From 60108e96bbc5432f4fe06ba313e64448e97a0e15 Mon Sep 17 00:00:00 2001 From: Evan Cheng Date: Thu, 15 Jul 2010 22:07:12 +0000 Subject: [PATCH] Split -enable-finite-only-fp-math to two options: -enable-no-nans-fp-math and -enable-no-infs-fp-math. All of the current codegen fp math optimizations only care whether the fp arithmetics arguments and results can never be NaN. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@108465 91177308-0d34-0410-b5e6-96231b3b80d8 --- include/llvm/Target/TargetOptions.h | 19 ++++++++++++------- lib/CodeGen/Analysis.cpp | 2 +- lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 2 +- lib/Target/ARM/ARMISelLowering.cpp | 7 ++++--- lib/Target/ARM/AsmPrinter/ARMAsmPrinter.cpp | 2 +- lib/Target/TargetMachine.cpp | 20 ++++++++++---------- lib/Target/X86/X86ISelLowering.cpp | 15 +++++---------- test/CodeGen/ARM/fpcmp-opt.ll | 2 +- test/CodeGen/X86/2006-05-22-FPSetEQ.ll | 2 +- test/CodeGen/X86/fabs.ll | 2 +- test/CodeGen/X86/sse-minmax.ll | 4 ++-- 11 files changed, 39 insertions(+), 38 deletions(-) diff --git a/include/llvm/Target/TargetOptions.h b/include/llvm/Target/TargetOptions.h index b36988097cf..f69778f232f 100644 --- a/include/llvm/Target/TargetOptions.h +++ b/include/llvm/Target/TargetOptions.h @@ -71,13 +71,18 @@ namespace llvm { /// UnsafeFPMath implies LessPreciseFPMAD. extern bool UnsafeFPMath; - /// FiniteOnlyFPMath - This returns true when the -enable-finite-only-fp-math - /// option is specified on the command line. If this returns false (default), - /// the code generator is not allowed to assume that FP arithmetic arguments - /// and results are never NaNs or +-Infs. - extern bool FiniteOnlyFPMathOption; - extern bool FiniteOnlyFPMath(); - + /// NoInfsFPMath - This flag is enabled when the + /// -enable-no-infs-fp-math flag is specified on the command line. When + /// this flag is off (the default), the code generator is not allowed to + /// assume the FP arithmetic arguments and results are never +-Infs. + extern bool NoInfsFPMath; + + /// NoNaNsFPMath - This flag is enabled when the + /// -enable-no-nans-fp-math flag is specified on the command line. When + /// this flag is off (the default), the code generator is not allowed to + /// assume the FP arithmetic arguments and results are never NaNs. + extern bool NoNaNsFPMath; + /// HonorSignDependentRoundingFPMath - This returns true when the /// -enable-sign-dependent-rounding-fp-math is specified. If this returns /// false (the default), the code generator is allowed to assume that the diff --git a/lib/CodeGen/Analysis.cpp b/lib/CodeGen/Analysis.cpp index f71eee5d01b..a20d8107dc4 100644 --- a/lib/CodeGen/Analysis.cpp +++ b/lib/CodeGen/Analysis.cpp @@ -171,7 +171,7 @@ ISD::CondCode llvm::getFCmpCondCode(FCmpInst::Predicate Pred) { FOC = FPC = ISD::SETFALSE; break; } - if (FiniteOnlyFPMath()) + if (NoNaNsFPMath) return FOC; else return FPC; diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp index e83a0346b53..27d7a829d9f 100644 --- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp +++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp @@ -2236,7 +2236,7 @@ unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{ bool SelectionDAG::isKnownNeverNaN(SDValue Op) const { // If we're told that NaNs won't happen, assume they won't. - if (FiniteOnlyFPMath()) + if (NoNaNsFPMath) return true; // If the value is a constant, we can obviously see if it is a NaN or not. diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp index 0091df753eb..444e30e8e1b 100644 --- a/lib/Target/ARM/ARMISelLowering.cpp +++ b/lib/Target/ARM/ARMISelLowering.cpp @@ -2403,8 +2403,9 @@ ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { bool SeenZero = false; if (canChangeToInt(LHS, SeenZero, Subtarget) && canChangeToInt(RHS, SeenZero, Subtarget) && - // If one of the operand is zero, it's safe to ignore the NaN case. - (FiniteOnlyFPMath() || SeenZero)) { + // If one of the operand is zero, it's safe to ignore the NaN case since + // we only care about