 2965 | /* save result */ |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 2966 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, False/*pop*/, 8)); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 2967 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 8, dst, m8_rsp)); |
| 2968 | return dst; |
| 2969 | } |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 2970 | |
sewardj | 6c299f3 | 2009-12-31 18:00:12 +0000 | [diff] [blame] | 2971 | if (e->tag == Iex_Binop && e->Iex.Binop.op == Iop_I64StoF64) { |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 2972 | HReg dst = newVRegV(env); |
| 2973 | HReg src = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 2974 | set_SSE_rounding_mode( env, e->Iex.Binop.arg1 ); |
| 2975 | addInstr(env, AMD64Instr_SseSI2SF( 8, 8, src, dst )); |
| 2976 | set_SSE_rounding_default( env ); |
| 2977 | return dst; |
| 2978 | } |
| 2979 | |
sewardj | 6c299f3 | 2009-12-31 18:00:12 +0000 | [diff] [blame] | 2980 | if (e->tag == Iex_Unop && e->Iex.Unop.op == Iop_I32StoF64) { |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 2981 | HReg dst = newVRegV(env); |
| 2982 | HReg src = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 2983 | set_SSE_rounding_default( env ); |
| 2984 | addInstr(env, AMD64Instr_SseSI2SF( 4, 8, src, dst )); |
| 2985 | return dst; |
| 2986 | } |
| 2987 | |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 2988 | if (e->tag == Iex_Unop |
| 2989 | && (e->Iex.Unop.op == Iop_NegF64 |
| 2990 | || e->Iex.Unop.op == Iop_AbsF64)) { |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2991 | /* Sigh ... very rough code. Could do much better. */ |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 2992 | /* Get the 128-bit literal 00---0 10---0 into a register |
| 2993 | and xor/nand it with the value to be negated. */ |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2994 | HReg r1 = newVRegI(env); |
| 2995 | HReg dst = newVRegV(env); |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 2996 | HReg tmp = newVRegV(env); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 2997 | HReg src = iselDblExpr(env, e->Iex.Unop.arg); |
| 2998 | AMD64AMode* rsp0 = AMD64AMode_IR(0, hregAMD64_RSP()); |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 2999 | addInstr(env, mk_vMOVsd_RR(src,tmp)); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3000 | addInstr(env, AMD64Instr_Push(AMD64RMI_Imm(0))); |
| 3001 | addInstr(env, AMD64Instr_Imm64( 1ULL<<63, r1 )); |
| 3002 | addInstr(env, AMD64Instr_Push(AMD64RMI_Reg(r1))); |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 3003 | addInstr(env, AMD64Instr_SseLdSt(True, 16, dst, rsp0)); |
| 3004 | |
| 3005 | if (e->Iex.Unop.op == Iop_NegF64) |
| 3006 | addInstr(env, AMD64Instr_SseReRg(Asse_XOR, tmp, dst)); |
| 3007 | else |
| 3008 | addInstr(env, AMD64Instr_SseReRg(Asse_ANDN, tmp, dst)); |
| 3009 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3010 | add_to_rsp(env, 16); |
| 3011 | return dst; |
| 3012 | } |
| 3013 | |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 3014 | if (e->tag == Iex_Binop) { |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 3015 | A87FpOp fpop = Afp_INVALID; |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 3016 | switch (e->Iex.Binop.op) { |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 3017 | case Iop_SqrtF64: fpop = Afp_SQRT; break; |
sewardj | 5e20537 | 2005-05-09 02:57:08 +0000 | [diff] [blame] | 3018 | case Iop_SinF64: fpop = Afp_SIN; break; |
| 3019 | case Iop_CosF64: fpop = Afp_COS; break; |
| 3020 | case Iop_TanF64: fpop = Afp_TAN; break; |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 3021 | case Iop_2xm1F64: fpop = Afp_2XM1; break; |
| 3022 | default: break; |
| 3023 | } |
| 3024 | if (fpop != Afp_INVALID) { |
| 3025 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 3026 | HReg arg = iselDblExpr(env, e->Iex.Binop.arg2); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 3027 | HReg dst = newVRegV(env); |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 3028 | Int nNeeded = e->Iex.Binop.op==Iop_TanF64 ? 2 : 1; |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 3029 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 8, arg, m8_rsp)); |
sewardj | 5e20537 | 2005-05-09 02:57:08 +0000 | [diff] [blame] | 3030 | addInstr(env, AMD64Instr_A87Free(nNeeded)); |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 3031 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, True/*push*/, 8)); |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 3032 | /* XXXROUNDINGFIXME */ |
| 3033 | /* set roundingmode here */ |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 3034 | addInstr(env, AMD64Instr_A87FpOp(fpop)); |
sewardj | 4796d66 | 2006-02-05 16:06:26 +0000 | [diff] [blame] | 3035 | if (e->Iex.Binop.op==Iop_TanF64) { |
sewardj | 5e20537 | 2005-05-09 02:57:08 +0000 | [diff] [blame] | 3036 | /* get rid of the extra 1.0 that fptan pushes */ |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 3037 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, False/*pop*/, 8)); |
sewardj | 5e20537 | 2005-05-09 02:57:08 +0000 | [diff] [blame] | 3038 | } |
sewardj | d15b597 | 2010-06-27 09:06:34 +0000 | [diff] [blame] | 3039 | addInstr(env, AMD64Instr_A87PushPop(m8_rsp, False/*pop*/, 8)); |
sewardj | 25a8581 | 2005-05-08 23:03:48 +0000 | [diff] [blame] | 3040 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 8, dst, m8_rsp)); |
| 3041 | return dst; |
| 3042 | } |
| 3043 | } |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 3044 | |
| 3045 | if (e->tag == Iex_Unop) { |
| 3046 | switch (e->Iex.Unop.op) { |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 3047 | //.. case Iop_I32toF64: { |
| 3048 | //.. HReg dst = newVRegF(env); |
| 3049 | //.. HReg ri = iselIntExpr_R(env, e->Iex.Unop.arg); |
| 3050 | //.. addInstr(env, X86Instr_Push(X86RMI_Reg(ri))); |
| 3051 | //.. set_FPU_rounding_default(env); |
| 3052 | //.. addInstr(env, X86Instr_FpLdStI( |
| 3053 | //.. True/*load*/, 4, dst, |
| 3054 | //.. X86AMode_IR(0, hregX86_ESP()))); |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 3055 | //.. add_to_esp(env, 4); |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 3056 | //.. return dst; |
| 3057 | //.. } |
sewardj | 924215b | 2005-03-26 21:50:31 +0000 | [diff] [blame] | 3058 | case Iop_ReinterpI64asF64: { |
| 3059 | /* Given an I64, produce an IEEE754 double with the same |
| 3060 | bit pattern. */ |
| 3061 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, hregAMD64_RSP()); |
| 3062 | HReg dst = newVRegV(env); |
| 3063 | AMD64RI* src = iselIntExpr_RI(env, e->Iex.Unop.arg); |
| 3064 | /* paranoia */ |
| 3065 | set_SSE_rounding_default(env); |
| 3066 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, src, m8_rsp)); |
| 3067 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 8, dst, m8_rsp)); |
| 3068 | return dst; |
| 3069 | } |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 3070 | case Iop_F32toF64: { |
sewardj | 9a036bf | 2005-03-14 18:19:08 +0000 | [diff] [blame] | 3071 | HReg f32; |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 3072 | HReg f64 = newVRegV(env); |
| 3073 | /* this shouldn't be necessary, but be paranoid ... */ |
| 3074 | set_SSE_rounding_default(env); |
sewardj | 9a036bf | 2005-03-14 18:19:08 +0000 | [diff] [blame] | 3075 | f32 = iselFltExpr(env, e->Iex.Unop.arg); |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 3076 | addInstr(env, AMD64Instr_SseSDSS(False/*S->D*/, f32, f64)); |
| 3077 | return f64; |
| 3078 | } |
| 3079 | default: |
| 3080 | break; |
| 3081 | } |
| 3082 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3083 | |
| 3084 | /* --------- MULTIPLEX --------- */ |
florian | 99dd03e | 2013-01-29 03:56:06 +0000 | [diff] [blame] | 3085 | if (e->tag == Iex_ITE) { // VFD |
| 3086 | HReg r1, r0, dst; |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3087 | vassert(ty == Ity_F64); |
florian | 99dd03e | 2013-01-29 03:56:06 +0000 | [diff] [blame] | 3088 | vassert(typeOfIRExpr(env->type_env,e->Iex.ITE.cond) == Ity_I1); |
| 3089 | r1 = iselDblExpr(env, e->Iex.ITE.iftrue); |
| 3090 | r0 = iselDblExpr(env, e->Iex.ITE.iffalse); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3091 | dst = newVRegV(env); |
florian | 99dd03e | 2013-01-29 03:56:06 +0000 | [diff] [blame] | 3092 | addInstr(env, mk_vMOVsd_RR(r1,dst)); |
| 3093 | AMD64CondCode cc = iselCondCode(env, e->Iex.ITE.cond); |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 3094 | addInstr(env, AMD64Instr_SseCMov(cc ^ 1, r0, dst)); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3095 | return dst; |
| 3096 | } |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 3097 | |
| 3098 | ppIRExpr(e); |
| 3099 | vpanic("iselDblExpr_wrk"); |
| 3100 | } |
sewardj | c2bcb6f | 2005-02-07 00:17:12 +0000 | [diff] [blame] | 3101 | |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3102 | |
| 3103 | /*---------------------------------------------------------*/ |
| 3104 | /*--- ISEL: SIMD (Vector) expressions, 128 bit. ---*/ |
| 3105 | /*---------------------------------------------------------*/ |
| 3106 | |
| 3107 | static HReg iselVecExpr ( ISelEnv* env, IRExpr* e ) |
| 3108 | { |
| 3109 | HReg r = iselVecExpr_wrk( env, e ); |
| 3110 | # if 0 |
| 3111 | vex_printf("\n"); ppIRExpr(e); vex_printf("\n"); |
| 3112 | # endif |
| 3113 | vassert(hregClass(r) == HRcVec128); |
| 3114 | vassert(hregIsVirtual(r)); |
| 3115 | return r; |
| 3116 | } |
| 3117 | |
| 3118 | |
| 3119 | /* DO NOT CALL THIS DIRECTLY */ |
| 3120 | static HReg iselVecExpr_wrk ( ISelEnv* env, IRExpr* e ) |
| 3121 | { |
sewardj | 69d98e3 | 2010-06-18 08:17:41 +0000 | [diff] [blame] | 3122 | HWord fn = 0; /* address of helper fn, if required */ |
| 3123 | Bool arg1isEReg = False; |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3124 | AMD64SseOp op = Asse_INVALID; |
| 3125 | IRType ty = typeOfIRExpr(env->type_env,e); |
| 3126 | vassert(e); |
| 3127 | vassert(ty == Ity_V128); |
| 3128 | |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 3129 | if (e->tag == Iex_RdTmp) { |
| 3130 | return lookupIRTemp(env, e->Iex.RdTmp.tmp); |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3131 | } |
| 3132 | |
| 3133 | if (e->tag == Iex_Get) { |
| 3134 | HReg dst = newVRegV(env); |
| 3135 | addInstr(env, AMD64Instr_SseLdSt( |
| 3136 | True/*load*/, |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 3137 | 16, |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3138 | dst, |
| 3139 | AMD64AMode_IR(e->Iex.Get.offset, hregAMD64_RBP()) |
| 3140 | ) |
| 3141 | ); |
| 3142 | return dst; |
| 3143 | } |
| 3144 | |
sewardj | e768e92 | 2009-11-26 17:17:37 +0000 | [diff] [blame] | 3145 | if (e->tag == Iex_Load && e->Iex.Load.end == Iend_LE) { |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3146 | HReg dst = newVRegV(env); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 3147 | AMD64AMode* am = iselIntExpr_AMode(env, e->Iex.Load.addr); |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3148 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 16, dst, am )); |
| 3149 | return dst; |
| 3150 | } |
| 3151 | |
| 3152 | if (e->tag == Iex_Const) { |
| 3153 | HReg dst = newVRegV(env); |
| 3154 | vassert(e->Iex.Const.con->tag == Ico_V128); |
sewardj | 9ba870d | 2010-04-02 11:29:23 +0000 | [diff] [blame] | 3155 | switch (e->Iex.Const.con->Ico.V128) { |
| 3156 | case 0x0000: |
| 3157 | dst = generate_zeroes_V128(env); |
sewardj | acfbd7d | 2010-08-17 22:52:08 +0000 | [diff] [blame] | 3158 | break; |
sewardj | 9ba870d | 2010-04-02 11:29:23 +0000 | [diff] [blame] | 3159 | case 0xFFFF: |
| 3160 | dst = generate_ones_V128(env); |
sewardj | 9ba870d | 2010-04-02 11:29:23 +0000 | [diff] [blame] | 3161 | break; |
sewardj | acfbd7d | 2010-08-17 22:52:08 +0000 | [diff] [blame] | 3162 | default: { |
| 3163 | AMD64AMode* rsp0 = AMD64AMode_IR(0, hregAMD64_RSP()); |
| 3164 | /* do push_uimm64 twice, first time for the high-order half. */ |
| 3165 | push_uimm64(env, bitmask8_to_bytemask64( |
| 3166 | (e->Iex.Const.con->Ico.V128 >> 8) & 0xFF |
| 3167 | )); |
| 3168 | push_uimm64(env, bitmask8_to_bytemask64( |
| 3169 | (e->Iex.Const.con->Ico.V128 >> 0) & 0xFF |
| 3170 | )); |
| 3171 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 16, dst, rsp0 )); |
| 3172 | add_to_rsp(env, 16); |
| 3173 | break; |
| 3174 | } |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3175 | } |
sewardj | 9ba870d | 2010-04-02 11:29:23 +0000 | [diff] [blame] | 3176 | return dst; |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3177 | } |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3178 | |
| 3179 | if (e->tag == Iex_Unop) { |
| 3180 | switch (e->Iex.Unop.op) { |
| 3181 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3182 | case Iop_NotV128: { |
| 3183 | HReg arg = iselVecExpr(env, e->Iex.Unop.arg); |
| 3184 | return do_sse_NotV128(env, arg); |
| 3185 | } |
| 3186 | |
