[llvm] 99dc396 - [InstCombine] Make fptrunc combine use intersection of fast math flags (#118808)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Dec 6 03:40:33 PST 2024
Author: John Brawn
Date: 2024-12-06T11:40:30Z
New Revision: 99dc3967595c472b6edbe789a1346b0350294567
URL: https://github.com/llvm/llvm-project/commit/99dc3967595c472b6edbe789a1346b0350294567
DIFF: https://github.com/llvm/llvm-project/commit/99dc3967595c472b6edbe789a1346b0350294567.diff
LOG: [InstCombine] Make fptrunc combine use intersection of fast math flags (#118808)
These combines involve swapping the fptrunc with its operand, and using
the intersection of fast math flags is the safest option as e.g. if we
have (fptrunc (fneg ninf x)) then (fneg ninf (fptrunc x)) will not be
correct as if x is a not within the range of the destination type the
result of (fptrunc x) will be inf.
Added:
Modified:
llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
llvm/test/Transforms/InstCombine/fpcast.ll
llvm/test/Transforms/InstCombine/fptrunc.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
index d21714b10155d9..102b784169ca7d 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -1847,15 +1847,16 @@ Instruction *InstCombinerImpl::visitFPTrunc(FPTruncInst &FPT) {
Value *X;
Instruction *Op = dyn_cast<Instruction>(FPT.getOperand(0));
if (Op && Op->hasOneUse()) {
- // FIXME: The FMF should propagate from the fptrunc, not the source op.
IRBuilder<>::FastMathFlagGuard FMFG(Builder);
- if (isa<FPMathOperator>(Op))
- Builder.setFastMathFlags(Op->getFastMathFlags());
+ FastMathFlags FMF = FPT.getFastMathFlags();
+ if (auto *FPMO = dyn_cast<FPMathOperator>(Op))
+ FMF &= FPMO->getFastMathFlags();
+ Builder.setFastMathFlags(FMF);
if (match(Op, m_FNeg(m_Value(X)))) {
Value *InnerTrunc = Builder.CreateFPTrunc(X, Ty);
-
- return UnaryOperator::CreateFNegFMF(InnerTrunc, Op);
+ Value *Neg = Builder.CreateFNeg(InnerTrunc);
+ return replaceInstUsesWith(FPT, Neg);
}
// If we are truncating a select that has an extended operand, we can
diff --git a/llvm/test/Transforms/InstCombine/fpcast.ll b/llvm/test/Transforms/InstCombine/fpcast.ll
index 029e513ceafbcd..d5290b572aefd9 100644
--- a/llvm/test/Transforms/InstCombine/fpcast.ll
+++ b/llvm/test/Transforms/InstCombine/fpcast.ll
@@ -29,6 +29,17 @@ define half @test3(float %a) {
ret half %c
}
+define half @test3_fast(float %a) {
+; CHECK-LABEL: @test3_fast(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT: [[C:%.*]] = call half @llvm.fabs.f16(half [[TMP1]])
+; CHECK-NEXT: ret half [[C]]
+;
+ %b = call float @llvm.fabs.f32(float %a)
+ %c = fptrunc fast float %b to half
+ ret half %c
+}
+
define half @fneg_fptrunc(float %a) {
; CHECK-LABEL: @fneg_fptrunc(
; CHECK-NEXT: [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
@@ -78,6 +89,28 @@ define half @test4-fast(float %a) {
; CHECK-NEXT: [[TMP1:%.*]] = fptrunc fast float [[A:%.*]] to half
; CHECK-NEXT: [[C:%.*]] = fneg fast half [[TMP1]]
; CHECK-NEXT: ret half [[C]]
+;
+ %b = fsub fast float -0.0, %a
+ %c = fptrunc fast float %b to half
+ ret half %c
+}
+
+define half @test4-mixed-fast-1(float %a) {
+; CHECK-LABEL: @test4-mixed-fast-1(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT: [[C:%.*]] = fneg half [[TMP1]]
+; CHECK-NEXT: ret half [[C]]
+;
+ %b = fsub float -0.0, %a
+ %c = fptrunc fast float %b to half
+ ret half %c
+}
+
+define half @test4-mixed-fast-2(float %a) {
+; CHECK-LABEL: @test4-mixed-fast-2(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT: [[C:%.*]] = fneg half [[TMP1]]
+; CHECK-NEXT: ret half [[C]]
;
%b = fsub fast float -0.0, %a
%c = fptrunc float %b to half
@@ -89,12 +122,67 @@ define half @test4_unary_fneg-fast(float %a) {
; CHECK-NEXT: [[TMP1:%.*]] = fptrunc fast float [[A:%.*]] to half
; CHECK-NEXT: [[C:%.*]] = fneg fast half [[TMP1]]
; CHECK-NEXT: ret half [[C]]
+;
+ %b = fneg fast float %a
+ %c = fptrunc fast float %b to half
