RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-19 11:02:59 +02:00
Code Issues Projects Releases Wiki Activity

Split -enable-finite-only-fp-math to two options:

Browse Source 
-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.

llvm-svn: 108465
This commit is contained in:
Evan Cheng 2010-07-15 22:07:12 +00:00
parent e44c00b33b
commit ffbae6ad52
11 changed files with 39 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
include/llvm/Target/TargetOptions.h
View File

@ -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

2
lib/CodeGen/Analysis.cpp
View File

@ -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;

2
lib/CodeGen/SelectionDAG/SelectionDAG.cpp
View File

@ -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.

7
lib/Target/ARM/ARMISelLowering.cpp
View File

@ -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.

2
lib/Target/ARM/AsmPrinter/ARMAsmPrinter.cpp
View File

@ -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

20
lib/Target/TargetMachine.cpp
View File

@ -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<bool, true>
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<bool, true>
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<bool, true>
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() {

15
lib/Target/X86/X86ISelLowering.cpp
View File

@ -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);
}

2
test/CodeGen/ARM/fpcmp-opt.ll
View File

@ -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

2
test/CodeGen/X86/2006-05-22-FPSetEQ.ll
View File

@ -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) {

2
test/CodeGen/X86/fabs.ll
View File

@ -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)

4
test/CodeGen/X86/sse-minmax.ll
View File

@ -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.