sewardj | 0971734 | 2005-05-05 21:34:02 +0000 | [diff] [blame] | 3187 | case Iop_CmpNEZ64x2: { |
| 3188 | /* We can use SSE2 instructions for this. */ |
| 3189 | /* Ideally, we want to do a 64Ix2 comparison against zero of |
| 3190 | the operand. Problem is no such insn exists. Solution |
| 3191 | therefore is to do a 32Ix4 comparison instead, and bitwise- |
| 3192 | negate (NOT) the result. Let a,b,c,d be 32-bit lanes, and |
| 3193 | let the not'd result of this initial comparison be a:b:c:d. |
| 3194 | What we need to compute is (a|b):(a|b):(c|d):(c|d). So, use |
| 3195 | pshufd to create a value b:a:d:c, and OR that with a:b:c:d, |
| 3196 | giving the required result. |
| 3197 | |
| 3198 | The required selection sequence is 2,3,0,1, which |
| 3199 | according to Intel's documentation means the pshufd |
| 3200 | literal value is 0xB1, that is, |
| 3201 | (2 << 6) | (3 << 4) | (0 << 2) | (1 << 0) |
| 3202 | */ |
| 3203 | HReg arg = iselVecExpr(env, e->Iex.Unop.arg); |
sewardj | ac53044 | 2005-05-11 16:13:37 +0000 | [diff] [blame] | 3204 | HReg tmp = generate_zeroes_V128(env); |
sewardj | 0971734 | 2005-05-05 21:34:02 +0000 | [diff] [blame] | 3205 | HReg dst = newVRegV(env); |
sewardj | 0971734 | 2005-05-05 21:34:02 +0000 | [diff] [blame] | 3206 | addInstr(env, AMD64Instr_SseReRg(Asse_CMPEQ32, arg, tmp)); |
| 3207 | tmp = do_sse_NotV128(env, tmp); |
| 3208 | addInstr(env, AMD64Instr_SseShuf(0xB1, tmp, dst)); |
| 3209 | addInstr(env, AMD64Instr_SseReRg(Asse_OR, tmp, dst)); |
| 3210 | return dst; |
| 3211 | } |
| 3212 | |
sewardj | ac53044 | 2005-05-11 16:13:37 +0000 | [diff] [blame] | 3213 | case Iop_CmpNEZ32x4: op = Asse_CMPEQ32; goto do_CmpNEZ_vector; |
| 3214 | case Iop_CmpNEZ16x8: op = Asse_CMPEQ16; goto do_CmpNEZ_vector; |
| 3215 | case Iop_CmpNEZ8x16: op = Asse_CMPEQ8; goto do_CmpNEZ_vector; |
| 3216 | do_CmpNEZ_vector: |
| 3217 | { |
| 3218 | HReg arg = iselVecExpr(env, e->Iex.Unop.arg); |
| 3219 | HReg tmp = newVRegV(env); |
| 3220 | HReg zero = generate_zeroes_V128(env); |
| 3221 | HReg dst; |
| 3222 | addInstr(env, mk_vMOVsd_RR(arg, tmp)); |
| 3223 | addInstr(env, AMD64Instr_SseReRg(op, zero, tmp)); |
| 3224 | dst = do_sse_NotV128(env, tmp); |
| 3225 | return dst; |
| 3226 | } |
sewardj | a7ba8c4 | 2005-05-10 20:08:34 +0000 | [diff] [blame] | 3227 | |
| 3228 | case Iop_Recip32Fx4: op = Asse_RCPF; goto do_32Fx4_unary; |
| 3229 | case Iop_RSqrt32Fx4: op = Asse_RSQRTF; goto do_32Fx4_unary; |
| 3230 | case Iop_Sqrt32Fx4: op = Asse_SQRTF; goto do_32Fx4_unary; |
| 3231 | do_32Fx4_unary: |
| 3232 | { |
| 3233 | HReg arg = iselVecExpr(env, e->Iex.Unop.arg); |
| 3234 | HReg dst = newVRegV(env); |
| 3235 | addInstr(env, AMD64Instr_Sse32Fx4(op, arg, dst)); |
| 3236 | return dst; |
| 3237 | } |
| 3238 | |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3239 | case Iop_Sqrt64Fx2: op = Asse_SQRTF; goto do_64Fx2_unary; |
| 3240 | do_64Fx2_unary: |
| 3241 | { |
| 3242 | HReg arg = iselVecExpr(env, e->Iex.Unop.arg); |
| 3243 | HReg dst = newVRegV(env); |
| 3244 | addInstr(env, AMD64Instr_Sse64Fx2(op, arg, dst)); |
| 3245 | return dst; |
| 3246 | } |
sewardj | a7ba8c4 | 2005-05-10 20:08:34 +0000 | [diff] [blame] | 3247 | |
| 3248 | case Iop_Recip32F0x4: op = Asse_RCPF; goto do_32F0x4_unary; |
| 3249 | case Iop_RSqrt32F0x4: op = Asse_RSQRTF; goto do_32F0x4_unary; |
| 3250 | case Iop_Sqrt32F0x4: op = Asse_SQRTF; goto do_32F0x4_unary; |
| 3251 | do_32F0x4_unary: |
| 3252 | { |
| 3253 | /* A bit subtle. We have to copy the arg to the result |
| 3254 | register first, because actually doing the SSE scalar insn |
| 3255 | leaves the upper 3/4 of the destination register |
| 3256 | unchanged. Whereas the required semantics of these |
| 3257 | primops is that the upper 3/4 is simply copied in from the |
| 3258 | argument. */ |
| 3259 | HReg arg = iselVecExpr(env, e->Iex.Unop.arg); |
| 3260 | HReg dst = newVRegV(env); |
| 3261 | addInstr(env, mk_vMOVsd_RR(arg, dst)); |
| 3262 | addInstr(env, AMD64Instr_Sse32FLo(op, arg, dst)); |
| 3263 | return dst; |
| 3264 | } |
| 3265 | |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3266 | case Iop_Sqrt64F0x2: op = Asse_SQRTF; goto do_64F0x2_unary; |
| 3267 | do_64F0x2_unary: |
| 3268 | { |
| 3269 | /* A bit subtle. We have to copy the arg to the result |
| 3270 | register first, because actually doing the SSE scalar insn |
| 3271 | leaves the upper half of the destination register |
| 3272 | unchanged. Whereas the required semantics of these |
| 3273 | primops is that the upper half is simply copied in from the |
| 3274 | argument. */ |
| 3275 | HReg arg = iselVecExpr(env, e->Iex.Unop.arg); |
| 3276 | HReg dst = newVRegV(env); |
| 3277 | addInstr(env, mk_vMOVsd_RR(arg, dst)); |
| 3278 | addInstr(env, AMD64Instr_Sse64FLo(op, arg, dst)); |
| 3279 | return dst; |
| 3280 | } |
| 3281 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3282 | case Iop_32UtoV128: { |
| 3283 | HReg dst = newVRegV(env); |
| 3284 | AMD64AMode* rsp_m32 = AMD64AMode_IR(-32, hregAMD64_RSP()); |
| 3285 | AMD64RI* ri = iselIntExpr_RI(env, e->Iex.Unop.arg); |
| 3286 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, ri, rsp_m32)); |
| 3287 | addInstr(env, AMD64Instr_SseLdzLO(4, dst, rsp_m32)); |
| 3288 | return dst; |
| 3289 | } |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3290 | |
| 3291 | case Iop_64UtoV128: { |
| 3292 | HReg dst = newVRegV(env); |
| 3293 | AMD64AMode* rsp0 = AMD64AMode_IR(0, hregAMD64_RSP()); |
| 3294 | AMD64RMI* rmi = iselIntExpr_RMI(env, e->Iex.Unop.arg); |
| 3295 | addInstr(env, AMD64Instr_Push(rmi)); |
| 3296 | addInstr(env, AMD64Instr_SseLdzLO(8, dst, rsp0)); |
| 3297 | add_to_rsp(env, 8); |
| 3298 | return dst; |
| 3299 | } |
| 3300 | |
sewardj | 4b1cc83 | 2012-06-13 11:10:20 +0000 | [diff] [blame] | 3301 | case Iop_V256toV128_0: |
| 3302 | case Iop_V256toV128_1: { |
| 3303 | HReg vHi, vLo; |
| 3304 | iselDVecExpr(&vHi, &vLo, env, e->Iex.Unop.arg); |
| 3305 | return (e->Iex.Unop.op == Iop_V256toV128_1) ? vHi : vLo; |
| 3306 | } |
| 3307 | |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3308 | default: |
| 3309 | break; |
| 3310 | } /* switch (e->Iex.Unop.op) */ |
| 3311 | } /* if (e->tag == Iex_Unop) */ |
| 3312 | |
| 3313 | if (e->tag == Iex_Binop) { |
| 3314 | switch (e->Iex.Binop.op) { |
| 3315 | |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3316 | /* FIXME: could we generate MOVQ here? */ |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 3317 | case Iop_SetV128lo64: { |
| 3318 | HReg dst = newVRegV(env); |
| 3319 | HReg srcV = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3320 | HReg srcI = iselIntExpr_R(env, e->Iex.Binop.arg2); |
sewardj | 478fe70 | 2005-04-23 01:15:47 +0000 | [diff] [blame] | 3321 | AMD64AMode* rsp_m16 = AMD64AMode_IR(-16, hregAMD64_RSP()); |
| 3322 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, srcV, rsp_m16)); |
| 3323 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, AMD64RI_Reg(srcI), rsp_m16)); |
| 3324 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, dst, rsp_m16)); |
| 3325 | return dst; |
| 3326 | } |
| 3327 | |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3328 | /* FIXME: could we generate MOVD here? */ |
sewardj | 478fe70 | 2005-04-23 01:15:47 +0000 | [diff] [blame] | 3329 | case Iop_SetV128lo32: { |
| 3330 | HReg dst = newVRegV(env); |
| 3331 | HReg srcV = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3332 | HReg srcI = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 3333 | AMD64AMode* rsp_m16 = AMD64AMode_IR(-16, hregAMD64_RSP()); |
| 3334 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, srcV, rsp_m16)); |
| 3335 | addInstr(env, AMD64Instr_Store(4, srcI, rsp_m16)); |
| 3336 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, dst, rsp_m16)); |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 3337 | return dst; |
| 3338 | } |
| 3339 | |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3340 | case Iop_64HLtoV128: { |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3341 | HReg rsp = hregAMD64_RSP(); |
| 3342 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, rsp); |
| 3343 | AMD64AMode* m16_rsp = AMD64AMode_IR(-16, rsp); |
| 3344 | AMD64RI* qHi = iselIntExpr_RI(env, e->Iex.Binop.arg1); |
| 3345 | AMD64RI* qLo = iselIntExpr_RI(env, e->Iex.Binop.arg2); |
| 3346 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, qHi, m8_rsp)); |
| 3347 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, qLo, m16_rsp)); |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3348 | HReg dst = newVRegV(env); |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3349 | /* One store-forwarding stall coming up, oh well :-( */ |
| 3350 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, dst, m16_rsp)); |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3351 | return dst; |
| 3352 | } |
| 3353 | |
sewardj | 432f8b6 | 2005-05-10 02:50:05 +0000 | [diff] [blame] | 3354 | case Iop_CmpEQ32Fx4: op = Asse_CMPEQF; goto do_32Fx4; |
| 3355 | case Iop_CmpLT32Fx4: op = Asse_CMPLTF; goto do_32Fx4; |
| 3356 | case Iop_CmpLE32Fx4: op = Asse_CMPLEF; goto do_32Fx4; |
sewardj | b928263 | 2005-11-05 02:33:25 +0000 | [diff] [blame] | 3357 | case Iop_CmpUN32Fx4: op = Asse_CMPUNF; goto do_32Fx4; |
sewardj | 432f8b6 | 2005-05-10 02:50:05 +0000 | [diff] [blame] | 3358 | case Iop_Add32Fx4: op = Asse_ADDF; goto do_32Fx4; |
| 3359 | case Iop_Div32Fx4: op = Asse_DIVF; goto do_32Fx4; |
| 3360 | case Iop_Max32Fx4: op = Asse_MAXF; goto do_32Fx4; |
| 3361 | case Iop_Min32Fx4: op = Asse_MINF; goto do_32Fx4; |
| 3362 | case Iop_Mul32Fx4: op = Asse_MULF; goto do_32Fx4; |
| 3363 | case Iop_Sub32Fx4: op = Asse_SUBF; goto do_32Fx4; |
| 3364 | do_32Fx4: |
| 3365 | { |
| 3366 | HReg argL = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3367 | HReg argR = iselVecExpr(env, e->Iex.Binop.arg2); |
| 3368 | HReg dst = newVRegV(env); |
| 3369 | addInstr(env, mk_vMOVsd_RR(argL, dst)); |
| 3370 | addInstr(env, AMD64Instr_Sse32Fx4(op, argR, dst)); |
| 3371 | return dst; |
| 3372 | } |
| 3373 | |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3374 | case Iop_CmpEQ64Fx2: op = Asse_CMPEQF; goto do_64Fx2; |
| 3375 | case Iop_CmpLT64Fx2: op = Asse_CMPLTF; goto do_64Fx2; |
| 3376 | case Iop_CmpLE64Fx2: op = Asse_CMPLEF; goto do_64Fx2; |
sewardj | b928263 | 2005-11-05 02:33:25 +0000 | [diff] [blame] | 3377 | case Iop_CmpUN64Fx2: op = Asse_CMPUNF; goto do_64Fx2; |
sewardj | 4c328cf | 2005-05-05 12:05:54 +0000 | [diff] [blame] | 3378 | case Iop_Add64Fx2: op = Asse_ADDF; goto do_64Fx2; |
sewardj | 5992bd0 | 2005-05-11 02:13:42 +0000 | [diff] [blame] | 3379 | case Iop_Div64Fx2: op = Asse_DIVF; goto do_64Fx2; |
| 3380 | case Iop_Max64Fx2: op = Asse_MAXF; goto do_64Fx2; |
| 3381 | case Iop_Min64Fx2: op = Asse_MINF; goto do_64Fx2; |
sewardj | 4c328cf | 2005-05-05 12:05:54 +0000 | [diff] [blame] | 3382 | case Iop_Mul64Fx2: op = Asse_MULF; goto do_64Fx2; |
| 3383 | case Iop_Sub64Fx2: op = Asse_SUBF; goto do_64Fx2; |
| 3384 | do_64Fx2: |
| 3385 | { |
| 3386 | HReg argL = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3387 | HReg argR = iselVecExpr(env, e->Iex.Binop.arg2); |
| 3388 | HReg dst = newVRegV(env); |
| 3389 | addInstr(env, mk_vMOVsd_RR(argL, dst)); |
| 3390 | addInstr(env, AMD64Instr_Sse64Fx2(op, argR, dst)); |
| 3391 | return dst; |
| 3392 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3393 | |
sewardj | 432f8b6 | 2005-05-10 02:50:05 +0000 | [diff] [blame] | 3394 | case Iop_CmpEQ32F0x4: op = Asse_CMPEQF; goto do_32F0x4; |
sewardj | 3aba9eb | 2005-03-30 23:20:47 +0000 | [diff] [blame] | 3395 | case Iop_CmpLT32F0x4: op = Asse_CMPLTF; goto do_32F0x4; |
sewardj | 4c328cf | 2005-05-05 12:05:54 +0000 | [diff] [blame] | 3396 | case Iop_CmpLE32F0x4: op = Asse_CMPLEF; goto do_32F0x4; |
sewardj | b928263 | 2005-11-05 02:33:25 +0000 | [diff] [blame] | 3397 | case Iop_CmpUN32F0x4: op = Asse_CMPUNF; goto do_32F0x4; |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3398 | case Iop_Add32F0x4: op = Asse_ADDF; goto do_32F0x4; |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 3399 | case Iop_Div32F0x4: op = Asse_DIVF; goto do_32F0x4; |
sewardj | 37d5257 | 2005-02-25 14:22:12 +0000 | [diff] [blame] | 3400 | case Iop_Max32F0x4: op = Asse_MAXF; goto do_32F0x4; |
| 3401 | case Iop_Min32F0x4: op = Asse_MINF; goto do_32F0x4; |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3402 | case Iop_Mul32F0x4: op = Asse_MULF; goto do_32F0x4; |
| 3403 | case Iop_Sub32F0x4: op = Asse_SUBF; goto do_32F0x4; |
| 3404 | do_32F0x4: { |
| 3405 | HReg argL = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3406 | HReg argR = iselVecExpr(env, e->Iex.Binop.arg2); |
| 3407 | HReg dst = newVRegV(env); |
| 3408 | addInstr(env, mk_vMOVsd_RR(argL, dst)); |
| 3409 | addInstr(env, AMD64Instr_Sse32FLo(op, argR, dst)); |
| 3410 | return dst; |
| 3411 | } |
| 3412 | |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 3413 | case Iop_CmpEQ64F0x2: op = Asse_CMPEQF; goto do_64F0x2; |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3414 | case Iop_CmpLT64F0x2: op = Asse_CMPLTF; goto do_64F0x2; |
