Long double patch 8 of N: make it partially work in
SSE mode (all but conversions <-> other FP types, I think):
>>Do not mark all-80-bit operations as "Requires[FPStack]"
(which really means "not SSE").
>>Refactor load-and-extend to facilitate this.
>>Update comments.
>>Handle long double in SSE when computing FP_REG_KILL.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40906 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86InstrFPStack.td b/lib/Target/X86/X86InstrFPStack.td
index eaec18b..b9148fe 100644
--- a/lib/Target/X86/X86InstrFPStack.td
+++ b/lib/Target/X86/X86InstrFPStack.td
@@ -2,7 +2,7 @@
//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the Evan Cheng and is distributed under
+// This file was developed by Evan Cheng and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
@@ -48,10 +48,6 @@
def X86fp_to_i64mem : SDNode<"X86ISD::FP_TO_INT64_IN_MEM", SDTX86FpToIMem,
[SDNPHasChain]>;
-def extloadf80f32 : PatFrag<(ops node:$ptr), (f80 (extloadf32 node:$ptr))>;
-def extloadf80f64 : PatFrag<(ops node:$ptr), (f80 (extloadf64 node:$ptr))>;
-def extloadf64f32 : PatFrag<(ops node:$ptr), (f64 (extloadf32 node:$ptr))>;
-
//===----------------------------------------------------------------------===//
// FPStack pattern fragments
//===----------------------------------------------------------------------===//
@@ -116,18 +112,18 @@
let Defs = [FP0, FP1, FP2, FP3, FP4, FP5, FP6] in
def FP_REG_KILL : I<0, Pseudo, (outs), (ins), "#FP_REG_KILL", []>;
-// All FP Stack operations are represented with three instructions here. The
-// first two instructions, generated by the instruction selector, uses "RFP32"
-// or "RFP64" registers: traditional register files to reference 32-bit or
-// 64-bit floating point values. These sizes apply to the values, not the
-// registers, which are always 64 bits; RFP32 and RFP64 can be copied to
-// each other without losing information. These instructions are all psuedo
-// instructions and use the "_Fp" suffix.
-// In some cases there are additional variants with a mixture of 32-bit and
-// 64-bit registers.
+// All FP Stack operations are represented with four instructions here. The
+// first three instructions, generated by the instruction selector, use "RFP32"
+// "RFP64" or "RFP80" registers: traditional register files to reference 32-bit,
+// 64-bit or 80-bit floating point values. These sizes apply to the values,
+// not the registers, which are always 80 bits; RFP32, RFP64 and RFP80 can be
+// copied to each other without losing information. These instructions are all
+// pseudo instructions and use the "_Fp" suffix.
+// In some cases there are additional variants with a mixture of different
+// register sizes.
// The second instruction is defined with FPI, which is the actual instruction
// emitted by the assembler. These use "RST" registers, although frequently
-// the actual register(s) used are implicit. These are always 64-bits.
+// the actual register(s) used are implicit. These are always 80 bits.
// The FP stackifier pass converts one to the other after register allocation
// occurs.
//
@@ -135,7 +131,7 @@
// a pattern) and the FPI instruction should have emission info (e.g. opcode
// encoding and asm printing info).
-// Random Pseudo Instructions.
+// Pseudo Instructions for FP stack return values.
def FpGETRESULT32 : FpI_<(outs RFP32:$dst), (ins), SpecialFP,
[(set RFP32:$dst, X86fpget)]>; // FPR = ST(0)
@@ -155,6 +151,8 @@
[(X86fpset RFP80:$src)]>, Imp<[], [ST0]>;// ST(0) = FPR
// FpI - Floating Point Psuedo Instruction template. Predicated on FPStack.
+// Note that f80-only instructions are used even in SSE mode and use FpI_
+// not this predicate.
class FpI<dag outs, dag ins, FPFormat fp, list<dag> pattern> :
FpI_<outs, ins, fp, pattern>, Requires<[FPStack]>;
@@ -167,7 +165,7 @@
def MOV_Fp3280 : FpI<(outs RFP80:$dst), (ins RFP32:$src), SpecialFP, []>;
def MOV_Fp8064 : FpI<(outs RFP64:$dst), (ins RFP80:$src), SpecialFP, []>;
def MOV_Fp6480 : FpI<(outs RFP80:$dst), (ins RFP64:$src), SpecialFP, []>;
-def MOV_Fp8080 : FpI<(outs RFP80:$dst), (ins RFP80:$src), SpecialFP, []>;
+def MOV_Fp8080 : FpI_<(outs RFP80:$dst), (ins RFP80:$src), SpecialFP, []>;
// Factoring for arithmetic.
