blob: d06542d6b110ab0ad5d0da91916ccd948976d233 [file] [log] [blame]
Chris Lattner7c5a3d32005-08-16 17:14:42 +00001//===-- PPC32ISelLowering.cpp - PPC32 DAG Lowering Implementation ---------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file was developed by Chris Lattner and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implements the PPC32ISelLowering class.
11//
12//===----------------------------------------------------------------------===//
13
14#include "PPC32ISelLowering.h"
15#include "PPC32TargetMachine.h"
16#include "llvm/CodeGen/MachineFrameInfo.h"
17#include "llvm/CodeGen/MachineFunction.h"
18#include "llvm/CodeGen/SelectionDAG.h"
19#include "llvm/Function.h"
20
21using namespace llvm;
22
23PPC32TargetLowering::PPC32TargetLowering(TargetMachine &TM)
24 : TargetLowering(TM) {
25
26 // Fold away setcc operations if possible.
27 setSetCCIsExpensive();
28
29 // Set up the register classes.
30 addRegisterClass(MVT::i32, PPC32::GPRCRegisterClass);
31 addRegisterClass(MVT::f32, PPC32::FPRCRegisterClass);
32 addRegisterClass(MVT::f64, PPC32::FPRCRegisterClass);
33
34 // PowerPC has no intrinsics for these particular operations
35 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
36 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
37 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
38
39 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
40 setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand);
41 setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand);
42
43 // PowerPC has no SREM/UREM instructions
44 setOperationAction(ISD::SREM, MVT::i32, Expand);
45 setOperationAction(ISD::UREM, MVT::i32, Expand);
46
47 // We don't support sin/cos/sqrt/fmod
48 setOperationAction(ISD::FSIN , MVT::f64, Expand);
49 setOperationAction(ISD::FCOS , MVT::f64, Expand);
50 setOperationAction(ISD::SREM , MVT::f64, Expand);
51 setOperationAction(ISD::FSIN , MVT::f32, Expand);
52 setOperationAction(ISD::FCOS , MVT::f32, Expand);
53 setOperationAction(ISD::SREM , MVT::f32, Expand);
54
55 // If we're enabling GP optimizations, use hardware square root
56 if (!TM.getSubtarget<PPCSubtarget>().isGigaProcessor()) {
57 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
58 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
59 }
60
61 // PowerPC does not have CTPOP or CTTZ
62 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
63 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
64
65 // PowerPC does not have Select
66 setOperationAction(ISD::SELECT, MVT::i32, Expand);
67 setOperationAction(ISD::SELECT, MVT::f32, Expand);
68 setOperationAction(ISD::SELECT, MVT::f64, Expand);
69
70 // PowerPC does not have FP_TO_UINT
71 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
72
73 setSetCCResultContents(ZeroOrOneSetCCResult);
74 addLegalFPImmediate(+0.0); // Necessary for FSEL
75 addLegalFPImmediate(-0.0); //
76
77 computeRegisterProperties();
78}
79
80std::vector<SDOperand>
81PPC32TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
82 //
83 // add beautiful description of PPC stack frame format, or at least some docs
84 //
85 MachineFunction &MF = DAG.getMachineFunction();
86 MachineFrameInfo *MFI = MF.getFrameInfo();
87 MachineBasicBlock& BB = MF.front();
88 std::vector<SDOperand> ArgValues;
89
90 // Due to the rather complicated nature of the PowerPC ABI, rather than a
91 // fixed size array of physical args, for the sake of simplicity let the STL
92 // handle tracking them for us.
93 std::vector<unsigned> argVR, argPR, argOp;
94 unsigned ArgOffset = 24;
95 unsigned GPR_remaining = 8;
96 unsigned FPR_remaining = 13;
97 unsigned GPR_idx = 0, FPR_idx = 0;
98 static const unsigned GPR[] = {
99 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
100 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
101 };
102 static const unsigned FPR[] = {
103 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
104 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
105 };
106
107 // Add DAG nodes to load the arguments... On entry to a function on PPC,
108 // the arguments start at offset 24, although they are likely to be passed
109 // in registers.