sewardj | 137015d | 2005-03-27 04:01:15 +0000 | [diff] [blame] | 3415 | case Iop_CmpLE64F0x2: op = Asse_CMPLEF; goto do_64F0x2; |
sewardj | b928263 | 2005-11-05 02:33:25 +0000 | [diff] [blame] | 3416 | case Iop_CmpUN64F0x2: op = Asse_CMPUNF; goto do_64F0x2; |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3417 | case Iop_Add64F0x2: op = Asse_ADDF; goto do_64F0x2; |
| 3418 | case Iop_Div64F0x2: op = Asse_DIVF; goto do_64F0x2; |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3419 | case Iop_Max64F0x2: op = Asse_MAXF; goto do_64F0x2; |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 3420 | case Iop_Min64F0x2: op = Asse_MINF; goto do_64F0x2; |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3421 | case Iop_Mul64F0x2: op = Asse_MULF; goto do_64F0x2; |
| 3422 | case Iop_Sub64F0x2: op = Asse_SUBF; goto do_64F0x2; |
| 3423 | do_64F0x2: { |
| 3424 | HReg argL = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3425 | HReg argR = iselVecExpr(env, e->Iex.Binop.arg2); |
| 3426 | HReg dst = newVRegV(env); |
| 3427 | addInstr(env, mk_vMOVsd_RR(argL, dst)); |
| 3428 | addInstr(env, AMD64Instr_Sse64FLo(op, argR, dst)); |
| 3429 | return dst; |
| 3430 | } |
| 3431 | |
sewardj | 5f438dd | 2011-06-16 11:36:23 +0000 | [diff] [blame] | 3432 | case Iop_QNarrowBin32Sto16Sx8: |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3433 | op = Asse_PACKSSD; arg1isEReg = True; goto do_SseReRg; |
sewardj | 5f438dd | 2011-06-16 11:36:23 +0000 | [diff] [blame] | 3434 | case Iop_QNarrowBin16Sto8Sx16: |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3435 | op = Asse_PACKSSW; arg1isEReg = True; goto do_SseReRg; |
sewardj | 5f438dd | 2011-06-16 11:36:23 +0000 | [diff] [blame] | 3436 | case Iop_QNarrowBin16Sto8Ux16: |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3437 | op = Asse_PACKUSW; arg1isEReg = True; goto do_SseReRg; |
| 3438 | |
| 3439 | case Iop_InterleaveHI8x16: |
| 3440 | op = Asse_UNPCKHB; arg1isEReg = True; goto do_SseReRg; |
| 3441 | case Iop_InterleaveHI16x8: |
| 3442 | op = Asse_UNPCKHW; arg1isEReg = True; goto do_SseReRg; |
| 3443 | case Iop_InterleaveHI32x4: |
| 3444 | op = Asse_UNPCKHD; arg1isEReg = True; goto do_SseReRg; |
| 3445 | case Iop_InterleaveHI64x2: |
| 3446 | op = Asse_UNPCKHQ; arg1isEReg = True; goto do_SseReRg; |
| 3447 | |
| 3448 | case Iop_InterleaveLO8x16: |
| 3449 | op = Asse_UNPCKLB; arg1isEReg = True; goto do_SseReRg; |
| 3450 | case Iop_InterleaveLO16x8: |
| 3451 | op = Asse_UNPCKLW; arg1isEReg = True; goto do_SseReRg; |
| 3452 | case Iop_InterleaveLO32x4: |
| 3453 | op = Asse_UNPCKLD; arg1isEReg = True; goto do_SseReRg; |
| 3454 | case Iop_InterleaveLO64x2: |
| 3455 | op = Asse_UNPCKLQ; arg1isEReg = True; goto do_SseReRg; |
| 3456 | |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 3457 | case Iop_AndV128: op = Asse_AND; goto do_SseReRg; |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 3458 | case Iop_OrV128: op = Asse_OR; goto do_SseReRg; |
sewardj | 9da1697 | 2005-02-21 13:58:26 +0000 | [diff] [blame] | 3459 | case Iop_XorV128: op = Asse_XOR; goto do_SseReRg; |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3460 | case Iop_Add8x16: op = Asse_ADD8; goto do_SseReRg; |
sewardj | 5992bd0 | 2005-05-11 02:13:42 +0000 | [diff] [blame] | 3461 | case Iop_Add16x8: op = Asse_ADD16; goto do_SseReRg; |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3462 | case Iop_Add32x4: op = Asse_ADD32; goto do_SseReRg; |
sewardj | 0971734 | 2005-05-05 21:34:02 +0000 | [diff] [blame] | 3463 | case Iop_Add64x2: op = Asse_ADD64; goto do_SseReRg; |
sewardj | 5992bd0 | 2005-05-11 02:13:42 +0000 | [diff] [blame] | 3464 | case Iop_QAdd8Sx16: op = Asse_QADD8S; goto do_SseReRg; |
| 3465 | case Iop_QAdd16Sx8: op = Asse_QADD16S; goto do_SseReRg; |
| 3466 | case Iop_QAdd8Ux16: op = Asse_QADD8U; goto do_SseReRg; |
| 3467 | case Iop_QAdd16Ux8: op = Asse_QADD16U; goto do_SseReRg; |
| 3468 | case Iop_Avg8Ux16: op = Asse_AVG8U; goto do_SseReRg; |
| 3469 | case Iop_Avg16Ux8: op = Asse_AVG16U; goto do_SseReRg; |
| 3470 | case Iop_CmpEQ8x16: op = Asse_CMPEQ8; goto do_SseReRg; |
| 3471 | case Iop_CmpEQ16x8: op = Asse_CMPEQ16; goto do_SseReRg; |
| 3472 | case Iop_CmpEQ32x4: op = Asse_CMPEQ32; goto do_SseReRg; |
| 3473 | case Iop_CmpGT8Sx16: op = Asse_CMPGT8S; goto do_SseReRg; |
| 3474 | case Iop_CmpGT16Sx8: op = Asse_CMPGT16S; goto do_SseReRg; |
| 3475 | case Iop_CmpGT32Sx4: op = Asse_CMPGT32S; goto do_SseReRg; |
sewardj | adffcef | 2005-05-11 00:03:06 +0000 | [diff] [blame] | 3476 | case Iop_Max16Sx8: op = Asse_MAX16S; goto do_SseReRg; |
| 3477 | case Iop_Max8Ux16: op = Asse_MAX8U; goto do_SseReRg; |
| 3478 | case Iop_Min16Sx8: op = Asse_MIN16S; goto do_SseReRg; |
| 3479 | case Iop_Min8Ux16: op = Asse_MIN8U; goto do_SseReRg; |
| 3480 | case Iop_MulHi16Ux8: op = Asse_MULHI16U; goto do_SseReRg; |
| 3481 | case Iop_MulHi16Sx8: op = Asse_MULHI16S; goto do_SseReRg; |
| 3482 | case Iop_Mul16x8: op = Asse_MUL16; goto do_SseReRg; |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3483 | case Iop_Sub8x16: op = Asse_SUB8; goto do_SseReRg; |
| 3484 | case Iop_Sub16x8: op = Asse_SUB16; goto do_SseReRg; |
| 3485 | case Iop_Sub32x4: op = Asse_SUB32; goto do_SseReRg; |
sewardj | 0971734 | 2005-05-05 21:34:02 +0000 | [diff] [blame] | 3486 | case Iop_Sub64x2: op = Asse_SUB64; goto do_SseReRg; |
sewardj | 9762859 | 2005-05-10 22:42:54 +0000 | [diff] [blame] | 3487 | case Iop_QSub8Sx16: op = Asse_QSUB8S; goto do_SseReRg; |
| 3488 | case Iop_QSub16Sx8: op = Asse_QSUB16S; goto do_SseReRg; |
| 3489 | case Iop_QSub8Ux16: op = Asse_QSUB8U; goto do_SseReRg; |
| 3490 | case Iop_QSub16Ux8: op = Asse_QSUB16U; goto do_SseReRg; |
sewardj | 9da1697 | 2005-02-21 13:58:26 +0000 | [diff] [blame] | 3491 | do_SseReRg: { |
| 3492 | HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3493 | HReg arg2 = iselVecExpr(env, e->Iex.Binop.arg2); |
| 3494 | HReg dst = newVRegV(env); |
| 3495 | if (arg1isEReg) { |
sewardj | 9da1697 | 2005-02-21 13:58:26 +0000 | [diff] [blame] | 3496 | addInstr(env, mk_vMOVsd_RR(arg2, dst)); |
| 3497 | addInstr(env, AMD64Instr_SseReRg(op, arg1, dst)); |
| 3498 | } else { |
| 3499 | addInstr(env, mk_vMOVsd_RR(arg1, dst)); |
| 3500 | addInstr(env, AMD64Instr_SseReRg(op, arg2, dst)); |
| 3501 | } |
| 3502 | return dst; |
| 3503 | } |
| 3504 | |
sewardj | adffcef | 2005-05-11 00:03:06 +0000 | [diff] [blame] | 3505 | case Iop_ShlN16x8: op = Asse_SHL16; goto do_SseShift; |
| 3506 | case Iop_ShlN32x4: op = Asse_SHL32; goto do_SseShift; |
| 3507 | case Iop_ShlN64x2: op = Asse_SHL64; goto do_SseShift; |
| 3508 | case Iop_SarN16x8: op = Asse_SAR16; goto do_SseShift; |
| 3509 | case Iop_SarN32x4: op = Asse_SAR32; goto do_SseShift; |
| 3510 | case Iop_ShrN16x8: op = Asse_SHR16; goto do_SseShift; |
| 3511 | case Iop_ShrN32x4: op = Asse_SHR32; goto do_SseShift; |
sewardj | 0971734 | 2005-05-05 21:34:02 +0000 | [diff] [blame] | 3512 | case Iop_ShrN64x2: op = Asse_SHR64; goto do_SseShift; |
| 3513 | do_SseShift: { |
| 3514 | HReg greg = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3515 | AMD64RMI* rmi = iselIntExpr_RMI(env, e->Iex.Binop.arg2); |
| 3516 | AMD64AMode* rsp0 = AMD64AMode_IR(0, hregAMD64_RSP()); |
| 3517 | HReg ereg = newVRegV(env); |
| 3518 | HReg dst = newVRegV(env); |
| 3519 | addInstr(env, AMD64Instr_Push(AMD64RMI_Imm(0))); |
| 3520 | addInstr(env, AMD64Instr_Push(rmi)); |
| 3521 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, ereg, rsp0)); |
| 3522 | addInstr(env, mk_vMOVsd_RR(greg, dst)); |
| 3523 | addInstr(env, AMD64Instr_SseReRg(op, ereg, dst)); |
| 3524 | add_to_rsp(env, 16); |
| 3525 | return dst; |
| 3526 | } |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3527 | |
sewardj | 69d98e3 | 2010-06-18 08:17:41 +0000 | [diff] [blame] | 3528 | case Iop_Mul32x4: fn = (HWord)h_generic_calc_Mul32x4; |
| 3529 | goto do_SseAssistedBinary; |
| 3530 | case Iop_Max32Sx4: fn = (HWord)h_generic_calc_Max32Sx4; |
| 3531 | goto do_SseAssistedBinary; |
| 3532 | case Iop_Min32Sx4: fn = (HWord)h_generic_calc_Min32Sx4; |
| 3533 | goto do_SseAssistedBinary; |
| 3534 | case Iop_Max32Ux4: fn = (HWord)h_generic_calc_Max32Ux4; |
| 3535 | goto do_SseAssistedBinary; |
| 3536 | case Iop_Min32Ux4: fn = (HWord)h_generic_calc_Min32Ux4; |
| 3537 | goto do_SseAssistedBinary; |
| 3538 | case Iop_Max16Ux8: fn = (HWord)h_generic_calc_Max16Ux8; |
| 3539 | goto do_SseAssistedBinary; |
| 3540 | case Iop_Min16Ux8: fn = (HWord)h_generic_calc_Min16Ux8; |
| 3541 | goto do_SseAssistedBinary; |
| 3542 | case Iop_Max8Sx16: fn = (HWord)h_generic_calc_Max8Sx16; |
| 3543 | goto do_SseAssistedBinary; |
| 3544 | case Iop_Min8Sx16: fn = (HWord)h_generic_calc_Min8Sx16; |
| 3545 | goto do_SseAssistedBinary; |
sewardj | d881562 | 2011-10-19 15:24:01 +0000 | [diff] [blame] | 3546 | case Iop_CmpEQ64x2: fn = (HWord)h_generic_calc_CmpEQ64x2; |
| 3547 | goto do_SseAssistedBinary; |
sewardj | 69d98e3 | 2010-06-18 08:17:41 +0000 | [diff] [blame] | 3548 | case Iop_CmpGT64Sx2: fn = (HWord)h_generic_calc_CmpGT64Sx2; |
| 3549 | goto do_SseAssistedBinary; |
sewardj | d8bca7e | 2012-06-20 11:46:19 +0000 | [diff] [blame] | 3550 | case Iop_Perm32x4: fn = (HWord)h_generic_calc_Perm32x4; |
| 3551 | goto do_SseAssistedBinary; |
sewardj | 5f438dd | 2011-06-16 11:36:23 +0000 | [diff] [blame] | 3552 | case Iop_QNarrowBin32Sto16Ux8: |
| 3553 | fn = (HWord)h_generic_calc_QNarrowBin32Sto16Ux8; |
sewardj | 2260b99 | 2011-06-15 16:05:07 +0000 | [diff] [blame] | 3554 | goto do_SseAssistedBinary; |
sewardj | ad2c9ea | 2011-10-22 09:32:16 +0000 | [diff] [blame] | 3555 | case Iop_NarrowBin16to8x16: |
| 3556 | fn = (HWord)h_generic_calc_NarrowBin16to8x16; |
| 3557 | goto do_SseAssistedBinary; |
| 3558 | case Iop_NarrowBin32to16x8: |
| 3559 | fn = (HWord)h_generic_calc_NarrowBin32to16x8; |
| 3560 | goto do_SseAssistedBinary; |
sewardj | 69d98e3 | 2010-06-18 08:17:41 +0000 | [diff] [blame] | 3561 | do_SseAssistedBinary: { |
| 3562 | /* RRRufff! RRRufff code is what we're generating here. Oh |
| 3563 | well. */ |
| 3564 | vassert(fn != 0); |
| 3565 | HReg dst = newVRegV(env); |
| 3566 | HReg argL = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3567 | HReg argR = iselVecExpr(env, e->Iex.Binop.arg2); |
| 3568 | HReg argp = newVRegI(env); |
| 3569 | /* subq $112, %rsp -- make a space*/ |
| 3570 | sub_from_rsp(env, 112); |
| 3571 | /* leaq 48(%rsp), %r_argp -- point into it */ |
| 3572 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(48, hregAMD64_RSP()), |
| 3573 | argp)); |
| 3574 | /* andq $-16, %r_argp -- 16-align the pointer */ |
| 3575 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, |
| 3576 | AMD64RMI_Imm( ~(UInt)15 ), |
| 3577 | argp)); |
| 3578 | /* Prepare 3 arg regs: |
| 3579 | leaq 0(%r_argp), %rdi |
| 3580 | leaq 16(%r_argp), %rsi |
| 3581 | leaq 32(%r_argp), %rdx |
| 3582 | */ |
| 3583 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(0, argp), |
| 3584 | hregAMD64_RDI())); |
| 3585 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(16, argp), |
| 3586 | hregAMD64_RSI())); |
| 3587 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(32, argp), |
| 3588 | hregAMD64_RDX())); |
| 3589 | /* Store the two args, at (%rsi) and (%rdx): |
| 3590 | movupd %argL, 0(%rsi) |
| 3591 | movupd %argR, 0(%rdx) |
| 3592 | */ |
| 3593 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argL, |
| 3594 | AMD64AMode_IR(0, hregAMD64_RSI()))); |
| 3595 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argR, |
| 3596 | AMD64AMode_IR(0, hregAMD64_RDX()))); |
| 3597 | /* call the helper */ |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 3598 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 3599 | 3, mk_RetLoc_simple(RLPri_None) )); |
sewardj | 69d98e3 | 2010-06-18 08:17:41 +0000 | [diff] [blame] | 3600 | /* fetch the result from memory, using %r_argp, which the |
| 3601 | register allocator will keep alive across the call. */ |
| 3602 | addInstr(env, AMD64Instr_SseLdSt(True/*isLoad*/, 16, dst, |
| 3603 | AMD64AMode_IR(0, argp))); |
| 3604 | /* and finally, clear the space */ |
| 3605 | add_to_rsp(env, 112); |
| 3606 | return dst; |
| 3607 | } |
| 3608 | |
sewardj | 0874bee | 2011-01-17 10:32:18 +0000 | [diff] [blame] | 3609 | case Iop_SarN64x2: fn = (HWord)h_generic_calc_SarN64x2; |
| 3610 | goto do_SseAssistedVectorAndScalar; |
| 3611 | case Iop_SarN8x16: fn = (HWord)h_generic_calc_SarN8x16; |
| 3612 | goto do_SseAssistedVectorAndScalar; |
| 3613 | do_SseAssistedVectorAndScalar: { |
| 3614 | /* RRRufff! RRRufff code is what we're generating here. Oh |
| 3615 | well. */ |
| 3616 | vassert(fn != 0); |
| 3617 | HReg dst = newVRegV(env); |
| 3618 | HReg argL = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3619 | HReg argR = iselIntExpr_R(env, e->Iex.Binop.arg2); |
| 3620 | HReg argp = newVRegI(env); |
| 3621 | /* subq $112, %rsp -- make a space*/ |
| 3622 | sub_from_rsp(env, 112); |