equality comparisons. + (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) { // If unsafe fp math optimization is enabled and there are no othter uses of // the CMP operands, and the condition code is EQ oe NE, we can optimize it // to an integer comparison. @@ -4561,7 +4562,7 @@ static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG, static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG, const ARMSubtarget *ST) { // If the target supports NEON, try to use vmax/vmin instructions for f32 - // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set, + // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set, // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is // a NaN; only do the transformation when it matches that behavior. diff --git a/lib/Target/ARM/AsmPrinter/ARMAsmPrinter.cpp b/lib/Target/ARM/AsmPrinter/ARMAsmPrinter.cpp index 946f4744f5b..7eb9b37108c 100644 --- a/lib/Target/ARM/AsmPrinter/ARMAsmPrinter.cpp +++ b/lib/Target/ARM/AsmPrinter/ARMAsmPrinter.cpp @@ -1174,7 +1174,7 @@ void ARMAsmPrinter::EmitStartOfAsmFile(Module &M) { Twine(ARMBuildAttrs::ABI_FP_exceptions) + ", 1"); } - if (FiniteOnlyFPMath()) + if (NoInfsFPMath && NoNaNsFPMath) OutStreamer.EmitRawText("\t.eabi_attribute " + Twine(ARMBuildAttrs::ABI_FP_number_model)+ ", 1"); else diff --git a/lib/Target/TargetMachine.cpp b/lib/Target/TargetMachine.cpp index 47c91df1400..705b1c097e5 100644 --- a/lib/Target/TargetMachine.cpp +++ b/lib/Target/TargetMachine.cpp @@ -30,7 +30,8 @@ namespace llvm { bool NoFramePointerElimNonLeaf; bool NoExcessFPPrecision; bool UnsafeFPMath; - bool FiniteOnlyFPMathOption; + bool NoInfsFPMath; + bool NoNaNsFPMath; bool HonorSignDependentRoundingFPMathOption; bool UseSoftFloat; FloatABI::ABIType FloatABIType; @@ -80,9 +81,14 @@ EnableUnsafeFPMath("enable-unsafe-fp-math", cl::location(UnsafeFPMath), cl::init(false)); static cl::opt -EnableFiniteOnlyFPMath("enable-finite-only-fp-math", - cl::desc("Enable optimizations that assumes non- NaNs / +-Infs"), - cl::location(FiniteOnlyFPMathOption), +EnableNoInfsFPMath("enable-no-infs-fp-math", + cl::desc("Enable FP math optimizations that assume no +-Infs"), + cl::location(NoInfsFPMath), + cl::init(false)); +static cl::opt +EnableNoNaNsFPMath("enable-no-nans-fp-math", + cl::desc("Enable FP math optimizations that assume no NaNs"), + cl::location(NoNaNsFPMath), cl::init(false)); static cl::opt EnableHonorSignDependentRoundingFPMath("enable-sign-dependent-rounding-fp-math", @@ -290,12 +296,6 @@ namespace llvm { /// result is "less precise" than doing those operations individually. bool LessPreciseFPMAD() { return UnsafeFPMath || LessPreciseFPMADOption; } - /// FiniteOnlyFPMath - This returns true when the -enable-finite-only-fp-math - /// option is specified on the command line. If this returns false (default), - /// the code generator is not allowed to assume that FP arithmetic arguments - /// and results are never NaNs or +-Infs. - bool FiniteOnlyFPMath() { return FiniteOnlyFPMathOption; } - /// HonorSignDependentRoundingFPMath - Return true if the codegen must assume /// that the rounding mode of the FPU can change from its default. bool HonorSignDependentRoundingFPMath() { diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp index b3c48862898..d10f4ec2f92 100644 --- a/lib/Target/X86/X86ISelLowering.cpp +++ b/lib/Target/X86/X86ISelLowering.cpp @@ -9087,8 +9087,7 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, // Converting this to a min would handle NaNs incorrectly, and swapping // the operands would cause it to handle comparisons between positive // and negative zero incorrectly. - if (!FiniteOnlyFPMath() && - (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) { + if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) { if (!UnsafeFPMath && !