multiclass FPBinary_rr<SDNode OpNode> {
@@ -177,7 +175,7 @@
[(set RFP32:$dst, (OpNode RFP32:$src1, RFP32:$src2))]>;
def _Fp64 : FpI<(outs RFP64:$dst), (ins RFP64:$src1, RFP64:$src2), TwoArgFP,
[(set RFP64:$dst, (OpNode RFP64:$src1, RFP64:$src2))]>;
-def _Fp80 : FpI<(outs RFP80:$dst), (ins RFP80:$src1, RFP80:$src2), TwoArgFP,
+def _Fp80 : FpI_<(outs RFP80:$dst), (ins RFP80:$src1, RFP80:$src2), TwoArgFP,
[(set RFP80:$dst, (OpNode RFP80:$src1, RFP80:$src2))]>;
}
// The FopST0 series are not included here because of the irregularities
@@ -193,13 +191,13 @@
(OpNode RFP64:$src1, (loadf64 addr:$src2)))]>;
def _Fp64m32: FpI<(outs RFP64:$dst), (ins RFP64:$src1, f32mem:$src2), OneArgFPRW,
[(set RFP64:$dst,
- (OpNode RFP64:$src1, (extloadf64f32 addr:$src2)))]>;
-def _Fp80m32: FpI<(outs RFP80:$dst), (ins RFP80:$src1, f32mem:$src2), OneArgFPRW,
+ (OpNode RFP64:$src1, (f64 (extloadf32 addr:$src2))))]>;
+def _Fp80m32: FpI_<(outs RFP80:$dst), (ins RFP80:$src1, f32mem:$src2), OneArgFPRW,
[(set RFP80:$dst,
- (OpNode RFP80:$src1, (extloadf80f32 addr:$src2)))]>;
-def _Fp80m64: FpI<(outs RFP80:$dst), (ins RFP80:$src1, f64mem:$src2), OneArgFPRW,
+ (OpNode RFP80:$src1, (f80 (extloadf32 addr:$src2))))]>;
+def _Fp80m64: FpI_<(outs RFP80:$dst), (ins RFP80:$src1, f64mem:$src2), OneArgFPRW,
[(set RFP80:$dst,
- (OpNode RFP80:$src1, (extloadf80f64 addr:$src2)))]>;
+ (OpNode RFP80:$src1, (f80 (extloadf64 addr:$src2))))]>;
def _F32m : FPI<0xD8, fp, (outs), (ins f32mem:$src),
!strconcat("f", !strconcat(asmstring, "{s}\t$src"))>;
def _F64m : FPI<0xDC, fp, (outs), (ins f64mem:$src),
@@ -217,10 +215,10 @@
def _FpI32m64 : FpI<(outs RFP64:$dst), (ins RFP64:$src1, i32mem:$src2), OneArgFPRW,
[(set RFP64:$dst, (OpNode RFP64:$src1,
(X86fild addr:$src2, i32)))]>;
-def _FpI16m80 : FpI<(outs RFP80:$dst), (ins RFP80:$src1, i16mem:$src2), OneArgFPRW,
+def _FpI16m80 : FpI_<(outs RFP80:$dst), (ins RFP80:$src1, i16mem:$src2), OneArgFPRW,
[(set RFP80:$dst, (OpNode RFP80:$src1,
(X86fild addr:$src2, i16)))]>;
-def _FpI32m80 : FpI<(outs RFP80:$dst), (ins RFP80:$src1, i32mem:$src2), OneArgFPRW,
+def _FpI32m80 : FpI_<(outs RFP80:$dst), (ins RFP80:$src1, i32mem:$src2), OneArgFPRW,
[(set RFP80:$dst, (OpNode RFP80:$src1,
(X86fild addr:$src2, i32)))]>;
def _FI16m : FPI<0xDE, fp, (outs), (ins i16mem:$src),
@@ -275,7 +273,7 @@
[(set RFP32:$dst, (OpNode RFP32:$src))]>;
def _Fp64 : FpI<(outs RFP64:$dst), (ins RFP64:$src), OneArgFPRW,
[(set RFP64:$dst, (OpNode RFP64:$src))]>;
-def _Fp80 : FpI<(outs RFP80:$dst), (ins RFP80:$src), OneArgFPRW,
+def _Fp80 : FpI_<(outs RFP80:$dst), (ins RFP80:$src), OneArgFPRW,
[(set RFP80:$dst, (OpNode RFP80:$src))]>;
def _F : FPI<opcode, RawFrm, (outs), (ins), asmstring>, D9;
}
@@ -290,7 +288,7 @@
[]>;
def TST_Fp64 : FpI<(outs), (ins RFP64:$src), OneArgFP,
[]>;
-def TST_Fp80 : FpI<(outs), (ins RFP80:$src), OneArgFP,
