Skip to content

release/21.x: [IR] Handle fabs LHS in fcmpImpliesClass (#152913) #152921

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Aug 11, 2025
Merged

release/21.x: [IR] Handle fabs LHS in fcmpImpliesClass (#152913) #152921

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 14 additions & 10 deletions llvm/include/llvm/IR/GenericFloatingPointPredicateUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -135,6 +135,12 @@ template <typename ContextT> class GenericFloatingPointPredicateUtils {
if (Mode.Input != DenormalMode::IEEE)
return {Invalid, fcAllFlags, fcAllFlags};

auto ExactClass = [IsFabs, Src](FPClassTest Mask) {
if (IsFabs)
Mask = llvm::inverse_fabs(Mask);
return exactClass(Src, Mask);
};

switch (Pred) {
case FCmpInst::FCMP_OEQ: // Match x == 0.0
return exactClass(Src, fcZero);
Expand All @@ -151,26 +157,24 @@ template <typename ContextT> class GenericFloatingPointPredicateUtils {
case FCmpInst::FCMP_UNO:
return exactClass(Src, fcNan);
case FCmpInst::FCMP_OGT: // x > 0
return exactClass(Src, fcPosSubnormal | fcPosNormal | fcPosInf);
return ExactClass(fcPosSubnormal | fcPosNormal | fcPosInf);
case FCmpInst::FCMP_UGT: // isnan(x) || x > 0
return exactClass(Src, fcPosSubnormal | fcPosNormal | fcPosInf | fcNan);
return ExactClass(fcPosSubnormal | fcPosNormal | fcPosInf | fcNan);
case FCmpInst::FCMP_OGE: // x >= 0
return exactClass(Src, fcPositive | fcNegZero);
return ExactClass(fcPositive | fcNegZero);
case FCmpInst::FCMP_UGE: // isnan(x) || x >= 0
return exactClass(Src, fcPositive | fcNegZero | fcNan);
return ExactClass(fcPositive | fcNegZero | fcNan);
case FCmpInst::FCMP_OLT: // x < 0
return exactClass(Src, fcNegSubnormal | fcNegNormal | fcNegInf);
return ExactClass(fcNegSubnormal | fcNegNormal | fcNegInf);
case FCmpInst::FCMP_ULT: // isnan(x) || x < 0
return exactClass(Src, fcNegSubnormal | fcNegNormal | fcNegInf | fcNan);
return ExactClass(fcNegSubnormal | fcNegNormal | fcNegInf | fcNan);
case FCmpInst::FCMP_OLE: // x <= 0
return exactClass(Src, fcNegative | fcPosZero);
return ExactClass(fcNegative | fcPosZero);
case FCmpInst::FCMP_ULE: // isnan(x) || x <= 0
return exactClass(Src, fcNegative | fcPosZero | fcNan);
return ExactClass(fcNegative | fcPosZero | fcNan);
default:
llvm_unreachable("all compare types are handled");
}

return {Invalid, fcAllFlags, fcAllFlags};
}

const bool IsDenormalRHS = (OrigClass & fcSubnormal) == OrigClass;
Expand Down
21 changes: 19 additions & 2 deletions llvm/test/Transforms/InstSimplify/floating-point-arithmetic.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -213,15 +213,15 @@ define double @fmul_nnan_ninf_nneg_n0.0_commute(i127 %x) {

define float @fmul_ninf_nnan_mul_zero_nsz(float nofpclass(inf nan) %f) {
; CHECK-LABEL: @fmul_ninf_nnan_mul_zero_nsz(
; CHECK-NEXT: ret float 0.000000e+00
; CHECK-NEXT: ret float 0.000000e+00
;
%r = fmul nsz float %f, 0.0
ret float %r
}

define float @fmul_ninf_nnan_mul_nzero_nsz(float nofpclass(inf nan) %f) {
; CHECK-LABEL: @fmul_ninf_nnan_mul_nzero_nsz(
; CHECK-NEXT: ret float 0.000000e+00
; CHECK-NEXT: ret float 0.000000e+00
;
%r = fmul nsz float %f, -0.0
ret float %r
Expand Down Expand Up @@ -1255,3 +1255,20 @@ define i1 @fptrunc_round_unknown_positive(double %unknown) {
%cmp = fcmp nnan oge float %op, 0.0
ret i1 %cmp
}

define half @fabs_select_fabs(half noundef %x) {
; CHECK-LABEL: @fabs_select_fabs(
; CHECK-NEXT: entry:
; CHECK-NEXT: [[ABS1:%.*]] = call half @llvm.fabs.f16(half [[X:%.*]])
; CHECK-NEXT: [[CMP:%.*]] = fcmp ogt half [[ABS1]], 0xH0000
; CHECK-NEXT: [[SEL:%.*]] = select i1 [[CMP]], half [[X]], half 0xH0000
; CHECK-NEXT: [[ABS2:%.*]] = call half @llvm.fabs.f16(half [[SEL]])
; CHECK-NEXT: ret half [[ABS2]]
;
entry:
%abs1 = call half @llvm.fabs.f16(half %x)
%cmp = fcmp ogt half %abs1, 0xH0000
%sel = select i1 %cmp, half %x, half 0xH0000
%abs2 = call half @llvm.fabs.f16(half %sel)
ret half %abs2
}
36 changes: 36 additions & 0 deletions llvm/unittests/Analysis/ValueTrackingTest.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@
//===----------------------------------------------------------------------===//

#include "llvm/Analysis/ValueTracking.h"
#include "llvm/ADT/FloatingPointMode.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/FloatingPointPredicateUtils.h"
#include "llvm/AsmParser/Parser.h"
Expand Down Expand Up @@ -2208,6 +2209,41 @@ TEST_F(ComputeKnownFPClassTest, Constants) {
}
}

TEST_F(ComputeKnownFPClassTest, fcmpImpliesClass_fabs_zero) {
parseAssembly("define float @test(float %x) {\n"
" %A = call float @llvm.fabs.f32(float %x)\n"
" ret float %A\n"
"}\n");
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_OEQ, *F, A, fcZero)),
fcZero);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_UEQ, *F, A, fcZero)),
fcZero | fcNan);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_UNE, *F, A, fcZero)),
~fcZero);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_ONE, *F, A, fcZero)),
~fcNan & ~fcZero);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_ORD, *F, A, fcZero)),
~fcNan);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_UNO, *F, A, fcZero)),
fcNan);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_OGT, *F, A, fcZero)),
fcSubnormal | fcNormal | fcInf);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_UGT, *F, A, fcZero)),
fcSubnormal | fcNormal | fcInf | fcNan);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_OGE, *F, A, fcZero)),
~fcNan);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_UGE, *F, A, fcZero)),
fcAllFlags);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_OLT, *F, A, fcZero)),
fcNone);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_ULT, *F, A, fcZero)),
fcNan);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_OLE, *F, A, fcZero)),
fcZero);
EXPECT_EQ(std::get<1>(fcmpImpliesClass(FCmpInst::FCMP_ULE, *F, A, fcZero)),
fcZero | fcNan);
}

TEST_F(ValueTrackingTest, isNonZeroRecurrence) {
parseAssembly(R"(
define i1 @test(i8 %n, i8 %r) {
Expand Down
Loading