| 3623 | /* leaq 48(%rsp), %r_argp -- point into it */ |
| 3624 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(48, hregAMD64_RSP()), |
| 3625 | argp)); |
| 3626 | /* andq $-16, %r_argp -- 16-align the pointer */ |
| 3627 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, |
| 3628 | AMD64RMI_Imm( ~(UInt)15 ), |
| 3629 | argp)); |
| 3630 | /* Prepare 2 vector arg regs: |
| 3631 | leaq 0(%r_argp), %rdi |
| 3632 | leaq 16(%r_argp), %rsi |
| 3633 | */ |
| 3634 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(0, argp), |
| 3635 | hregAMD64_RDI())); |
| 3636 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(16, argp), |
| 3637 | hregAMD64_RSI())); |
| 3638 | /* Store the vector arg, at (%rsi): |
| 3639 | movupd %argL, 0(%rsi) |
| 3640 | */ |
| 3641 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argL, |
| 3642 | AMD64AMode_IR(0, hregAMD64_RSI()))); |
| 3643 | /* And get the scalar value into rdx */ |
| 3644 | addInstr(env, mk_iMOVsd_RR(argR, hregAMD64_RDX())); |
| 3645 | |
| 3646 | /* call the helper */ |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 3647 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 3648 | 3, mk_RetLoc_simple(RLPri_None) )); |
sewardj | 0874bee | 2011-01-17 10:32:18 +0000 | [diff] [blame] | 3649 | /* fetch the result from memory, using %r_argp, which the |
| 3650 | register allocator will keep alive across the call. */ |
| 3651 | addInstr(env, AMD64Instr_SseLdSt(True/*isLoad*/, 16, dst, |
| 3652 | AMD64AMode_IR(0, argp))); |
| 3653 | /* and finally, clear the space */ |
| 3654 | add_to_rsp(env, 112); |
| 3655 | return dst; |
| 3656 | } |
| 3657 | |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3658 | default: |
| 3659 | break; |
| 3660 | } /* switch (e->Iex.Binop.op) */ |
| 3661 | } /* if (e->tag == Iex_Binop) */ |
| 3662 | |
florian | 99dd03e | 2013-01-29 03:56:06 +0000 | [diff] [blame] | 3663 | if (e->tag == Iex_ITE) { // VFD |
| 3664 | HReg r1 = iselVecExpr(env, e->Iex.ITE.iftrue); |
| 3665 | HReg r0 = iselVecExpr(env, e->Iex.ITE.iffalse); |
sewardj | adffcef | 2005-05-11 00:03:06 +0000 | [diff] [blame] | 3666 | HReg dst = newVRegV(env); |
florian | 99dd03e | 2013-01-29 03:56:06 +0000 | [diff] [blame] | 3667 | addInstr(env, mk_vMOVsd_RR(r1,dst)); |
florian | e6be61f | 2013-02-01 16:11:51 +0000 | [diff] [blame] | 3668 | AMD64CondCode cc = iselCondCode(env, e->Iex.ITE.cond); |
sewardj | 009230b | 2013-01-26 11:47:55 +0000 | [diff] [blame] | 3669 | addInstr(env, AMD64Instr_SseCMov(cc ^ 1, r0, dst)); |
sewardj | adffcef | 2005-05-11 00:03:06 +0000 | [diff] [blame] | 3670 | return dst; |
| 3671 | } |
| 3672 | |
sewardj | acfbd7d | 2010-08-17 22:52:08 +0000 | [diff] [blame] | 3673 | //vec_fail: |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3674 | vex_printf("iselVecExpr (amd64, subarch = %s): can't reduce\n", |
sewardj | 5117ce1 | 2006-01-27 21:20:15 +0000 | [diff] [blame] | 3675 | LibVEX_ppVexHwCaps(VexArchAMD64, env->hwcaps)); |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 3676 | ppIRExpr(e); |
| 3677 | vpanic("iselVecExpr_wrk"); |
| 3678 | } |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 3679 | |
| 3680 | |
| 3681 | /*---------------------------------------------------------*/ |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3682 | /*--- ISEL: SIMD (V256) expressions, into 2 XMM regs. --*/ |
| 3683 | /*---------------------------------------------------------*/ |
| 3684 | |
sewardj | 56c3031 | 2012-06-12 08:45:39 +0000 | [diff] [blame] | 3685 | static void iselDVecExpr ( /*OUT*/HReg* rHi, /*OUT*/HReg* rLo, |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3686 | ISelEnv* env, IRExpr* e ) |
| 3687 | { |
| 3688 | iselDVecExpr_wrk( rHi, rLo, env, e ); |
| 3689 | # if 0 |
| 3690 | vex_printf("\n"); ppIRExpr(e); vex_printf("\n"); |
| 3691 | # endif |
| 3692 | vassert(hregClass(*rHi) == HRcVec128); |
| 3693 | vassert(hregClass(*rLo) == HRcVec128); |
| 3694 | vassert(hregIsVirtual(*rHi)); |
| 3695 | vassert(hregIsVirtual(*rLo)); |
| 3696 | } |
| 3697 | |
| 3698 | |
| 3699 | /* DO NOT CALL THIS DIRECTLY */ |
sewardj | 56c3031 | 2012-06-12 08:45:39 +0000 | [diff] [blame] | 3700 | static void iselDVecExpr_wrk ( /*OUT*/HReg* rHi, /*OUT*/HReg* rLo, |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3701 | ISelEnv* env, IRExpr* e ) |
| 3702 | { |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 3703 | HWord fn = 0; /* address of helper fn, if required */ |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3704 | vassert(e); |
| 3705 | IRType ty = typeOfIRExpr(env->type_env,e); |
| 3706 | vassert(ty == Ity_V256); |
| 3707 | |
sewardj | 56c3031 | 2012-06-12 08:45:39 +0000 | [diff] [blame] | 3708 | AMD64SseOp op = Asse_INVALID; |
| 3709 | |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 3710 | /* read 256-bit IRTemp */ |
| 3711 | if (e->tag == Iex_RdTmp) { |
| 3712 | lookupIRTempPair( rHi, rLo, env, e->Iex.RdTmp.tmp); |
| 3713 | return; |
| 3714 | } |
| 3715 | |
| 3716 | if (e->tag == Iex_Get) { |
| 3717 | HReg vHi = newVRegV(env); |
| 3718 | HReg vLo = newVRegV(env); |
| 3719 | HReg rbp = hregAMD64_RBP(); |
| 3720 | AMD64AMode* am0 = AMD64AMode_IR(e->Iex.Get.offset + 0, rbp); |
| 3721 | AMD64AMode* am16 = AMD64AMode_IR(e->Iex.Get.offset + 16, rbp); |
| 3722 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, vLo, am0)); |
| 3723 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, vHi, am16)); |
| 3724 | *rHi = vHi; |
| 3725 | *rLo = vLo; |
| 3726 | return; |
| 3727 | } |
| 3728 | |
| 3729 | if (e->tag == Iex_Load) { |
| 3730 | HReg vHi = newVRegV(env); |
| 3731 | HReg vLo = newVRegV(env); |
| 3732 | HReg rA = iselIntExpr_R(env, e->Iex.Load.addr); |
| 3733 | AMD64AMode* am0 = AMD64AMode_IR(0, rA); |
| 3734 | AMD64AMode* am16 = AMD64AMode_IR(16, rA); |
| 3735 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, vLo, am0)); |
| 3736 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, vHi, am16)); |
| 3737 | *rHi = vHi; |
| 3738 | *rLo = vLo; |
| 3739 | return; |
| 3740 | } |
| 3741 | |
sewardj | 37a505b | 2012-06-29 15:28:24 +0000 | [diff] [blame] | 3742 | if (e->tag == Iex_Const) { |
| 3743 | vassert(e->Iex.Const.con->tag == Ico_V256); |
| 3744 | switch (e->Iex.Const.con->Ico.V256) { |
| 3745 | case 0x00000000: { |
| 3746 | HReg vHi = generate_zeroes_V128(env); |
| 3747 | HReg vLo = newVRegV(env); |
| 3748 | addInstr(env, mk_vMOVsd_RR(vHi, vLo)); |
| 3749 | *rHi = vHi; |
| 3750 | *rLo = vLo; |
| 3751 | return; |
| 3752 | } |
| 3753 | default: |
| 3754 | break; /* give up. Until such time as is necessary. */ |
| 3755 | } |
| 3756 | } |
| 3757 | |
sewardj | 2a2bda9 | 2012-06-14 23:32:02 +0000 | [diff] [blame] | 3758 | if (e->tag == Iex_Unop) { |
| 3759 | switch (e->Iex.Unop.op) { |
| 3760 | |
| 3761 | case Iop_NotV256: { |
| 3762 | HReg argHi, argLo; |
| 3763 | iselDVecExpr(&argHi, &argLo, env, e->Iex.Unop.arg); |
| 3764 | *rHi = do_sse_NotV128(env, argHi); |
| 3765 | *rLo = do_sse_NotV128(env, argLo); |
| 3766 | return; |
| 3767 | } |
| 3768 | |
sewardj | 8209692 | 2012-06-24 14:57:59 +0000 | [diff] [blame] | 3769 | case Iop_Recip32Fx8: op = Asse_RCPF; goto do_32Fx8_unary; |
sewardj | 66becf3 | 2012-06-18 23:15:16 +0000 | [diff] [blame] | 3770 | case Iop_Sqrt32Fx8: op = Asse_SQRTF; goto do_32Fx8_unary; |
| 3771 | case Iop_RSqrt32Fx8: op = Asse_RSQRTF; goto do_32Fx8_unary; |
| 3772 | do_32Fx8_unary: |
| 3773 | { |
| 3774 | HReg argHi, argLo; |
| 3775 | iselDVecExpr(&argHi, &argLo, env, e->Iex.Unop.arg); |
| 3776 | HReg dstHi = newVRegV(env); |
| 3777 | HReg dstLo = newVRegV(env); |
| 3778 | addInstr(env, AMD64Instr_Sse32Fx4(op, argHi, dstHi)); |
| 3779 | addInstr(env, AMD64Instr_Sse32Fx4(op, argLo, dstLo)); |
| 3780 | *rHi = dstHi; |
| 3781 | *rLo = dstLo; |
| 3782 | return; |
| 3783 | } |
| 3784 | |
| 3785 | case Iop_Sqrt64Fx4: op = Asse_SQRTF; goto do_64Fx4_unary; |
| 3786 | do_64Fx4_unary: |
| 3787 | { |
| 3788 | HReg argHi, argLo; |
| 3789 | iselDVecExpr(&argHi, &argLo, env, e->Iex.Unop.arg); |
| 3790 | HReg dstHi = newVRegV(env); |
| 3791 | HReg dstLo = newVRegV(env); |
| 3792 | addInstr(env, AMD64Instr_Sse64Fx2(op, argHi, dstHi)); |
| 3793 | addInstr(env, AMD64Instr_Sse64Fx2(op, argLo, dstLo)); |
| 3794 | *rHi = dstHi; |
| 3795 | *rLo = dstLo; |
| 3796 | return; |
| 3797 | } |
| 3798 | |
sewardj | 23db8a0 | 2012-06-25 07:46:18 +0000 | [diff] [blame] | 3799 | case Iop_CmpNEZ64x4: { |
| 3800 | /* We can use SSE2 instructions for this. */ |
| 3801 | /* Same scheme as Iop_CmpNEZ64x2, except twice as wide |
| 3802 | (obviously). See comment on Iop_CmpNEZ64x2 for |
| 3803 | explanation of what's going on here. */ |
| 3804 | HReg argHi, argLo; |
| 3805 | iselDVecExpr(&argHi, &argLo, env, e->Iex.Unop.arg); |
| 3806 | HReg tmpHi = generate_zeroes_V128(env); |
| 3807 | HReg tmpLo = newVRegV(env); |
| 3808 | addInstr(env, mk_vMOVsd_RR(tmpHi, tmpLo)); |
| 3809 | HReg dstHi = newVRegV(env); |
| 3810 | HReg dstLo = newVRegV(env); |
| 3811 | addInstr(env, AMD64Instr_SseReRg(Asse_CMPEQ32, argHi, tmpHi)); |
| 3812 | addInstr(env, AMD64Instr_SseReRg(Asse_CMPEQ32, argLo, tmpLo)); |
| 3813 | tmpHi = do_sse_NotV128(env, tmpHi); |
| 3814 | tmpLo = do_sse_NotV128(env, tmpLo); |
| 3815 | addInstr(env, AMD64Instr_SseShuf(0xB1, tmpHi, dstHi)); |
| 3816 | addInstr(env, AMD64Instr_SseShuf(0xB1, tmpLo, dstLo)); |
| 3817 | addInstr(env, AMD64Instr_SseReRg(Asse_OR, tmpHi, dstHi)); |
| 3818 | addInstr(env, AMD64Instr_SseReRg(Asse_OR, tmpLo, dstLo)); |
| 3819 | *rHi = dstHi; |
| 3820 | *rLo = dstLo; |
| 3821 | return; |
| 3822 | } |
| 3823 | |
| 3824 | case Iop_CmpNEZ32x8: op = Asse_CMPEQ32; goto do_CmpNEZ_vector; |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 3825 | case Iop_CmpNEZ16x16: op = Asse_CMPEQ16; goto do_CmpNEZ_vector; |
| 3826 | case Iop_CmpNEZ8x32: op = Asse_CMPEQ8; goto do_CmpNEZ_vector; |
sewardj | 23db8a0 | 2012-06-25 07:46:18 +0000 | [diff] [blame] | 3827 | do_CmpNEZ_vector: |
| 3828 | { |
| 3829 | HReg argHi, argLo; |
| 3830 | iselDVecExpr(&argHi, &argLo, env, e->Iex.Unop.arg); |
| 3831 | HReg tmpHi = newVRegV(env); |
| 3832 | HReg tmpLo = newVRegV(env); |
| 3833 | HReg zero = generate_zeroes_V128(env); |
| 3834 | HReg dstHi, dstLo; |
| 3835 | addInstr(env, mk_vMOVsd_RR(argHi, tmpHi)); |
| 3836 | addInstr(env, mk_vMOVsd_RR(argLo, tmpLo)); |
| 3837 | addInstr(env, AMD64Instr_SseReRg(op, zero, tmpHi)); |
| 3838 | addInstr(env, AMD64Instr_SseReRg(op, zero, tmpLo)); |
| 3839 | dstHi = do_sse_NotV128(env, tmpHi); |
| 3840 | dstLo = do_sse_NotV128(env, tmpLo); |
| 3841 | *rHi = dstHi; |
| 3842 | *rLo = dstLo; |
| 3843 | return; |
| 3844 | } |
| 3845 | |
sewardj | 2a2bda9 | 2012-06-14 23:32:02 +0000 | [diff] [blame] | 3846 | default: |
| 3847 | break; |
| 3848 | } /* switch (e->Iex.Unop.op) */ |
| 3849 | } /* if (e->tag == Iex_Unop) */ |
| 3850 | |
sewardj | 56c3031 | 2012-06-12 08:45:39 +0000 | [diff] [blame] | 3851 | if (e->tag == Iex_Binop) { |
| 3852 | switch (e->Iex.Binop.op) { |
| 3853 | |
| 3854 | case Iop_Add64Fx4: op = Asse_ADDF; goto do_64Fx4; |
| 3855 | case Iop_Sub64Fx4: op = Asse_SUBF; goto do_64Fx4; |
| 3856 | case Iop_Mul64Fx4: op = Asse_MULF; goto do_64Fx4; |
| 3857 | case Iop_Div64Fx4: op = Asse_DIVF; goto do_64Fx4; |
sewardj | 8eb7ae8 | 2012-06-24 14:00:27 +0000 | [diff] [blame] | 3858 | case Iop_Max64Fx4: op = Asse_MAXF; goto do_64Fx4; |
| 3859 | case Iop_Min64Fx4: op = Asse_MINF; goto do_64Fx4; |
sewardj | 56c3031 | 2012-06-12 08:45:39 +0000 | [diff] [blame] | 3860 | do_64Fx4: |
| 3861 | { |
| 3862 | HReg argLhi, argLlo, argRhi, argRlo; |
| 3863 | iselDVecExpr(&argLhi, &argLlo, env, e->Iex.Binop.arg1); |
| 3864 | iselDVecExpr(&argRhi, &argRlo, env, e->Iex.Binop.arg2); |
| 3865 | HReg dstHi = newVRegV(env); |
| 3866 | HReg dstLo = newVRegV(env); |
| 3867 | addInstr(env, mk_vMOVsd_RR(argLhi, dstHi)); |
| 3868 | addInstr(env, mk_vMOVsd_RR(argLlo, dstLo)); |
| 3869 | addInstr(env, AMD64Instr_Sse64Fx2(op, argRhi, dstHi)); |
| 3870 | addInstr(env, AMD64Instr_Sse64Fx2(op, argRlo, dstLo)); |
| 3871 | *rHi = dstHi; |
| 3872 | *rLo = dstLo; |
| 3873 | return; |
| 3874 | } |
| 3875 | |
| 3876 | case Iop_Add32Fx8: op = Asse_ADDF; goto do_32Fx8; |
| 3877 | case Iop_Sub32Fx8: op = Asse_SUBF; goto do_32Fx8; |
| 3878 | case Iop_Mul32Fx8: op = Asse_MULF; goto do_32Fx8; |
| 3879 | case Iop_Div32Fx8: op = Asse_DIVF; goto do_32Fx8; |
sewardj | 8eb7ae8 | 2012-06-24 14:00:27 +0000 | [diff] [blame] | 3880 | case Iop_Max32Fx8: op = Asse_MAXF; goto do_32Fx8; |
| 3881 | case Iop_Min32Fx8: op = Asse_MINF; goto do_32Fx8; |
sewardj | 56c3031 | 2012-06-12 08:45:39 +0000 | [diff] [blame] | 3882 | do_32Fx8: |
| 3883 | { |
| 3884 | HReg argLhi, argLlo, argRhi, argRlo; |
| 3885 | iselDVecExpr(&argLhi, &argLlo, env, e->Iex.Binop.arg1); |
| 3886 | iselDVecExpr(&argRhi, &argRlo, env, e->Iex.Binop.arg2); |
| 3887 | HReg dstHi = newVRegV(env); |
| 3888 | HReg dstLo = newVRegV(env); |
| 3889 | addInstr(env, mk_vMOVsd_RR(argLhi, dstHi)); |
| 3890 | addInstr(env, mk_vMOVsd_RR(argLlo, dstLo)); |