+def TST_Fp80 : FpI_<(outs), (ins RFP80:$src), OneArgFP,
[]>;
def TST_F : FPI<0xE4, RawFrm, (outs), (ins), "ftst">, D9;
@@ -302,7 +300,7 @@
def _Fp64 : FpI<(outs RFP64:$dst), (ins RFP64:$src1, RFP64:$src2), CondMovFP,
[(set RFP64:$dst, (X86cmov RFP64:$src1, RFP64:$src2,
cc))]>;
- def _Fp80 : FpI<(outs RFP80:$dst), (ins RFP80:$src1, RFP80:$src2), CondMovFP,
+ def _Fp80 : FpI_<(outs RFP80:$dst), (ins RFP80:$src1, RFP80:$src2), CondMovFP,
[(set RFP80:$dst, (X86cmov RFP80:$src1, RFP80:$src2,
cc))]>;
}
@@ -340,8 +338,14 @@
[(set RFP32:$dst, (loadf32 addr:$src))]>;
def LD_Fp64m : FpI<(outs RFP64:$dst), (ins f64mem:$src), ZeroArgFP,
[(set RFP64:$dst, (loadf64 addr:$src))]>;
-def LD_Fp80m : FpI<(outs RFP80:$dst), (ins f80mem:$src), ZeroArgFP,
+def LD_Fp80m : FpI_<(outs RFP80:$dst), (ins f80mem:$src), ZeroArgFP,
[(set RFP80:$dst, (loadf80 addr:$src))]>;
+def LD_Fp32m64 : FpI<(outs RFP64:$dst), (ins f32mem:$src), ZeroArgFP,
+ [(set RFP64:$dst, (f64 (extloadf32 addr:$src)))]>;
+def LD_Fp64m80 : FpI_<(outs RFP80:$dst), (ins f64mem:$src), ZeroArgFP,
+ [(set RFP80:$dst, (f80 (extloadf64 addr:$src)))]>;
+def LD_Fp32m80 : FpI_<(outs RFP80:$dst), (ins f32mem:$src), ZeroArgFP,
+ [(set RFP80:$dst, (f80 (extloadf32 addr:$src)))]>;
def ILD_Fp16m32: FpI<(outs RFP32:$dst), (ins i16mem:$src), ZeroArgFP,
[(set RFP32:$dst, (X86fild addr:$src, i16))]>;
def ILD_Fp32m32: FpI<(outs RFP32:$dst), (ins i32mem:$src), ZeroArgFP,
@@ -354,11 +358,11 @@
[(set RFP64:$dst, (X86fild addr:$src, i32))]>;
def ILD_Fp64m64: FpI<(outs RFP64:$dst), (ins i64mem:$src), ZeroArgFP,
[(set RFP64:$dst, (X86fild addr:$src, i64))]>;
-def ILD_Fp16m80: FpI<(outs RFP80:$dst), (ins i16mem:$src), ZeroArgFP,
+def ILD_Fp16m80: FpI_<(outs RFP80:$dst), (ins i16mem:$src), ZeroArgFP,
[(set RFP80:$dst, (X86fild addr:$src, i16))]>;
-def ILD_Fp32m80: FpI<(outs RFP80:$dst), (ins i32mem:$src), ZeroArgFP,
+def ILD_Fp32m80: FpI_<(outs RFP80:$dst), (ins i32mem:$src), ZeroArgFP,
[(set RFP80:$dst, (X86fild addr:$src, i32))]>;
-def ILD_Fp64m80: FpI<(outs RFP80:$dst), (ins i64mem:$src), ZeroArgFP,
+def ILD_Fp64m80: FpI_<(outs RFP80:$dst), (ins i64mem:$src), ZeroArgFP,
[(set RFP80:$dst, (X86fild addr:$src, i64))]>;
def ST_Fp32m : FpI<(outs), (ins f32mem:$op, RFP32:$src), OneArgFP,
@@ -367,9 +371,9 @@
[(truncstoref32 RFP64:$src, addr:$op)]>;
def ST_Fp64m : FpI<(outs), (ins f64mem:$op, RFP64:$src), OneArgFP,
[(store RFP64:$src, addr:$op)]>;
-def ST_Fp80m32 : FpI<(outs), (ins f32mem:$op, RFP80:$src), OneArgFP,
+def ST_Fp80m32 : FpI_<(outs), (ins f32mem:$op, RFP80:$src), OneArgFP,
[(truncstoref32 RFP80:$src, addr:$op)]>;
-def ST_Fp80m64 : FpI<(outs), (ins f64mem:$op, RFP80:$src), OneArgFP,
+def ST_Fp80m64 : FpI_<(outs), (ins f64mem:$op, RFP80:$src), OneArgFP,
[(truncstoref64 RFP80:$src, addr:$op)]>;
// FST does not support 80-bit memory target; FSTP must be used.