+ ret half %c
+}
+
+define half @test4_unary_fneg-mixed-fast-1(float %a) {
+; CHECK-LABEL: @test4_unary_fneg-mixed-fast-1(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT: [[C:%.*]] = fneg half [[TMP1]]
+; CHECK-NEXT: ret half [[C]]
+;
+ %b = fneg float %a
+ %c = fptrunc fast float %b to half
+ ret half %c
+}
+
+define half @test4_unary_fneg-mixed-fast-2(float %a) {
+; CHECK-LABEL: @test4_unary_fneg-mixed-fast-2(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT: [[C:%.*]] = fneg half [[TMP1]]
+; CHECK-NEXT: ret half [[C]]
;
%b = fneg fast float %a
%c = fptrunc float %b to half
ret half %c
}
+define <2 x half> @test4_unary_fneg-vec-fast(<2 x float> %a) {
+; CHECK-LABEL: @test4_unary_fneg-vec-fast(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc fast <2 x float> [[A:%.*]] to <2 x half>
+; CHECK-NEXT: [[C:%.*]] = fneg fast <2 x half> [[TMP1]]
+; CHECK-NEXT: ret <2 x half> [[C]]
+;
+ %b = fneg fast <2 x float> %a
+ %c = fptrunc fast <2 x float> %b to <2 x half>
+ ret <2 x half> %c
+}
+
+define <2 x half> @test4_unary_fneg-vec-mixed-fast-1(<2 x float> %a) {
+; CHECK-LABEL: @test4_unary_fneg-vec-mixed-fast-1(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc <2 x float> [[A:%.*]] to <2 x half>
+; CHECK-NEXT: [[C:%.*]] = fneg <2 x half> [[TMP1]]
+; CHECK-NEXT: ret <2 x half> [[C]]
+;
+ %b = fneg <2 x float> %a
+ %c = fptrunc fast <2 x float> %b to <2 x half>
+ ret <2 x half> %c
+}
+
+define <2 x half> @test4_unary_fneg-vec-mixed-fast-2(<2 x float> %a) {
+; CHECK-LABEL: @test4_unary_fneg-vec-mixed-fast-2(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc <2 x float> [[A:%.*]] to <2 x half>
+; CHECK-NEXT: [[C:%.*]] = fneg <2 x half> [[TMP1]]
+; CHECK-NEXT: ret <2 x half> [[C]]
+;
+ %b = fneg fast <2 x float> %a
+ %c = fptrunc <2 x float> %b to <2 x half>
+ ret <2 x half> %c
+}
+
define half @test5(float %a, float %b, float %c) {
; CHECK-LABEL: @test5(
; CHECK-NEXT: [[D:%.*]] = fcmp ogt float [[A:%.*]], [[B:%.*]]
diff --git a/llvm/test/Transforms/InstCombine/fptrunc.ll b/llvm/test/Transforms/InstCombine/fptrunc.ll
index a4296a326c4bc6..0b5d8b3cd06e07 100644
--- a/llvm/test/Transforms/InstCombine/fptrunc.ll
+++ b/llvm/test/Transforms/InstCombine/fptrunc.ll
@@ -61,6 +61,18 @@ define float @fptrunc_select_true_val(float %x, double %y, i1 %cond) {
ret float %r
}
+define float @fptrunc_fast_select_true_val(float %x, double %y, i1 %cond) {
+; CHECK-LABEL: @fptrunc_fast_select_true_val(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc fast double [[Y:%.*]] to float
+; CHECK-NEXT: [[NARROW_SEL:%.*]] = select i1 [[COND:%.*]], float [[TMP1]], float [[X:%.*]]
+; CHECK-NEXT: ret float [[NARROW_SEL]]
+;
+ %e = fpext float %x to double
+ %sel = select fast i1 %cond, double %y, double %e
+ %r = fptrunc fast double %sel to float
+ ret float %r
+}
+
define <2 x float> @fptrunc_select_false_val(<2 x float> %x, <2 x double> %y, <2 x i1> %cond) {
; CHECK-LABEL: @fptrunc_select_false_val(
; CHECK-NEXT: [[TMP1:%.*]] = fptrunc <2 x double> [[Y:%.*]] to <2 x float>
@@ -73,6 +85,18 @@ define <2 x float> @fptrunc_select_false_val(<2 x float> %x, <2 x double> %y, <2
ret <2 x float> %r
}
+define <2 x float> @fptrunc_nnan_select_false_val(<2 x float> %x, <2 x double> %y, <2 x i1> %cond) {
+; CHECK-LABEL: @fptrunc_nnan_select_false_val(
+; CHECK-NEXT: [[TMP1:%.*]] = fptrunc nnan <2 x double> [[Y:%.*]] to <2 x float>
+; CHECK-NEXT: [[NARROW_SEL:%.*]] = select <2 x i1> [[COND:%.*]], <2 x float> [[X:%.*]], <2 x float> [[TMP1]]
+; CHECK-NEXT: ret <2 x float> [[NARROW_SEL]]
+;
+ %e = fpext <2 x float> %x to <2 x double>
+ %sel = select nnan <2 x i1> %cond, <2 x double> %e, <2 x double> %y
+ %r = fptrunc nnan <2 x double> %sel to <2 x float>
+ ret <2 x float> %r
+}
+
declare void @use(float)
define half @fptrunc_select_true_val_extra_use(half %x, float %y, i1 %cond) {
More information about the llvm-commits
mailing list