| 3891 | addInstr(env, AMD64Instr_Sse32Fx4(op, argRhi, dstHi)); |
| 3892 | addInstr(env, AMD64Instr_Sse32Fx4(op, argRlo, dstLo)); |
| 3893 | *rHi = dstHi; |
| 3894 | *rLo = dstLo; |
| 3895 | return; |
| 3896 | } |
| 3897 | |
sewardj | 4b1cc83 | 2012-06-13 11:10:20 +0000 | [diff] [blame] | 3898 | case Iop_AndV256: op = Asse_AND; goto do_SseReRg; |
sewardj | 2a2bda9 | 2012-06-14 23:32:02 +0000 | [diff] [blame] | 3899 | case Iop_OrV256: op = Asse_OR; goto do_SseReRg; |
sewardj | 4b1cc83 | 2012-06-13 11:10:20 +0000 | [diff] [blame] | 3900 | case Iop_XorV256: op = Asse_XOR; goto do_SseReRg; |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 3901 | case Iop_Add8x32: op = Asse_ADD8; goto do_SseReRg; |
| 3902 | case Iop_Add16x16: op = Asse_ADD16; goto do_SseReRg; |
| 3903 | case Iop_Add32x8: op = Asse_ADD32; goto do_SseReRg; |
| 3904 | case Iop_Add64x4: op = Asse_ADD64; goto do_SseReRg; |
| 3905 | case Iop_QAdd8Sx32: op = Asse_QADD8S; goto do_SseReRg; |
| 3906 | case Iop_QAdd16Sx16: op = Asse_QADD16S; goto do_SseReRg; |
| 3907 | case Iop_QAdd8Ux32: op = Asse_QADD8U; goto do_SseReRg; |
| 3908 | case Iop_QAdd16Ux16: op = Asse_QADD16U; goto do_SseReRg; |
| 3909 | case Iop_Avg8Ux32: op = Asse_AVG8U; goto do_SseReRg; |
| 3910 | case Iop_Avg16Ux16: op = Asse_AVG16U; goto do_SseReRg; |
| 3911 | case Iop_CmpEQ8x32: op = Asse_CMPEQ8; goto do_SseReRg; |
| 3912 | case Iop_CmpEQ16x16: op = Asse_CMPEQ16; goto do_SseReRg; |
| 3913 | case Iop_CmpEQ32x8: op = Asse_CMPEQ32; goto do_SseReRg; |
| 3914 | case Iop_CmpGT8Sx32: op = Asse_CMPGT8S; goto do_SseReRg; |
| 3915 | case Iop_CmpGT16Sx16: op = Asse_CMPGT16S; goto do_SseReRg; |
| 3916 | case Iop_CmpGT32Sx8: op = Asse_CMPGT32S; goto do_SseReRg; |
| 3917 | case Iop_Max16Sx16: op = Asse_MAX16S; goto do_SseReRg; |
| 3918 | case Iop_Max8Ux32: op = Asse_MAX8U; goto do_SseReRg; |
| 3919 | case Iop_Min16Sx16: op = Asse_MIN16S; goto do_SseReRg; |
| 3920 | case Iop_Min8Ux32: op = Asse_MIN8U; goto do_SseReRg; |
| 3921 | case Iop_MulHi16Ux16: op = Asse_MULHI16U; goto do_SseReRg; |
| 3922 | case Iop_MulHi16Sx16: op = Asse_MULHI16S; goto do_SseReRg; |
| 3923 | case Iop_Mul16x16: op = Asse_MUL16; goto do_SseReRg; |
| 3924 | case Iop_Sub8x32: op = Asse_SUB8; goto do_SseReRg; |
| 3925 | case Iop_Sub16x16: op = Asse_SUB16; goto do_SseReRg; |
| 3926 | case Iop_Sub32x8: op = Asse_SUB32; goto do_SseReRg; |
| 3927 | case Iop_Sub64x4: op = Asse_SUB64; goto do_SseReRg; |
| 3928 | case Iop_QSub8Sx32: op = Asse_QSUB8S; goto do_SseReRg; |
| 3929 | case Iop_QSub16Sx16: op = Asse_QSUB16S; goto do_SseReRg; |
| 3930 | case Iop_QSub8Ux32: op = Asse_QSUB8U; goto do_SseReRg; |
| 3931 | case Iop_QSub16Ux16: op = Asse_QSUB16U; goto do_SseReRg; |
sewardj | 4b1cc83 | 2012-06-13 11:10:20 +0000 | [diff] [blame] | 3932 | do_SseReRg: |
| 3933 | { |
| 3934 | HReg argLhi, argLlo, argRhi, argRlo; |
| 3935 | iselDVecExpr(&argLhi, &argLlo, env, e->Iex.Binop.arg1); |
| 3936 | iselDVecExpr(&argRhi, &argRlo, env, e->Iex.Binop.arg2); |
| 3937 | HReg dstHi = newVRegV(env); |
| 3938 | HReg dstLo = newVRegV(env); |
| 3939 | addInstr(env, mk_vMOVsd_RR(argLhi, dstHi)); |
| 3940 | addInstr(env, mk_vMOVsd_RR(argLlo, dstLo)); |
| 3941 | addInstr(env, AMD64Instr_SseReRg(op, argRhi, dstHi)); |
| 3942 | addInstr(env, AMD64Instr_SseReRg(op, argRlo, dstLo)); |
| 3943 | *rHi = dstHi; |
| 3944 | *rLo = dstLo; |
| 3945 | return; |
| 3946 | } |
| 3947 | |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 3948 | case Iop_ShlN16x16: op = Asse_SHL16; goto do_SseShift; |
| 3949 | case Iop_ShlN32x8: op = Asse_SHL32; goto do_SseShift; |
| 3950 | case Iop_ShlN64x4: op = Asse_SHL64; goto do_SseShift; |
| 3951 | case Iop_SarN16x16: op = Asse_SAR16; goto do_SseShift; |
| 3952 | case Iop_SarN32x8: op = Asse_SAR32; goto do_SseShift; |
| 3953 | case Iop_ShrN16x16: op = Asse_SHR16; goto do_SseShift; |
| 3954 | case Iop_ShrN32x8: op = Asse_SHR32; goto do_SseShift; |
| 3955 | case Iop_ShrN64x4: op = Asse_SHR64; goto do_SseShift; |
| 3956 | do_SseShift: { |
| 3957 | HReg gregHi, gregLo; |
| 3958 | iselDVecExpr(&gregHi, &gregLo, env, e->Iex.Binop.arg1); |
| 3959 | AMD64RMI* rmi = iselIntExpr_RMI(env, e->Iex.Binop.arg2); |
| 3960 | AMD64AMode* rsp0 = AMD64AMode_IR(0, hregAMD64_RSP()); |
| 3961 | HReg ereg = newVRegV(env); |
| 3962 | HReg dstHi = newVRegV(env); |
| 3963 | HReg dstLo = newVRegV(env); |
| 3964 | addInstr(env, AMD64Instr_Push(AMD64RMI_Imm(0))); |
| 3965 | addInstr(env, AMD64Instr_Push(rmi)); |
| 3966 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, ereg, rsp0)); |
| 3967 | addInstr(env, mk_vMOVsd_RR(gregHi, dstHi)); |
| 3968 | addInstr(env, AMD64Instr_SseReRg(op, ereg, dstHi)); |
| 3969 | addInstr(env, mk_vMOVsd_RR(gregLo, dstLo)); |
| 3970 | addInstr(env, AMD64Instr_SseReRg(op, ereg, dstLo)); |
| 3971 | add_to_rsp(env, 16); |
| 3972 | *rHi = dstHi; |
| 3973 | *rLo = dstLo; |
| 3974 | return; |
| 3975 | } |
| 3976 | |
sewardj | 4b1cc83 | 2012-06-13 11:10:20 +0000 | [diff] [blame] | 3977 | case Iop_V128HLtoV256: { |
| 3978 | *rHi = iselVecExpr(env, e->Iex.Binop.arg1); |
| 3979 | *rLo = iselVecExpr(env, e->Iex.Binop.arg2); |
| 3980 | return; |
| 3981 | } |
| 3982 | |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 3983 | case Iop_Mul32x8: fn = (HWord)h_generic_calc_Mul32x4; |
| 3984 | goto do_SseAssistedBinary; |
| 3985 | case Iop_Max32Sx8: fn = (HWord)h_generic_calc_Max32Sx4; |
| 3986 | goto do_SseAssistedBinary; |
| 3987 | case Iop_Min32Sx8: fn = (HWord)h_generic_calc_Min32Sx4; |
| 3988 | goto do_SseAssistedBinary; |
| 3989 | case Iop_Max32Ux8: fn = (HWord)h_generic_calc_Max32Ux4; |
| 3990 | goto do_SseAssistedBinary; |
| 3991 | case Iop_Min32Ux8: fn = (HWord)h_generic_calc_Min32Ux4; |
| 3992 | goto do_SseAssistedBinary; |
| 3993 | case Iop_Max16Ux16: fn = (HWord)h_generic_calc_Max16Ux8; |
| 3994 | goto do_SseAssistedBinary; |
| 3995 | case Iop_Min16Ux16: fn = (HWord)h_generic_calc_Min16Ux8; |
| 3996 | goto do_SseAssistedBinary; |
| 3997 | case Iop_Max8Sx32: fn = (HWord)h_generic_calc_Max8Sx16; |
| 3998 | goto do_SseAssistedBinary; |
| 3999 | case Iop_Min8Sx32: fn = (HWord)h_generic_calc_Min8Sx16; |
| 4000 | goto do_SseAssistedBinary; |
| 4001 | case Iop_CmpEQ64x4: fn = (HWord)h_generic_calc_CmpEQ64x2; |
| 4002 | goto do_SseAssistedBinary; |
| 4003 | case Iop_CmpGT64Sx4: fn = (HWord)h_generic_calc_CmpGT64Sx2; |
| 4004 | goto do_SseAssistedBinary; |
| 4005 | do_SseAssistedBinary: { |
| 4006 | /* RRRufff! RRRufff code is what we're generating here. Oh |
| 4007 | well. */ |
| 4008 | vassert(fn != 0); |
| 4009 | HReg dstHi = newVRegV(env); |
| 4010 | HReg dstLo = newVRegV(env); |
| 4011 | HReg argLhi, argLlo, argRhi, argRlo; |
| 4012 | iselDVecExpr(&argLhi, &argLlo, env, e->Iex.Binop.arg1); |
| 4013 | iselDVecExpr(&argRhi, &argRlo, env, e->Iex.Binop.arg2); |
| 4014 | HReg argp = newVRegI(env); |
| 4015 | /* subq $160, %rsp -- make a space*/ |
| 4016 | sub_from_rsp(env, 160); |
| 4017 | /* leaq 48(%rsp), %r_argp -- point into it */ |
| 4018 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(48, hregAMD64_RSP()), |
| 4019 | argp)); |
| 4020 | /* andq $-16, %r_argp -- 16-align the pointer */ |
| 4021 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, |
| 4022 | AMD64RMI_Imm( ~(UInt)15 ), |
| 4023 | argp)); |
| 4024 | /* Prepare 3 arg regs: |
| 4025 | leaq 0(%r_argp), %rdi |
| 4026 | leaq 16(%r_argp), %rsi |
| 4027 | leaq 32(%r_argp), %rdx |
| 4028 | */ |
| 4029 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(0, argp), |
| 4030 | hregAMD64_RDI())); |
| 4031 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(16, argp), |
| 4032 | hregAMD64_RSI())); |
| 4033 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(32, argp), |
| 4034 | hregAMD64_RDX())); |
| 4035 | /* Store the two high args, at (%rsi) and (%rdx): |
| 4036 | movupd %argLhi, 0(%rsi) |
| 4037 | movupd %argRhi, 0(%rdx) |
| 4038 | */ |
| 4039 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argLhi, |
| 4040 | AMD64AMode_IR(0, hregAMD64_RSI()))); |
| 4041 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argRhi, |
| 4042 | AMD64AMode_IR(0, hregAMD64_RDX()))); |
| 4043 | /* Store the two low args, at 48(%rsi) and 48(%rdx): |
| 4044 | movupd %argLlo, 48(%rsi) |
| 4045 | movupd %argRlo, 48(%rdx) |
| 4046 | */ |
| 4047 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argLlo, |
| 4048 | AMD64AMode_IR(48, hregAMD64_RSI()))); |
| 4049 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argRlo, |
| 4050 | AMD64AMode_IR(48, hregAMD64_RDX()))); |
| 4051 | /* call the helper */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4052 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, 3, |
| 4053 | mk_RetLoc_simple(RLPri_None) )); |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 4054 | /* Prepare 3 arg regs: |
| 4055 | leaq 48(%r_argp), %rdi |
| 4056 | leaq 64(%r_argp), %rsi |
| 4057 | leaq 80(%r_argp), %rdx |
| 4058 | */ |
| 4059 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(48, argp), |
| 4060 | hregAMD64_RDI())); |
| 4061 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(64, argp), |
| 4062 | hregAMD64_RSI())); |
| 4063 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(80, argp), |
| 4064 | hregAMD64_RDX())); |
| 4065 | /* call the helper */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4066 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, 3, |
| 4067 | mk_RetLoc_simple(RLPri_None) )); |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 4068 | /* fetch the result from memory, using %r_argp, which the |
| 4069 | register allocator will keep alive across the call. */ |
| 4070 | addInstr(env, AMD64Instr_SseLdSt(True/*isLoad*/, 16, dstHi, |
| 4071 | AMD64AMode_IR(0, argp))); |
| 4072 | addInstr(env, AMD64Instr_SseLdSt(True/*isLoad*/, 16, dstLo, |
| 4073 | AMD64AMode_IR(48, argp))); |
| 4074 | /* and finally, clear the space */ |
| 4075 | add_to_rsp(env, 160); |
| 4076 | *rHi = dstHi; |
| 4077 | *rLo = dstLo; |
| 4078 | return; |
| 4079 | } |
| 4080 | |
| 4081 | case Iop_Perm32x8: fn = (HWord)h_generic_calc_Perm32x8; |
| 4082 | goto do_SseAssistedBinary256; |
| 4083 | do_SseAssistedBinary256: { |
| 4084 | /* RRRufff! RRRufff code is what we're generating here. Oh |
| 4085 | well. */ |
| 4086 | vassert(fn != 0); |
| 4087 | HReg dstHi = newVRegV(env); |
| 4088 | HReg dstLo = newVRegV(env); |
| 4089 | HReg argLhi, argLlo, argRhi, argRlo; |
| 4090 | iselDVecExpr(&argLhi, &argLlo, env, e->Iex.Binop.arg1); |
| 4091 | iselDVecExpr(&argRhi, &argRlo, env, e->Iex.Binop.arg2); |
| 4092 | HReg argp = newVRegI(env); |
| 4093 | /* subq $160, %rsp -- make a space*/ |
| 4094 | sub_from_rsp(env, 160); |
| 4095 | /* leaq 48(%rsp), %r_argp -- point into it */ |
| 4096 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(48, hregAMD64_RSP()), |
| 4097 | argp)); |
| 4098 | /* andq $-16, %r_argp -- 16-align the pointer */ |
| 4099 | addInstr(env, AMD64Instr_Alu64R(Aalu_AND, |
| 4100 | AMD64RMI_Imm( ~(UInt)15 ), |
| 4101 | argp)); |
| 4102 | /* Prepare 3 arg regs: |
| 4103 | leaq 0(%r_argp), %rdi |
| 4104 | leaq 32(%r_argp), %rsi |
| 4105 | leaq 64(%r_argp), %rdx |
| 4106 | */ |
| 4107 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(0, argp), |
| 4108 | hregAMD64_RDI())); |
| 4109 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(32, argp), |
| 4110 | hregAMD64_RSI())); |
| 4111 | addInstr(env, AMD64Instr_Lea64(AMD64AMode_IR(64, argp), |
| 4112 | hregAMD64_RDX())); |
| 4113 | /* Store the two args, at (%rsi) and (%rdx): |
| 4114 | movupd %argLlo, 0(%rsi) |
| 4115 | movupd %argLhi, 16(%rsi) |
| 4116 | movupd %argRlo, 0(%rdx) |
| 4117 | movupd %argRhi, 16(%rdx) |
| 4118 | */ |
| 4119 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argLlo, |
| 4120 | AMD64AMode_IR(0, hregAMD64_RSI()))); |
| 4121 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argLhi, |
| 4122 | AMD64AMode_IR(16, hregAMD64_RSI()))); |
| 4123 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argRlo, |
| 4124 | AMD64AMode_IR(0, hregAMD64_RDX()))); |
| 4125 | addInstr(env, AMD64Instr_SseLdSt(False/*!isLoad*/, 16, argRhi, |
| 4126 | AMD64AMode_IR(16, hregAMD64_RDX()))); |
| 4127 | /* call the helper */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4128 | addInstr(env, AMD64Instr_Call( Acc_ALWAYS, (ULong)fn, 3, |
| 4129 | mk_RetLoc_simple(RLPri_None) )); |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 4130 | /* fetch the result from memory, using %r_argp, which the |
| 4131 | register allocator will keep alive across the call. */ |
| 4132 | addInstr(env, AMD64Instr_SseLdSt(True/*isLoad*/, 16, dstLo, |
| 4133 | AMD64AMode_IR(0, argp))); |
| 4134 | addInstr(env, AMD64Instr_SseLdSt(True/*isLoad*/, 16, dstHi, |
| 4135 | AMD64AMode_IR(16, argp))); |
| 4136 | /* and finally, clear the space */ |
| 4137 | add_to_rsp(env, 160); |
| 4138 | *rHi = dstHi; |
| 4139 | *rLo = dstLo; |
| 4140 | return; |
| 4141 | } |
| 4142 | |
sewardj | 56c3031 | 2012-06-12 08:45:39 +0000 | [diff] [blame] | 4143 | default: |
| 4144 | break; |