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) break; @@ -9126,8 +9125,7 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, // Converting this to a max would handle NaNs incorrectly, and swapping // the operands would cause it to handle comparisons between positive // and negative zero incorrectly. - if (!FiniteOnlyFPMath() && - (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) { + if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) { if (!UnsafeFPMath && !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) break; @@ -9156,8 +9154,7 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, // cause it to handle NaNs incorrectly. if (!UnsafeFPMath && !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) { - if (!FiniteOnlyFPMath() && - (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) + if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) break; std::swap(LHS, RHS); } @@ -9182,8 +9179,7 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, case ISD::SETULT: // Converting this to a max would handle NaNs incorrectly. - if (!FiniteOnlyFPMath() && - (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) + if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) break; Opcode = X86ISD::FMAX; break; @@ -9193,8 +9189,7 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, // cause it to handle NaNs incorrectly. if (!UnsafeFPMath && !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) { - if (!FiniteOnlyFPMath() && - (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) + if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) break; std::swap(LHS, RHS); } diff --git a/test/CodeGen/ARM/fpcmp-opt.ll b/test/CodeGen/ARM/fpcmp-opt.ll index 6875288304b..64350591b87 100644 --- a/test/CodeGen/ARM/fpcmp-opt.ll +++ b/test/CodeGen/ARM/fpcmp-opt.ll @@ -1,4 +1,4 @@ -; RUN: llc < %s -march=arm -mcpu=cortex-a8 -mattr=+vfp2 -enable-unsafe-fp-math -enable-finite-only-fp-math | FileCheck -check-prefix=FINITE %s +; RUN: llc < %s -march=arm -mcpu=cortex-a8 -mattr=+vfp2 -enable-unsafe-fp-math -enable-no-nans-fp-math | FileCheck -check-prefix=FINITE %s ; RUN: llc < %s -march=arm -mcpu=cortex-a8 -mattr=+vfp2 -enable-unsafe-fp-math | FileCheck -check-prefix=NAN %s ; rdar://7461510 diff --git a/test/CodeGen/X86/2006-05-22-FPSetEQ.ll b/test/CodeGen/X86/2006-05-22-FPSetEQ.ll index 2d7bd27d24b..35b0159d39c 100644 --- a/test/CodeGen/X86/2006-05-22-FPSetEQ.ll +++ b/test/CodeGen/X86/2006-05-22-FPSetEQ.ll @@ -1,5 +1,5 @@ ; RUN: llc < %s -march=x86 | grep setnp -; RUN: llc < %s -march=x86 -enable-unsafe-fp-math -enable-finite-only-fp-math | \ +; RUN: llc < %s -march=x86 -enable-unsafe-fp-math -enable-no-nans-fp-math | \ ; RUN: not grep setnp define i32 @test(float %f) { diff --git a/test/CodeGen/X86/fabs.ll b/test/CodeGen/X86/fabs.ll index 23b45ebb8d8..9ded7e05dc4 100644 --- a/test/CodeGen/X86/fabs.ll +++ b/test/CodeGen/X86/fabs.ll @@ -1,7 +1,7 @@ ; Make sure this testcase codegens to the fabs instruction, not a call to fabsf ; RUN: llc < %s -march=x86 -mattr=-sse2,-sse3,-sse | grep fabs\$ | \ ; RUN: count 2 -; RUN: llc < %s -march=x86 -mattr=-sse,-sse2,-sse3 -enable-unsafe-fp-math -enable-finite-only-fp-math | \ +; RUN: llc < %s -march=x86 -mattr=-sse,-sse2,-sse3 -enable-unsafe-fp-math -enable-no-nans-fp-math | \ ; RUN: grep fabs\$ | count 3 declare float @fabsf(float) diff --git a/test/CodeGen/X86/sse-minmax.ll b/test/CodeGen/X86/sse-minmax.ll index ebcdc655eed..348121ac8bc 100644 --- a/test/CodeGen/X86/sse-minmax.ll +++ b/test/CodeGen/X86/sse-minmax.ll @@ -1,6 +1,6 @@ ; RUN: llc < %s -march=x86-64 -asm-verbose=false | FileCheck %s -; RUN: llc < %s -march=x86-64 -asm-verbose=false -enable-unsafe-fp-math -enable-finite-only-fp-math | FileCheck -check-prefix=UNSAFE %s -; RUN: llc < %s -march=x86-64 -asm-verbose=false -enable-finite-only-fp-math | FileCheck -check-prefix=FINITE %s +; RUN: llc < %s -march=x86-64 -asm-verbose=false -enable-unsafe-fp-math -enable-no-nans-fp-math | FileCheck -check-prefix=UNSAFE %s +; RUN: llc < %s -march=x86-64 -asm-verbose=false -enable-no-nans-fp-math | FileCheck -check-prefix=FINITE %s ; Some of these patterns can be matched as SSE min or max. Some of ; then can be matched provided that the operands are swapped.