@@ -378,7 +382,7 @@
def ST_FpP64m : FpI<(outs), (ins f64mem:$op, RFP64:$src), OneArgFP, []>;
def ST_FpP80m32 : FpI<(outs), (ins f32mem:$op, RFP80:$src), OneArgFP, []>;
def ST_FpP80m64 : FpI<(outs), (ins f64mem:$op, RFP80:$src), OneArgFP, []>;
-def ST_FpP80m : FpI<(outs), (ins f80mem:$op, RFP80:$src), OneArgFP,
+def ST_FpP80m : FpI_<(outs), (ins f80mem:$op, RFP80:$src), OneArgFP,
[(store RFP80:$src, addr:$op)]>;
def IST_Fp16m32 : FpI<(outs), (ins i16mem:$op, RFP32:$src), OneArgFP, []>;
def IST_Fp32m32 : FpI<(outs), (ins i32mem:$op, RFP32:$src), OneArgFP, []>;
@@ -386,9 +390,9 @@
def IST_Fp16m64 : FpI<(outs), (ins i16mem:$op, RFP64:$src), OneArgFP, []>;
def IST_Fp32m64 : FpI<(outs), (ins i32mem:$op, RFP64:$src), OneArgFP, []>;
def IST_Fp64m64 : FpI<(outs), (ins i64mem:$op, RFP64:$src), OneArgFP, []>;
-def IST_Fp16m80 : FpI<(outs), (ins i16mem:$op, RFP80:$src), OneArgFP, []>;
-def IST_Fp32m80 : FpI<(outs), (ins i32mem:$op, RFP80:$src), OneArgFP, []>;
-def IST_Fp64m80 : FpI<(outs), (ins i64mem:$op, RFP80:$src), OneArgFP, []>;
+def IST_Fp16m80 : FpI_<(outs), (ins i16mem:$op, RFP80:$src), OneArgFP, []>;
+def IST_Fp32m80 : FpI_<(outs), (ins i32mem:$op, RFP80:$src), OneArgFP, []>;
+def IST_Fp64m80 : FpI_<(outs), (ins i64mem:$op, RFP80:$src), OneArgFP, []>;
def LD_F32m : FPI<0xD9, MRM0m, (outs), (ins f32mem:$src), "fld{s}\t$src">;
def LD_F64m : FPI<0xDD, MRM0m, (outs), (ins f64mem:$src), "fld{l}\t$src">;
@@ -456,9 +460,9 @@
[(set RFP64:$dst, fpimm0)]>;
def LD_Fp164 : FpI<(outs RFP64:$dst), (ins), ZeroArgFP,
[(set RFP64:$dst, fpimm1)]>;
-def LD_Fp080 : FpI<(outs RFP80:$dst), (ins), ZeroArgFP,
+def LD_Fp080 : FpI_<(outs RFP80:$dst), (ins), ZeroArgFP,
[(set RFP80:$dst, fpimm0)]>;
-def LD_Fp180 : FpI<(outs RFP80:$dst), (ins), ZeroArgFP,
+def LD_Fp180 : FpI_<(outs RFP80:$dst), (ins), ZeroArgFP,
[(set RFP80:$dst, fpimm1)]>;
}
@@ -475,9 +479,9 @@
[]>; // FPSW = cmp ST(0) with ST(i)
def UCOM_FpIr64: FpI<(outs), (ins RFP64:$lhs, RFP64:$rhs), CompareFP,
[(X86cmp RFP64:$lhs, RFP64:$rhs)]>; // CC = ST(0) cmp ST(i)
-def UCOM_Fpr80 : FpI<(outs), (ins RFP80:$lhs, RFP80:$rhs), CompareFP,
+def UCOM_Fpr80 : FpI_<(outs), (ins RFP80:$lhs, RFP80:$rhs), CompareFP,
[]>; // FPSW = cmp ST(0) with ST(i)