| 4145 | } /* switch (e->Iex.Binop.op) */ |
| 4146 | } /* if (e->tag == Iex_Binop) */ |
| 4147 | |
florian | 96d7cc3 | 2012-06-01 20:41:24 +0000 | [diff] [blame] | 4148 | if (e->tag == Iex_Qop && e->Iex.Qop.details->op == Iop_64x4toV256) { |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4149 | HReg rsp = hregAMD64_RSP(); |
| 4150 | HReg vHi = newVRegV(env); |
| 4151 | HReg vLo = newVRegV(env); |
| 4152 | AMD64AMode* m8_rsp = AMD64AMode_IR(-8, rsp); |
| 4153 | AMD64AMode* m16_rsp = AMD64AMode_IR(-16, rsp); |
| 4154 | /* arg1 is the most significant (Q3), arg4 the least (Q0) */ |
| 4155 | /* Get all the args into regs, before messing with the stack. */ |
florian | 96d7cc3 | 2012-06-01 20:41:24 +0000 | [diff] [blame] | 4156 | AMD64RI* q3 = iselIntExpr_RI(env, e->Iex.Qop.details->arg1); |
| 4157 | AMD64RI* q2 = iselIntExpr_RI(env, e->Iex.Qop.details->arg2); |
| 4158 | AMD64RI* q1 = iselIntExpr_RI(env, e->Iex.Qop.details->arg3); |
| 4159 | AMD64RI* q0 = iselIntExpr_RI(env, e->Iex.Qop.details->arg4); |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4160 | /* less significant lane (Q2) at the lower address (-16(rsp)) */ |
| 4161 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, q3, m8_rsp)); |
| 4162 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, q2, m16_rsp)); |
| 4163 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, vHi, m16_rsp)); |
| 4164 | /* and then the lower half .. */ |
| 4165 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, q1, m8_rsp)); |
| 4166 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV, q0, m16_rsp)); |
| 4167 | addInstr(env, AMD64Instr_SseLdSt(True/*load*/, 16, vLo, m16_rsp)); |
| 4168 | *rHi = vHi; |
| 4169 | *rLo = vLo; |
| 4170 | return; |
| 4171 | } |
| 4172 | |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 4173 | if (e->tag == Iex_ITE) { |
| 4174 | HReg r1Hi, r1Lo, r0Hi, r0Lo; |
| 4175 | iselDVecExpr(&r1Hi, &r1Lo, env, e->Iex.ITE.iftrue); |
| 4176 | iselDVecExpr(&r0Hi, &r0Lo, env, e->Iex.ITE.iffalse); |
| 4177 | HReg dstHi = newVRegV(env); |
| 4178 | HReg dstLo = newVRegV(env); |
| 4179 | addInstr(env, mk_vMOVsd_RR(r1Hi,dstHi)); |
| 4180 | addInstr(env, mk_vMOVsd_RR(r1Lo,dstLo)); |
| 4181 | AMD64CondCode cc = iselCondCode(env, e->Iex.ITE.cond); |
| 4182 | addInstr(env, AMD64Instr_SseCMov(cc ^ 1, r0Hi, dstHi)); |
| 4183 | addInstr(env, AMD64Instr_SseCMov(cc ^ 1, r0Lo, dstLo)); |
| 4184 | *rHi = dstHi; |
| 4185 | *rLo = dstLo; |
| 4186 | return; |
| 4187 | } |
| 4188 | |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4189 | //avx_fail: |
| 4190 | vex_printf("iselDVecExpr (amd64, subarch = %s): can't reduce\n", |
| 4191 | LibVEX_ppVexHwCaps(VexArchAMD64, env->hwcaps)); |
| 4192 | ppIRExpr(e); |
| 4193 | vpanic("iselDVecExpr_wrk"); |
| 4194 | } |
| 4195 | |
| 4196 | |
| 4197 | /*---------------------------------------------------------*/ |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4198 | /*--- ISEL: Statements ---*/ |
| 4199 | /*---------------------------------------------------------*/ |
| 4200 | |
| 4201 | static void iselStmt ( ISelEnv* env, IRStmt* stmt ) |
| 4202 | { |
| 4203 | if (vex_traceflags & VEX_TRACE_VCODE) { |
| 4204 | vex_printf("\n-- "); |
| 4205 | ppIRStmt(stmt); |
| 4206 | vex_printf("\n"); |
| 4207 | } |
| 4208 | |
| 4209 | switch (stmt->tag) { |
| 4210 | |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 4211 | /* --------- STORE --------- */ |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4212 | case Ist_Store: { |
sewardj | e9d8a26 | 2009-07-01 08:06:34 +0000 | [diff] [blame] | 4213 | IRType tya = typeOfIRExpr(env->type_env, stmt->Ist.Store.addr); |
| 4214 | IRType tyd = typeOfIRExpr(env->type_env, stmt->Ist.Store.data); |
| 4215 | IREndness end = stmt->Ist.Store.end; |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4216 | |
sewardj | e768e92 | 2009-11-26 17:17:37 +0000 | [diff] [blame] | 4217 | if (tya != Ity_I64 || end != Iend_LE) |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4218 | goto stmt_fail; |
| 4219 | |
sewardj | 3119107 | 2005-02-05 18:24:47 +0000 | [diff] [blame] | 4220 | if (tyd == Ity_I64) { |
sewardj | bf0d86c | 2007-11-26 23:18:52 +0000 | [diff] [blame] | 4221 | AMD64AMode* am = iselIntExpr_AMode(env, stmt->Ist.Store.addr); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4222 | AMD64RI* ri = iselIntExpr_RI(env, stmt->Ist.Store.data); |
sewardj | 3119107 | 2005-02-05 18:24:47 +0000 | [diff] [blame] | 4223 | addInstr(env, AMD64Instr_Alu64M(Aalu_MOV,ri,am)); |
| 4224 | return; |
| 4225 | } |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 4226 | if (tyd == Ity_I8 || tyd == Ity_I16 || tyd == Ity_I32) { |
sewardj | bf0d86c | 2007-11-26 23:18:52 +0000 | [diff] [blame] | 4227 | AMD64AMode* am = iselIntExpr_AMode(env, stmt->Ist.Store.addr); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4228 | HReg r = iselIntExpr_R(env, stmt->Ist.Store.data); |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 4229 | addInstr(env, AMD64Instr_Store( |
| 4230 | toUChar(tyd==Ity_I8 ? 1 : (tyd==Ity_I16 ? 2 : 4)), |
| 4231 | r,am)); |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 4232 | return; |
| 4233 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4234 | if (tyd == Ity_F64) { |
sewardj | bf0d86c | 2007-11-26 23:18:52 +0000 | [diff] [blame] | 4235 | AMD64AMode* am = iselIntExpr_AMode(env, stmt->Ist.Store.addr); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4236 | HReg r = iselDblExpr(env, stmt->Ist.Store.data); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4237 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 8, r, am)); |
| 4238 | return; |
| 4239 | } |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 4240 | if (tyd == Ity_F32) { |
sewardj | bf0d86c | 2007-11-26 23:18:52 +0000 | [diff] [blame] | 4241 | AMD64AMode* am = iselIntExpr_AMode(env, stmt->Ist.Store.addr); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4242 | HReg r = iselFltExpr(env, stmt->Ist.Store.data); |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 4243 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 4, r, am)); |
| 4244 | return; |
| 4245 | } |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 4246 | if (tyd == Ity_V128) { |
sewardj | bf0d86c | 2007-11-26 23:18:52 +0000 | [diff] [blame] | 4247 | AMD64AMode* am = iselIntExpr_AMode(env, stmt->Ist.Store.addr); |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4248 | HReg r = iselVecExpr(env, stmt->Ist.Store.data); |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 4249 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, r, am)); |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 4250 | return; |
| 4251 | } |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4252 | if (tyd == Ity_V256) { |
| 4253 | HReg rA = iselIntExpr_R(env, stmt->Ist.Store.addr); |
| 4254 | AMD64AMode* am0 = AMD64AMode_IR(0, rA); |
| 4255 | AMD64AMode* am16 = AMD64AMode_IR(16, rA); |
| 4256 | HReg vHi, vLo; |
| 4257 | iselDVecExpr(&vHi, &vLo, env, stmt->Ist.Store.data); |
| 4258 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, vLo, am0)); |
| 4259 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, vHi, am16)); |
| 4260 | return; |
| 4261 | } |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 4262 | break; |
sewardj | 05b3b6a | 2005-02-04 01:44:33 +0000 | [diff] [blame] | 4263 | } |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 4264 | |
| 4265 | /* --------- PUT --------- */ |
| 4266 | case Ist_Put: { |
| 4267 | IRType ty = typeOfIRExpr(env->type_env, stmt->Ist.Put.data); |
| 4268 | if (ty == Ity_I64) { |
| 4269 | /* We're going to write to memory, so compute the RHS into an |
| 4270 | AMD64RI. */ |
| 4271 | AMD64RI* ri = iselIntExpr_RI(env, stmt->Ist.Put.data); |
| 4272 | addInstr(env, |
| 4273 | AMD64Instr_Alu64M( |
| 4274 | Aalu_MOV, |
| 4275 | ri, |
| 4276 | AMD64AMode_IR(stmt->Ist.Put.offset, |
| 4277 | hregAMD64_RBP()) |
| 4278 | )); |
| 4279 | return; |
| 4280 | } |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 4281 | if (ty == Ity_I8 || ty == Ity_I16 || ty == Ity_I32) { |
| 4282 | HReg r = iselIntExpr_R(env, stmt->Ist.Put.data); |
| 4283 | addInstr(env, AMD64Instr_Store( |
sewardj | 428fabd | 2005-03-21 03:11:17 +0000 | [diff] [blame] | 4284 | toUChar(ty==Ity_I8 ? 1 : (ty==Ity_I16 ? 2 : 4)), |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 4285 | r, |
| 4286 | AMD64AMode_IR(stmt->Ist.Put.offset, |
| 4287 | hregAMD64_RBP()))); |
| 4288 | return; |
| 4289 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4290 | if (ty == Ity_F32) { |
| 4291 | HReg f32 = iselFltExpr(env, stmt->Ist.Put.data); |
| 4292 | AMD64AMode* am = AMD64AMode_IR(stmt->Ist.Put.offset, hregAMD64_RBP()); |
| 4293 | set_SSE_rounding_default(env); /* paranoia */ |
| 4294 | addInstr(env, AMD64Instr_SseLdSt( False/*store*/, 4, f32, am )); |
| 4295 | return; |
| 4296 | } |
sewardj | 1a01e65 | 2005-02-23 11:39:21 +0000 | [diff] [blame] | 4297 | if (ty == Ity_F64) { |
| 4298 | HReg f64 = iselDblExpr(env, stmt->Ist.Put.data); |
| 4299 | AMD64AMode* am = AMD64AMode_IR( stmt->Ist.Put.offset, |
| 4300 | hregAMD64_RBP() ); |
| 4301 | addInstr(env, AMD64Instr_SseLdSt( False/*store*/, 8, f64, am )); |
| 4302 | return; |
| 4303 | } |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4304 | if (ty == Ity_V128) { |
| 4305 | HReg vec = iselVecExpr(env, stmt->Ist.Put.data); |
| 4306 | AMD64AMode* am = AMD64AMode_IR(stmt->Ist.Put.offset, |
| 4307 | hregAMD64_RBP()); |
| 4308 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, vec, am)); |
| 4309 | return; |
| 4310 | } |
| 4311 | if (ty == Ity_V256) { |
| 4312 | HReg vHi, vLo; |
| 4313 | iselDVecExpr(&vHi, &vLo, env, stmt->Ist.Put.data); |
| 4314 | HReg rbp = hregAMD64_RBP(); |
| 4315 | AMD64AMode* am0 = AMD64AMode_IR(stmt->Ist.Put.offset + 0, rbp); |
| 4316 | AMD64AMode* am16 = AMD64AMode_IR(stmt->Ist.Put.offset + 16, rbp); |
| 4317 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, vLo, am0)); |
| 4318 | addInstr(env, AMD64Instr_SseLdSt(False/*store*/, 16, vHi, am16)); |
| 4319 | return; |
| 4320 | } |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 4321 | break; |
| 4322 | } |
| 4323 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4324 | /* --------- Indexed PUT --------- */ |
| 4325 | case Ist_PutI: { |
florian | d6f38b3 | 2012-05-31 15:46:18 +0000 | [diff] [blame] | 4326 | IRPutI *puti = stmt->Ist.PutI.details; |
| 4327 | |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4328 | AMD64AMode* am |
| 4329 | = genGuestArrayOffset( |
florian | d6f38b3 | 2012-05-31 15:46:18 +0000 | [diff] [blame] | 4330 | env, puti->descr, |
| 4331 | puti->ix, puti->bias ); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4332 | |
florian | d6f38b3 | 2012-05-31 15:46:18 +0000 | [diff] [blame] | 4333 | IRType ty = typeOfIRExpr(env->type_env, puti->data); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4334 | if (ty == Ity_F64) { |
florian | d6f38b3 | 2012-05-31 15:46:18 +0000 | [diff] [blame] | 4335 | HReg val = iselDblExpr(env, puti->data); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4336 | addInstr(env, AMD64Instr_SseLdSt( False/*store*/, 8, val, am )); |
| 4337 | return; |
| 4338 | } |
| 4339 | if (ty == Ity_I8) { |
florian | d6f38b3 | 2012-05-31 15:46:18 +0000 | [diff] [blame] | 4340 | HReg r = iselIntExpr_R(env, puti->data); |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4341 | addInstr(env, AMD64Instr_Store( 1, r, am )); |
| 4342 | return; |
| 4343 | } |
sewardj | 1e015d8 | 2005-04-23 23:41:46 +0000 | [diff] [blame] | 4344 | if (ty == Ity_I64) { |
florian | d6f38b3 | 2012-05-31 15:46:18 +0000 | [diff] [blame] | 4345 | AMD64RI* ri = iselIntExpr_RI(env, puti->data); |
sewardj | 1e015d8 | 2005-04-23 23:41:46 +0000 | [diff] [blame] | 4346 | addInstr(env, AMD64Instr_Alu64M( Aalu_MOV, ri, am )); |
| 4347 | return; |
| 4348 | } |
sewardj | 8d96531 | 2005-02-25 02:48:47 +0000 | [diff] [blame] | 4349 | break; |
| 4350 | } |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 4351 | |
| 4352 | /* --------- TMP --------- */ |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 4353 | case Ist_WrTmp: { |
| 4354 | IRTemp tmp = stmt->Ist.WrTmp.tmp; |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 4355 | IRType ty = typeOfIRTemp(env->type_env, tmp); |
sewardj | 6ce1a23 | 2007-03-31 19:12:38 +0000 | [diff] [blame] | 4356 | |
| 4357 | /* optimisation: if stmt->Ist.WrTmp.data is Add64(..,..), |
| 4358 | compute it into an AMode and then use LEA. This usually |
| 4359 | produces fewer instructions, often because (for memcheck |
| 4360 | created IR) we get t = address-expression, (t is later used |
| 4361 | twice) and so doing this naturally turns address-expression |
| 4362 | back into an AMD64 amode. */ |
| 4363 | if (ty == Ity_I64 |
| 4364 | && stmt->Ist.WrTmp.data->tag == Iex_Binop |
| 4365 | && stmt->Ist.WrTmp.data->Iex.Binop.op == Iop_Add64) { |
| 4366 | AMD64AMode* am = iselIntExpr_AMode(env, stmt->Ist.WrTmp.data); |