110 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
111 SDOperand newroot, argt;
112 unsigned ObjSize;
113 bool needsLoad = false;
114 bool ArgLive = !I->use_empty();
115 MVT::ValueType ObjectVT = getValueType(I->getType());
116
117 switch (ObjectVT) {
118 default: assert(0 && "Unhandled argument type!");
119 case MVT::i1:
120 case MVT::i8:
121 case MVT::i16:
122 case MVT::i32:
123 ObjSize = 4;
124 if (!ArgLive) break;
125 if (GPR_remaining > 0) {
126 MF.addLiveIn(GPR[GPR_idx]);
127 argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32,
128 DAG.getRoot());
129 if (ObjectVT != MVT::i32)
130 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
131 } else {
132 needsLoad = true;
133 }
134 break;
135 case MVT::i64: ObjSize = 8;
136 if (!ArgLive) break;
137 if (GPR_remaining > 0) {
138 SDOperand argHi, argLo;
139 MF.addLiveIn(GPR[GPR_idx]);
140 argHi = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
141 // If we have two or more remaining argument registers, then both halves
142 // of the i64 can be sourced from there. Otherwise, the lower half will
143 // have to come off the stack. This can happen when an i64 is preceded
144 // by 28 bytes of arguments.
145 if (GPR_remaining > 1) {
146 MF.addLiveIn(GPR[GPR_idx+1]);
147 argLo = DAG.getCopyFromReg(GPR[GPR_idx+1], MVT::i32, argHi);
148 } else {
149 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
150 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
151 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
152 DAG.getSrcValue(NULL));
153 }
154 // Build the outgoing arg thingy
155 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
156 newroot = argLo;
157 } else {
158 needsLoad = true;
159 }
160 break;
161 case MVT::f32:
162 case MVT::f64:
163 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
164 if (!ArgLive) break;
165 if (FPR_remaining > 0) {
166 MF.addLiveIn(FPR[FPR_idx]);
167 argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT,
168 DAG.getRoot());
169 --FPR_remaining;
170 ++FPR_idx;
171 } else {
172 needsLoad = true;
173 }
174 break;
175 }
176
177 // We need to load the argument to a virtual register if we determined above
178 // that we ran out of physical registers of the appropriate type
179 if (needsLoad) {
180 unsigned SubregOffset = 0;
181 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
182 if (ObjectVT == MVT::i16) SubregOffset = 2;
183 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
184 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
185 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
186 DAG.getConstant(SubregOffset, MVT::i32));
187 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
188 DAG.getSrcValue(NULL));
189 }
190
191 // Every 4 bytes of argument space consumes one of the GPRs available for
192 // argument passing.
193 if (GPR_remaining > 0) {
194 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
195 GPR_remaining -= delta;
196 GPR_idx += delta;
197 }
198 ArgOffset += ObjSize;
199 if (newroot.Val)
200 DAG.setRoot(newroot.getValue(1));
201
202 ArgValues.push_back(argt);
203 }
204
205 // If the function takes variable number of arguments, make a frame index for
206 // the start of the first vararg value... for expansion of llvm.va_start.
207 if (F.isVarArg()) {
208 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
209 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
210 // If this function is vararg, store any remaining integer argument regs
211 // to their spots on the stack so that they may be loaded by deferencing the
212 // result of va_next.
213 std::vector<SDOperand> MemOps;
214 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
215 MF.addLiveIn(GPR[GPR_idx]);
216 SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
217 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
218 Val, FIN, DAG.getSrcValue(NULL));
219 MemOps.push_back(Store);
220 // Increment the address by four for the next argument to store
221 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
222 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
223 }
224 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
225 }
226
227 // Finally, inform the code generator which regs we return values in.