-def UCOM_FpIr80: FpI<(outs), (ins RFP80:$lhs, RFP80:$rhs), CompareFP,
+def UCOM_FpIr80: FpI_<(outs), (ins RFP80:$lhs, RFP80:$rhs), CompareFP,
[(X86cmp RFP80:$lhs, RFP80:$rhs)]>; // CC = ST(0) cmp ST(i)
def UCOM_Fr : FPI<0xE0, AddRegFrm, // FPSW = cmp ST(0) with ST(i)
@@ -510,12 +514,12 @@
// Non-Instruction Patterns
//===----------------------------------------------------------------------===//
-// Required for RET of f32 / f64 values.
+// Required for RET of f32 / f64 / f80 values.
def : Pat<(X86fld addr:$src, f32), (LD_Fp32m addr:$src)>;
def : Pat<(X86fld addr:$src, f64), (LD_Fp64m addr:$src)>;
def : Pat<(X86fld addr:$src, f80), (LD_Fp80m addr:$src)>;
-// Required for CALL which return f32 / f64 values.
+// Required for CALL which return f32 / f64 / f80 values.
def : Pat<(X86fst RFP32:$src, addr:$op, f32), (ST_Fp32m addr:$op, RFP32:$src)>;
def : Pat<(X86fst RFP64:$src, addr:$op, f32), (ST_Fp64m32 addr:$op, RFP64:$src)>;
def : Pat<(X86fst RFP64:$src, addr:$op, f64), (ST_Fp64m addr:$op, RFP64:$src)>;
@@ -528,19 +532,12 @@
def : Pat<(f32 fpimmneg1), (CHS_Fp32 (LD_Fp132))>, Requires<[FPStack]>;
def : Pat<(f64 fpimmneg0), (CHS_Fp64 (LD_Fp064))>, Requires<[FPStack]>;
def : Pat<(f64 fpimmneg1), (CHS_Fp64 (LD_Fp164))>, Requires<[FPStack]>;
-def : Pat<(f80 fpimmneg0), (CHS_Fp80 (LD_Fp080))>, Requires<[FPStack]>;
-def : Pat<(f80 fpimmneg1), (CHS_Fp80 (LD_Fp180))>, Requires<[FPStack]>;
+def : Pat<(f80 fpimmneg0), (CHS_Fp80 (LD_Fp080))>;
+def : Pat<(f80 fpimmneg1), (CHS_Fp80 (LD_Fp180))>;
// Used to conv. i64 to f64 since there isn't a SSE version.
def : Pat<(X86fildflag addr:$src, i64), (ILD_Fp64m64 addr:$src)>;
-def : Pat<(extloadf80f32 addr:$src),
- (MOV_Fp3280 (LD_Fp32m addr:$src))>, Requires<[FPStack]>;
-def : Pat<(extloadf80f64 addr:$src),
- (MOV_Fp6480 (LD_Fp64m addr:$src))>, Requires<[FPStack]>;
-def : Pat<(extloadf64f32 addr:$src),
- (MOV_Fp3264 (LD_Fp32m addr:$src))>, Requires<[FPStack]>;
-
def : Pat<(f64 (fextend RFP32:$src)), (MOV_Fp3264 RFP32:$src)>, Requires<[FPStack]>;
def : Pat<(f80 (fextend RFP32:$src)), (MOV_Fp3280 RFP32:$src)>, Requires<[FPStack]>;
def : Pat<(f80 (fextend RFP64:$src)), (MOV_Fp6480 RFP64:$src)>, Requires<[FPStack]>;