| 4367 | HReg dst = lookupIRTemp(env, tmp); |
| 4368 | if (am->tag == Aam_IR && am->Aam.IR.imm == 0) { |
| 4369 | /* Hmm, iselIntExpr_AMode wimped out and just computed the |
| 4370 | value into a register. Just emit a normal reg-reg move |
| 4371 | so reg-alloc can coalesce it away in the usual way. */ |
| 4372 | HReg src = am->Aam.IR.reg; |
| 4373 | addInstr(env, AMD64Instr_Alu64R(Aalu_MOV, AMD64RMI_Reg(src), dst)); |
| 4374 | } else { |
| 4375 | addInstr(env, AMD64Instr_Lea64(am,dst)); |
| 4376 | } |
| 4377 | return; |
| 4378 | } |
| 4379 | |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 4380 | if (ty == Ity_I64 || ty == Ity_I32 |
| 4381 | || ty == Ity_I16 || ty == Ity_I8) { |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 4382 | AMD64RMI* rmi = iselIntExpr_RMI(env, stmt->Ist.WrTmp.data); |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 4383 | HReg dst = lookupIRTemp(env, tmp); |
| 4384 | addInstr(env, AMD64Instr_Alu64R(Aalu_MOV,rmi,dst)); |
| 4385 | return; |
| 4386 | } |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 4387 | if (ty == Ity_I128) { |
| 4388 | HReg rHi, rLo, dstHi, dstLo; |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 4389 | iselInt128Expr(&rHi,&rLo, env, stmt->Ist.WrTmp.data); |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4390 | lookupIRTempPair( &dstHi, &dstLo, env, tmp); |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 4391 | addInstr(env, mk_iMOVsd_RR(rHi,dstHi) ); |
| 4392 | addInstr(env, mk_iMOVsd_RR(rLo,dstLo) ); |
| 4393 | return; |
| 4394 | } |
sewardj | a5bd0af | 2005-03-24 20:40:12 +0000 | [diff] [blame] | 4395 | if (ty == Ity_I1) { |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 4396 | AMD64CondCode cond = iselCondCode(env, stmt->Ist.WrTmp.data); |
sewardj | a5bd0af | 2005-03-24 20:40:12 +0000 | [diff] [blame] | 4397 | HReg dst = lookupIRTemp(env, tmp); |
| 4398 | addInstr(env, AMD64Instr_Set64(cond, dst)); |
| 4399 | return; |
| 4400 | } |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 4401 | if (ty == Ity_F64) { |
| 4402 | HReg dst = lookupIRTemp(env, tmp); |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 4403 | HReg src = iselDblExpr(env, stmt->Ist.WrTmp.data); |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 4404 | addInstr(env, mk_vMOVsd_RR(src, dst)); |
| 4405 | return; |
| 4406 | } |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 4407 | if (ty == Ity_F32) { |
| 4408 | HReg dst = lookupIRTemp(env, tmp); |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 4409 | HReg src = iselFltExpr(env, stmt->Ist.WrTmp.data); |
sewardj | c49ce23 | 2005-02-25 13:03:03 +0000 | [diff] [blame] | 4410 | addInstr(env, mk_vMOVsd_RR(src, dst)); |
| 4411 | return; |
| 4412 | } |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 4413 | if (ty == Ity_V128) { |
| 4414 | HReg dst = lookupIRTemp(env, tmp); |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 4415 | HReg src = iselVecExpr(env, stmt->Ist.WrTmp.data); |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 4416 | addInstr(env, mk_vMOVsd_RR(src, dst)); |
sewardj | 0852a13 | 2005-02-21 08:28:46 +0000 | [diff] [blame] | 4417 | return; |
| 4418 | } |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4419 | if (ty == Ity_V256) { |
| 4420 | HReg rHi, rLo, dstHi, dstLo; |
| 4421 | iselDVecExpr(&rHi,&rLo, env, stmt->Ist.WrTmp.data); |
| 4422 | lookupIRTempPair( &dstHi, &dstLo, env, tmp); |
| 4423 | addInstr(env, mk_vMOVsd_RR(rHi,dstHi) ); |
| 4424 | addInstr(env, mk_vMOVsd_RR(rLo,dstLo) ); |
| 4425 | return; |
| 4426 | } |
sewardj | 614b3fb | 2005-02-02 02:16:03 +0000 | [diff] [blame] | 4427 | break; |
| 4428 | } |
| 4429 | |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 4430 | /* --------- Call to DIRTY helper --------- */ |
| 4431 | case Ist_Dirty: { |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 4432 | IRDirty* d = stmt->Ist.Dirty.details; |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 4433 | |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 4434 | /* Figure out the return type, if any. */ |
| 4435 | IRType retty = Ity_INVALID; |
| 4436 | if (d->tmp != IRTemp_INVALID) |
| 4437 | retty = typeOfIRTemp(env->type_env, d->tmp); |
| 4438 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4439 | /* Throw out any return types we don't know about. */ |
| 4440 | Bool retty_ok = False; |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 4441 | switch (retty) { |
| 4442 | case Ity_INVALID: /* function doesn't return anything */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4443 | case Ity_I64: case Ity_I32: case Ity_I16: case Ity_I8: |
sewardj | 82cc37c | 2013-08-16 08:32:15 +0000 | [diff] [blame] | 4444 | case Ity_V128: case Ity_V256: |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4445 | retty_ok = True; break; |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 4446 | default: |
| 4447 | break; |
| 4448 | } |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4449 | if (!retty_ok) |
sewardj | cfe046e | 2013-01-17 14:23:53 +0000 | [diff] [blame] | 4450 | break; /* will go to stmt_fail: */ |
| 4451 | |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4452 | /* Marshal args, do the call, and set the return value to |
| 4453 | 0x555..555 if this is a conditional call that returns a value |
| 4454 | and the call is skipped. */ |
| 4455 | UInt addToSp = 0; |
| 4456 | RetLoc rloc = mk_RetLoc_INVALID(); |
| 4457 | doHelperCall( &addToSp, &rloc, env, d->guard, d->cee, retty, d->args ); |
| 4458 | vassert(is_sane_RetLoc(rloc)); |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 4459 | |
| 4460 | /* Now figure out what to do with the returned value, if any. */ |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4461 | switch (retty) { |
| 4462 | case Ity_INVALID: { |
| 4463 | /* No return value. Nothing to do. */ |
| 4464 | vassert(d->tmp == IRTemp_INVALID); |
| 4465 | vassert(rloc.pri == RLPri_None); |
| 4466 | vassert(addToSp == 0); |
| 4467 | return; |
| 4468 | } |
| 4469 | case Ity_I64: case Ity_I32: case Ity_I16: case Ity_I8: { |
| 4470 | /* The returned value is in %rax. Park it in the register |
| 4471 | associated with tmp. */ |
| 4472 | vassert(rloc.pri == RLPri_Int); |
| 4473 | vassert(addToSp == 0); |
| 4474 | HReg dst = lookupIRTemp(env, d->tmp); |
| 4475 | addInstr(env, mk_iMOVsd_RR(hregAMD64_RAX(),dst) ); |
| 4476 | return; |
| 4477 | } |
| 4478 | case Ity_V128: { |
sewardj | 54eea4e | 2013-09-02 13:17:49 +0000 | [diff] [blame] | 4479 | /* The returned value is on the stack, and rloc.spOff |
| 4480 | tells us where. Fish it off the stack and then move |
| 4481 | the stack pointer upwards to clear it, as directed by |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4482 | doHelperCall. */ |
| 4483 | vassert(rloc.pri == RLPri_V128SpRel); |
| 4484 | vassert(addToSp >= 16); |
| 4485 | HReg dst = lookupIRTemp(env, d->tmp); |
| 4486 | AMD64AMode* am = AMD64AMode_IR(rloc.spOff, hregAMD64_RSP()); |
| 4487 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 16, dst, am )); |
| 4488 | add_to_rsp(env, addToSp); |
| 4489 | return; |
| 4490 | } |
sewardj | 82cc37c | 2013-08-16 08:32:15 +0000 | [diff] [blame] | 4491 | case Ity_V256: { |
| 4492 | /* See comments for Ity_V128. */ |
| 4493 | vassert(rloc.pri == RLPri_V256SpRel); |
| 4494 | vassert(addToSp >= 32); |
| 4495 | HReg dstLo, dstHi; |
| 4496 | lookupIRTempPair(&dstHi, &dstLo, env, d->tmp); |
| 4497 | AMD64AMode* amLo = AMD64AMode_IR(rloc.spOff, hregAMD64_RSP()); |
| 4498 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 16, dstLo, amLo )); |
| 4499 | AMD64AMode* amHi = AMD64AMode_IR(rloc.spOff+16, hregAMD64_RSP()); |
| 4500 | addInstr(env, AMD64Instr_SseLdSt( True/*load*/, 16, dstHi, amHi )); |
| 4501 | add_to_rsp(env, addToSp); |
| 4502 | return; |
| 4503 | } |
sewardj | 74142b8 | 2013-08-08 10:28:59 +0000 | [diff] [blame] | 4504 | default: |
| 4505 | /*NOTREACHED*/ |
| 4506 | vassert(0); |
sewardj | d0a12df | 2005-02-10 02:07:43 +0000 | [diff] [blame] | 4507 | } |
| 4508 | break; |
| 4509 | } |
| 4510 | |
| 4511 | /* --------- MEM FENCE --------- */ |
sewardj | c4356f0 | 2007-11-09 21:15:04 +0000 | [diff] [blame] | 4512 | case Ist_MBE: |
| 4513 | switch (stmt->Ist.MBE.event) { |
| 4514 | case Imbe_Fence: |
| 4515 | addInstr(env, AMD64Instr_MFence()); |
| 4516 | return; |
sewardj | c4356f0 | 2007-11-09 21:15:04 +0000 | [diff] [blame] | 4517 | default: |
| 4518 | break; |
| 4519 | } |
| 4520 | break; |
sewardj | f8c37f7 | 2005-02-07 18:55:29 +0000 | [diff] [blame] | 4521 | |
sewardj | e9d8a26 | 2009-07-01 08:06:34 +0000 | [diff] [blame] | 4522 | /* --------- ACAS --------- */ |
| 4523 | case Ist_CAS: |
| 4524 | if (stmt->Ist.CAS.details->oldHi == IRTemp_INVALID) { |
| 4525 | /* "normal" singleton CAS */ |
| 4526 | UChar sz; |
| 4527 | IRCAS* cas = stmt->Ist.CAS.details; |
| 4528 | IRType ty = typeOfIRExpr(env->type_env, cas->dataLo); |
| 4529 | /* get: cas->expd into %rax, and cas->data into %rbx */ |
| 4530 | AMD64AMode* am = iselIntExpr_AMode(env, cas->addr); |
| 4531 | HReg rData = iselIntExpr_R(env, cas->dataLo); |
| 4532 | HReg rExpd = iselIntExpr_R(env, cas->expdLo); |
| 4533 | HReg rOld = lookupIRTemp(env, cas->oldLo); |
| 4534 | vassert(cas->expdHi == NULL); |
| 4535 | vassert(cas->dataHi == NULL); |
| 4536 | addInstr(env, mk_iMOVsd_RR(rExpd, rOld)); |
| 4537 | addInstr(env, mk_iMOVsd_RR(rExpd, hregAMD64_RAX())); |
| 4538 | addInstr(env, mk_iMOVsd_RR(rData, hregAMD64_RBX())); |
| 4539 | switch (ty) { |
| 4540 | case Ity_I64: sz = 8; break; |
| 4541 | case Ity_I32: sz = 4; break; |
| 4542 | case Ity_I16: sz = 2; break; |
| 4543 | case Ity_I8: sz = 1; break; |
| 4544 | default: goto unhandled_cas; |
| 4545 | } |
| 4546 | addInstr(env, AMD64Instr_ACAS(am, sz)); |
| 4547 | addInstr(env, AMD64Instr_CMov64( |
| 4548 | Acc_NZ, AMD64RM_Reg(hregAMD64_RAX()), rOld)); |
| 4549 | return; |
| 4550 | } else { |
| 4551 | /* double CAS */ |
| 4552 | UChar sz; |
| 4553 | IRCAS* cas = stmt->Ist.CAS.details; |
| 4554 | IRType ty = typeOfIRExpr(env->type_env, cas->dataLo); |
| 4555 | /* only 32-bit and 64-bit allowed in this case */ |
| 4556 | /* get: cas->expdLo into %rax, and cas->dataLo into %rbx */ |
| 4557 | /* get: cas->expdHi into %rdx, and cas->dataHi into %rcx */ |
| 4558 | AMD64AMode* am = iselIntExpr_AMode(env, cas->addr); |
| 4559 | HReg rDataHi = iselIntExpr_R(env, cas->dataHi); |
| 4560 | HReg rDataLo = iselIntExpr_R(env, cas->dataLo); |
| 4561 | HReg rExpdHi = iselIntExpr_R(env, cas->expdHi); |
| 4562 | HReg rExpdLo = iselIntExpr_R(env, cas->expdLo); |
| 4563 | HReg rOldHi = lookupIRTemp(env, cas->oldHi); |
| 4564 | HReg rOldLo = lookupIRTemp(env, cas->oldLo); |
| 4565 | switch (ty) { |
| 4566 | case Ity_I64: |
| 4567 | if (!(env->hwcaps & VEX_HWCAPS_AMD64_CX16)) |
| 4568 | goto unhandled_cas; /* we'd have to generate |
| 4569 | cmpxchg16b, but the host |
| 4570 | doesn't support that */ |
| 4571 | sz = 8; |
| 4572 | break; |
| 4573 | case Ity_I32: |
| 4574 | sz = 4; |
| 4575 | break; |
| 4576 | default: |
| 4577 | goto unhandled_cas; |
| 4578 | } |
| 4579 | addInstr(env, mk_iMOVsd_RR(rExpdHi, rOldHi)); |
| 4580 | addInstr(env, mk_iMOVsd_RR(rExpdLo, rOldLo)); |
| 4581 | addInstr(env, mk_iMOVsd_RR(rExpdHi, hregAMD64_RDX())); |
| 4582 | addInstr(env, mk_iMOVsd_RR(rExpdLo, hregAMD64_RAX())); |
| 4583 | addInstr(env, mk_iMOVsd_RR(rDataHi, hregAMD64_RCX())); |
| 4584 | addInstr(env, mk_iMOVsd_RR(rDataLo, hregAMD64_RBX())); |
| 4585 | addInstr(env, AMD64Instr_DACAS(am, sz)); |
| 4586 | addInstr(env, |
| 4587 | AMD64Instr_CMov64( |
| 4588 | Acc_NZ, AMD64RM_Reg(hregAMD64_RDX()), rOldHi)); |
| 4589 | addInstr(env, |
| 4590 | AMD64Instr_CMov64( |
| 4591 | Acc_NZ, AMD64RM_Reg(hregAMD64_RAX()), rOldLo)); |
| 4592 | return; |
| 4593 | } |
| 4594 | unhandled_cas: |
| 4595 | break; |
| 4596 | |
sewardj | d20b290 | 2005-03-22 00:15:00 +0000 | [diff] [blame] | 4597 | /* --------- INSTR MARK --------- */ |
| 4598 | /* Doesn't generate any executable code ... */ |
| 4599 | case Ist_IMark: |
| 4600 | return; |
| 4601 | |
sewardj | 5a9ffab | 2005-05-12 17:55:01 +0000 | [diff] [blame] | 4602 | /* --------- ABI HINT --------- */ |
| 4603 | /* These have no meaning (denotation in the IR) and so we ignore |
| 4604 | them ... if any actually made it this far. */ |
| 4605 | case Ist_AbiHint: |
| 4606 | return; |
| 4607 | |
sewardj | d20b290 | 2005-03-22 00:15:00 +0000 | [diff] [blame] | 4608 | /* --------- NO-OP --------- */ |
| 4609 | case Ist_NoOp: |
| 4610 | return; |
| 4611 | |
sewardj | f8c37f7 | 2005-02-07 18:55:29 +0000 | [diff] [blame] | 4612 | /* --------- EXIT --------- */ |
| 4613 | case Ist_Exit: { |
sewardj | f8c37f7 | 2005-02-07 18:55:29 +0000 | [diff] [blame] | 4614 | if (stmt->Ist.Exit.dst->tag != Ico_U64) |
| 4615 | vpanic("iselStmt(amd64): Ist_Exit: dst is not a 64-bit value"); |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4616 | |
| 4617 | AMD64CondCode cc = iselCondCode(env, stmt->Ist.Exit.guard); |