228 switch (getValueType(F.getReturnType())) {
229 default: assert(0 && "Unknown type!");
230 case MVT::isVoid: break;
231 case MVT::i1:
232 case MVT::i8:
233 case MVT::i16:
234 case MVT::i32:
235 MF.addLiveOut(PPC::R3);
236 break;
237 case MVT::i64:
238 MF.addLiveOut(PPC::R3);
239 MF.addLiveOut(PPC::R4);
240 break;
241 case MVT::f32:
242 case MVT::f64:
243 MF.addLiveOut(PPC::F1);
244 break;
245 }
246
247 return ArgValues;
248}
249
250std::pair<SDOperand, SDOperand>
251PPC32TargetLowering::LowerCallTo(SDOperand Chain,
252 const Type *RetTy, bool isVarArg,
253 unsigned CallingConv, bool isTailCall,
254 SDOperand Callee, ArgListTy &Args,
255 SelectionDAG &DAG) {
256 // args_to_use will accumulate outgoing args for the ISD::CALL case in
257 // SelectExpr to use to put the arguments in the appropriate registers.
258 std::vector<SDOperand> args_to_use;
259
260 // Count how many bytes are to be pushed on the stack, including the linkage
261 // area, and parameter passing area.
262 unsigned NumBytes = 24;
263
264 if (Args.empty()) {
265 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
266 DAG.getConstant(NumBytes, getPointerTy()));
267 } else {
268 for (unsigned i = 0, e = Args.size(); i != e; ++i)
269 switch (getValueType(Args[i].second)) {
270 default: assert(0 && "Unknown value type!");
271 case MVT::i1:
272 case MVT::i8:
273 case MVT::i16:
274 case MVT::i32:
275 case MVT::f32:
276 NumBytes += 4;
277 break;
278 case MVT::i64:
279 case MVT::f64:
280 NumBytes += 8;
281 break;
282 }
283
284 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
285 // plus 32 bytes of argument space in case any called code gets funky on us.
286 // (Required by ABI to support var arg)
287 if (NumBytes < 56) NumBytes = 56;
288
289 // Adjust the stack pointer for the new arguments...
290 // These operations are automatically eliminated by the prolog/epilog pass
291 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
292 DAG.getConstant(NumBytes, getPointerTy()));
293
294 // Set up a copy of the stack pointer for use loading and storing any
295 // arguments that may not fit in the registers available for argument
296 // passing.
297 SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
298 DAG.getEntryNode());
299
300 // Figure out which arguments are going to go in registers, and which in
301 // memory. Also, if this is a vararg function, floating point operations
302 // must be stored to our stack, and loaded into integer regs as well, if
303 // any integer regs are available for argument passing.
304 unsigned ArgOffset = 24;
305 unsigned GPR_remaining = 8;
306 unsigned FPR_remaining = 13;
307
308 std::vector<SDOperand> MemOps;
309 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
310 // PtrOff will be used to store the current argument to the stack if a
311 // register cannot be found for it.
312 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
313 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
314 MVT::ValueType ArgVT = getValueType(Args[i].second);
315
316 switch (ArgVT) {
317 default: assert(0 && "Unexpected ValueType for argument!");
318 case MVT::i1:
319 case MVT::i8:
320 case MVT::i16:
321 // Promote the integer to 32 bits. If the input type is signed use a
322 // sign extend, otherwise use a zero extend.
323 if (Args[i].second->isSigned())
324 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
325 else
326 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
327 // FALL THROUGH
328 case MVT::i32:
329 if (GPR_remaining > 0) {
330 args_to_use.push_back(Args[i].first);
331 --GPR_remaining;
332 } else {
333 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
334 Args[i].first, PtrOff,
335 DAG.getSrcValue(NULL)));
336 }
337 ArgOffset += 4;
338 break;
339 case MVT::i64:
340 // If we have one free GPR left, we can place the upper half of the i64
341 // in it, and store the other half to the stack. If we have two or more
342 // free GPRs, then we can pass both halves of the i64 in registers.