| 4618 | AMD64AMode* amRIP = AMD64AMode_IR(stmt->Ist.Exit.offsIP, |
| 4619 | hregAMD64_RBP()); |
| 4620 | |
| 4621 | /* Case: boring transfer to known address */ |
| 4622 | if (stmt->Ist.Exit.jk == Ijk_Boring) { |
| 4623 | if (env->chainingAllowed) { |
| 4624 | /* .. almost always true .. */ |
| 4625 | /* Skip the event check at the dst if this is a forwards |
| 4626 | edge. */ |
| 4627 | Bool toFastEP |
| 4628 | = ((Addr64)stmt->Ist.Exit.dst->Ico.U64) > env->max_ga; |
| 4629 | if (0) vex_printf("%s", toFastEP ? "Y" : ","); |
| 4630 | addInstr(env, AMD64Instr_XDirect(stmt->Ist.Exit.dst->Ico.U64, |
| 4631 | amRIP, cc, toFastEP)); |
| 4632 | } else { |
| 4633 | /* .. very occasionally .. */ |
| 4634 | /* We can't use chaining, so ask for an assisted transfer, |
| 4635 | as that's the only alternative that is allowable. */ |
| 4636 | HReg r = iselIntExpr_R(env, IRExpr_Const(stmt->Ist.Exit.dst)); |
| 4637 | addInstr(env, AMD64Instr_XAssisted(r, amRIP, cc, Ijk_Boring)); |
| 4638 | } |
| 4639 | return; |
| 4640 | } |
| 4641 | |
| 4642 | /* Case: assisted transfer to arbitrary address */ |
| 4643 | switch (stmt->Ist.Exit.jk) { |
sewardj | 2f6902b | 2012-04-23 09:48:14 +0000 | [diff] [blame] | 4644 | /* Keep this list in sync with that in iselNext below */ |
| 4645 | case Ijk_ClientReq: |
| 4646 | case Ijk_EmWarn: |
| 4647 | case Ijk_NoDecode: |
| 4648 | case Ijk_NoRedir: |
| 4649 | case Ijk_SigSEGV: |
| 4650 | case Ijk_SigTRAP: |
| 4651 | case Ijk_Sys_syscall: |
| 4652 | case Ijk_TInval: |
| 4653 | case Ijk_Yield: |
| 4654 | { |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4655 | HReg r = iselIntExpr_R(env, IRExpr_Const(stmt->Ist.Exit.dst)); |
| 4656 | addInstr(env, AMD64Instr_XAssisted(r, amRIP, cc, stmt->Ist.Exit.jk)); |
| 4657 | return; |
| 4658 | } |
| 4659 | default: |
| 4660 | break; |
| 4661 | } |
| 4662 | |
| 4663 | /* Do we ever expect to see any other kind? */ |
| 4664 | goto stmt_fail; |
sewardj | f8c37f7 | 2005-02-07 18:55:29 +0000 | [diff] [blame] | 4665 | } |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4666 | |
| 4667 | default: break; |
| 4668 | } |
sewardj | af1ceca | 2005-06-30 23:31:27 +0000 | [diff] [blame] | 4669 | stmt_fail: |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4670 | ppIRStmt(stmt); |
| 4671 | vpanic("iselStmt(amd64)"); |
| 4672 | } |
| 4673 | |
| 4674 | |
| 4675 | /*---------------------------------------------------------*/ |
| 4676 | /*--- ISEL: Basic block terminators (Nexts) ---*/ |
| 4677 | /*---------------------------------------------------------*/ |
| 4678 | |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4679 | static void iselNext ( ISelEnv* env, |
| 4680 | IRExpr* next, IRJumpKind jk, Int offsIP ) |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 4681 | { |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 4682 | if (vex_traceflags & VEX_TRACE_VCODE) { |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4683 | vex_printf( "\n-- PUT(%d) = ", offsIP); |
| 4684 | ppIRExpr( next ); |
| 4685 | vex_printf( "; exit-"); |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 4686 | ppIRJumpKind(jk); |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4687 | vex_printf( "\n"); |
sewardj | f67eadf | 2005-02-03 03:53:52 +0000 | [diff] [blame] | 4688 | } |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4689 | |
| 4690 | /* Case: boring transfer to known address */ |
| 4691 | if (next->tag == Iex_Const) { |
| 4692 | IRConst* cdst = next->Iex.Const.con; |
| 4693 | vassert(cdst->tag == Ico_U64); |
| 4694 | if (jk == Ijk_Boring || jk == Ijk_Call) { |
| 4695 | /* Boring transfer to known address */ |
| 4696 | AMD64AMode* amRIP = AMD64AMode_IR(offsIP, hregAMD64_RBP()); |
| 4697 | if (env->chainingAllowed) { |
| 4698 | /* .. almost always true .. */ |
| 4699 | /* Skip the event check at the dst if this is a forwards |
| 4700 | edge. */ |
| 4701 | Bool toFastEP |
| 4702 | = ((Addr64)cdst->Ico.U64) > env->max_ga; |
| 4703 | if (0) vex_printf("%s", toFastEP ? "X" : "."); |
| 4704 | addInstr(env, AMD64Instr_XDirect(cdst->Ico.U64, |
| 4705 | amRIP, Acc_ALWAYS, |
| 4706 | toFastEP)); |
| 4707 | } else { |
| 4708 | /* .. very occasionally .. */ |
| 4709 | /* We can't use chaining, so ask for an indirect transfer, |
| 4710 | as that's the cheapest alternative that is |
| 4711 | allowable. */ |
| 4712 | HReg r = iselIntExpr_R(env, next); |
| 4713 | addInstr(env, AMD64Instr_XAssisted(r, amRIP, Acc_ALWAYS, |
| 4714 | Ijk_Boring)); |
| 4715 | } |
| 4716 | return; |
| 4717 | } |
| 4718 | } |
| 4719 | |
| 4720 | /* Case: call/return (==boring) transfer to any address */ |
| 4721 | switch (jk) { |
| 4722 | case Ijk_Boring: case Ijk_Ret: case Ijk_Call: { |
| 4723 | HReg r = iselIntExpr_R(env, next); |
| 4724 | AMD64AMode* amRIP = AMD64AMode_IR(offsIP, hregAMD64_RBP()); |
| 4725 | if (env->chainingAllowed) { |
| 4726 | addInstr(env, AMD64Instr_XIndir(r, amRIP, Acc_ALWAYS)); |
| 4727 | } else { |
| 4728 | addInstr(env, AMD64Instr_XAssisted(r, amRIP, Acc_ALWAYS, |
| 4729 | Ijk_Boring)); |
| 4730 | } |
| 4731 | return; |
| 4732 | } |
| 4733 | default: |
| 4734 | break; |
| 4735 | } |
| 4736 | |
sewardj | 2f6902b | 2012-04-23 09:48:14 +0000 | [diff] [blame] | 4737 | /* Case: assisted transfer to arbitrary address */ |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4738 | switch (jk) { |
sewardj | 2f6902b | 2012-04-23 09:48:14 +0000 | [diff] [blame] | 4739 | /* Keep this list in sync with that for Ist_Exit above */ |
| 4740 | case Ijk_ClientReq: |
| 4741 | case Ijk_EmWarn: |
sewardj | 3d0e38e | 2012-04-21 07:38:29 +0000 | [diff] [blame] | 4742 | case Ijk_NoDecode: |
sewardj | 2f6902b | 2012-04-23 09:48:14 +0000 | [diff] [blame] | 4743 | case Ijk_NoRedir: |
| 4744 | case Ijk_SigSEGV: |
| 4745 | case Ijk_SigTRAP: |
| 4746 | case Ijk_Sys_syscall: |
| 4747 | case Ijk_TInval: |
| 4748 | case Ijk_Yield: { |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4749 | HReg r = iselIntExpr_R(env, next); |
| 4750 | AMD64AMode* amRIP = AMD64AMode_IR(offsIP, hregAMD64_RBP()); |
| 4751 | addInstr(env, AMD64Instr_XAssisted(r, amRIP, Acc_ALWAYS, jk)); |
| 4752 | return; |
| 4753 | } |
| 4754 | default: |
| 4755 | break; |
| 4756 | } |
| 4757 | |
| 4758 | vex_printf( "\n-- PUT(%d) = ", offsIP); |
| 4759 | ppIRExpr( next ); |
| 4760 | vex_printf( "; exit-"); |
| 4761 | ppIRJumpKind(jk); |
| 4762 | vex_printf( "\n"); |
| 4763 | vassert(0); // are we expecting any other kind? |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4764 | } |
| 4765 | |
| 4766 | |
| 4767 | /*---------------------------------------------------------*/ |
| 4768 | /*--- Insn selector top-level ---*/ |
| 4769 | /*---------------------------------------------------------*/ |
| 4770 | |
sewardj | dd40fdf | 2006-12-24 02:20:24 +0000 | [diff] [blame] | 4771 | /* Translate an entire SB to amd64 code. */ |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4772 | |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4773 | HInstrArray* iselSB_AMD64 ( IRSB* bb, |
| 4774 | VexArch arch_host, |
| 4775 | VexArchInfo* archinfo_host, |
| 4776 | VexAbiInfo* vbi/*UNUSED*/, |
| 4777 | Int offs_Host_EvC_Counter, |
| 4778 | Int offs_Host_EvC_FailAddr, |
| 4779 | Bool chainingAllowed, |
| 4780 | Bool addProfInc, |
| 4781 | Addr64 max_ga ) |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4782 | { |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4783 | Int i, j; |
| 4784 | HReg hreg, hregHI; |
| 4785 | ISelEnv* env; |
| 4786 | UInt hwcaps_host = archinfo_host->hwcaps; |
| 4787 | AMD64AMode *amCounter, *amFailAddr; |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4788 | |
| 4789 | /* sanity ... */ |
sewardj | 8f07359 | 2006-05-01 02:14:17 +0000 | [diff] [blame] | 4790 | vassert(arch_host == VexArchAMD64); |
sewardj | 536fbab | 2010-07-29 15:39:05 +0000 | [diff] [blame] | 4791 | vassert(0 == (hwcaps_host |
| 4792 | & ~(VEX_HWCAPS_AMD64_SSE3 |
| 4793 | | VEX_HWCAPS_AMD64_CX16 |
sewardj | f350a42 | 2012-04-26 14:16:52 +0000 | [diff] [blame] | 4794 | | VEX_HWCAPS_AMD64_LZCNT |
sewardj | 818c730 | 2013-03-26 13:53:18 +0000 | [diff] [blame] | 4795 | | VEX_HWCAPS_AMD64_AVX |
sewardj | cc3d219 | 2013-03-27 11:37:33 +0000 | [diff] [blame] | 4796 | | VEX_HWCAPS_AMD64_RDTSCP |
| 4797 | | VEX_HWCAPS_AMD64_BMI |
| 4798 | | VEX_HWCAPS_AMD64_AVX2))); |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4799 | |
| 4800 | /* Make up an initial environment to use. */ |
sewardj | 9a036bf | 2005-03-14 18:19:08 +0000 | [diff] [blame] | 4801 | env = LibVEX_Alloc(sizeof(ISelEnv)); |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4802 | env->vreg_ctr = 0; |
| 4803 | |
| 4804 | /* Set up output code array. */ |
| 4805 | env->code = newHInstrArray(); |
| 4806 | |
| 4807 | /* Copy BB's type env. */ |
| 4808 | env->type_env = bb->tyenv; |
| 4809 | |
| 4810 | /* Make up an IRTemp -> virtual HReg mapping. This doesn't |
| 4811 | change as we go along. */ |
| 4812 | env->n_vregmap = bb->tyenv->types_used; |
| 4813 | env->vregmap = LibVEX_Alloc(env->n_vregmap * sizeof(HReg)); |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 4814 | env->vregmapHI = LibVEX_Alloc(env->n_vregmap * sizeof(HReg)); |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4815 | |
| 4816 | /* and finally ... */ |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4817 | env->chainingAllowed = chainingAllowed; |
| 4818 | env->hwcaps = hwcaps_host; |
| 4819 | env->max_ga = max_ga; |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4820 | |
| 4821 | /* For each IR temporary, allocate a suitably-kinded virtual |
| 4822 | register. */ |
| 4823 | j = 0; |
| 4824 | for (i = 0; i < env->n_vregmap; i++) { |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 4825 | hregHI = hreg = INVALID_HREG; |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4826 | switch (bb->tyenv->types[i]) { |
| 4827 | case Ity_I1: |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4828 | case Ity_I8: case Ity_I16: case Ity_I32: case Ity_I64: |
| 4829 | hreg = mkHReg(j++, HRcInt64, True); |
| 4830 | break; |
| 4831 | case Ity_I128: |
| 4832 | hreg = mkHReg(j++, HRcInt64, True); |
| 4833 | hregHI = mkHReg(j++, HRcInt64, True); |
| 4834 | break; |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4835 | case Ity_F32: |
sewardj | 1830386 | 2005-02-21 12:36:54 +0000 | [diff] [blame] | 4836 | case Ity_F64: |
sewardj | c4530ae | 2012-05-21 10:18:49 +0000 | [diff] [blame] | 4837 | case Ity_V128: |
| 4838 | hreg = mkHReg(j++, HRcVec128, True); |
| 4839 | break; |
| 4840 | case Ity_V256: |
| 4841 | hreg = mkHReg(j++, HRcVec128, True); |
| 4842 | hregHI = mkHReg(j++, HRcVec128, True); |
| 4843 | break; |
| 4844 | default: |
| 4845 | ppIRType(bb->tyenv->types[i]); |
| 4846 | vpanic("iselBB(amd64): IRTemp type"); |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4847 | } |
| 4848 | env->vregmap[i] = hreg; |
sewardj | 9b96767 | 2005-02-08 11:13:09 +0000 | [diff] [blame] | 4849 | env->vregmapHI[i] = hregHI; |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4850 | } |
| 4851 | env->vreg_ctr = j; |
| 4852 | |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4853 | /* The very first instruction must be an event check. */ |
| 4854 | amCounter = AMD64AMode_IR(offs_Host_EvC_Counter, hregAMD64_RBP()); |
| 4855 | amFailAddr = AMD64AMode_IR(offs_Host_EvC_FailAddr, hregAMD64_RBP()); |
| 4856 | addInstr(env, AMD64Instr_EvCheck(amCounter, amFailAddr)); |
| 4857 | |
| 4858 | /* Possibly a block counter increment (for profiling). At this |
| 4859 | point we don't know the address of the counter, so just pretend |
| 4860 | it is zero. It will have to be patched later, but before this |
| 4861 | translation is used, by a call to LibVEX_patchProfCtr. */ |
| 4862 | if (addProfInc) { |
| 4863 | addInstr(env, AMD64Instr_ProfInc()); |
| 4864 | } |
| 4865 | |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4866 | /* Ok, finally we can iterate over the statements. */ |
| 4867 | for (i = 0; i < bb->stmts_used; i++) |
| 4868 | if (bb->stmts[i]) |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4869 | iselStmt(env, bb->stmts[i]); |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4870 | |
sewardj | c6f970f | 2012-04-02 21:54:49 +0000 | [diff] [blame] | 4871 | iselNext(env, bb->next, bb->jumpkind, bb->offsIP); |
sewardj | c33671d | 2005-02-01 20:30:00 +0000 | [diff] [blame] | 4872 | |
| 4873 | /* record the number of vregs we used. */ |
| 4874 | env->code->n_vregs = env->vreg_ctr; |
| 4875 | return env->code; |
| 4876 | } |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 4877 | |
| 4878 | |
| 4879 | /*---------------------------------------------------------------*/ |
sewardj | cef7d3e | 2009-07-02 12:21:59 +0000 | [diff] [blame] | 4880 | /*--- end host_amd64_isel.c ---*/ |
sewardj | a3e9830 | 2005-02-01 15:55:05 +0000 | [diff] [blame] | 4881 | /*---------------------------------------------------------------*/ |