343 if (GPR_remaining > 0) {
344 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
345 Args[i].first, DAG.getConstant(1, MVT::i32));
346 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
347 Args[i].first, DAG.getConstant(0, MVT::i32));
348 args_to_use.push_back(Hi);
349 --GPR_remaining;
350 if (GPR_remaining > 0) {
351 args_to_use.push_back(Lo);
352 --GPR_remaining;
353 } else {
354 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
355 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
356 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
357 Lo, PtrOff, DAG.getSrcValue(NULL)));
358 }
359 } else {
360 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
361 Args[i].first, PtrOff,
362 DAG.getSrcValue(NULL)));
363 }
364 ArgOffset += 8;
365 break;
366 case MVT::f32:
367 case MVT::f64:
368 if (FPR_remaining > 0) {
369 args_to_use.push_back(Args[i].first);
370 --FPR_remaining;
371 if (isVarArg) {
372 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
373 Args[i].first, PtrOff,
374 DAG.getSrcValue(NULL));
375 MemOps.push_back(Store);
376 // Float varargs are always shadowed in available integer registers
377 if (GPR_remaining > 0) {
378 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
379 DAG.getSrcValue(NULL));
380 MemOps.push_back(Load);
381 args_to_use.push_back(Load);
382 --GPR_remaining;
383 }
384 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
385 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
386 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
387 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
388 DAG.getSrcValue(NULL));
389 MemOps.push_back(Load);
390 args_to_use.push_back(Load);
391 --GPR_remaining;
392 }
393 } else {
394 // If we have any FPRs remaining, we may also have GPRs remaining.
395 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
396 // GPRs.
397 if (GPR_remaining > 0) {
398 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
399 --GPR_remaining;
400 }
401 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
402 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
403 --GPR_remaining;
404 }
405 }
406 } else {
407 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
408 Args[i].first, PtrOff,
409 DAG.getSrcValue(NULL)));
410 }
411 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
412 break;
413 }
414 }
415 if (!MemOps.empty())
416 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
417 }
418
419 std::vector<MVT::ValueType> RetVals;
420 MVT::ValueType RetTyVT = getValueType(RetTy);
421 if (RetTyVT != MVT::isVoid)
422 RetVals.push_back(RetTyVT);
423 RetVals.push_back(MVT::Other);
424
425 SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
426 Chain, Callee, args_to_use), 0);
427 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
428 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
429 DAG.getConstant(NumBytes, getPointerTy()));
430 return std::make_pair(TheCall, Chain);
431}
432
433SDOperand PPC32TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
434 Value *VAListV, SelectionDAG &DAG) {
435 // vastart just stores the address of the VarArgsFrameIndex slot into the
436 // memory location argument.
437 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
438 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR, VAListP,
439 DAG.getSrcValue(VAListV));
440}
441
442std::pair<SDOperand,SDOperand>
443PPC32TargetLowering::LowerVAArg(SDOperand Chain,
444 SDOperand VAListP, Value *VAListV,
445 const Type *ArgTy, SelectionDAG &DAG) {
446 MVT::ValueType ArgVT = getValueType(ArgTy);
447
448 SDOperand VAList =
449 DAG.getLoad(MVT::i32, Chain, VAListP, DAG.getSrcValue(VAListV));
450 SDOperand Result = DAG.getLoad(ArgVT, Chain, VAList, DAG.getSrcValue(NULL));
451 unsigned Amt;
452 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
453 Amt = 4;
454 else {
455 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
456 "Other types should have been promoted for varargs!");
457 Amt = 8;
458 }
459 VAList = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
460 DAG.getConstant(Amt, VAList.getValueType()));
461 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
462 VAList, VAListP, DAG.getSrcValue(VAListV));
463 return std::make_pair(Result, Chain);
464}
465
466
467std::pair<SDOperand, SDOperand> PPC32TargetLowering::
468LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
469 SelectionDAG &DAG) {
470 assert(0 && "LowerFrameReturnAddress unimplemented");
471 abort();
472}