Skip to content

[InstCombine] Propagate FMF from select instead of fcmp #141010

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

Jump to bottom
Merged
merged 4 commits into from
May 31, 2025
Merged

[InstCombine] Propagate FMF from select instead of fcmp #141010

merged 4 commits into from
May 31, 2025

Conversation

dtcxzyw
Copy link
Member

@dtcxzyw dtcxzyw commented May 22, 2025 •
edited
Loading

Previously, 3d6b539 propagates FMF from fcmp to avoid performance regressions. With the help of #139861, #141015, and #141914, we can still convert SPF into fabs/minnum/maxnum intrinsics even if some flags are missing. This patch propagates FMF from select to address the long-standing issue.

Closes #140994.

This was referenced May 22, 2025
Open
Closed
Merged
dtcxzyw added a commit that referenced this pull request May 28, 2025
@dtcxzyw
[ValueTracking][InstCombine] Generalize ignoreSignBitOfZero/NaN to ha…
6c86b7d 
…ndle more cases (#141015)

This patch was originally part of
#139861. It generalizes
`ignoreSignBitOfZero/NaN` to handle more instructions/intrinsics.

BTW, I find it mitigates performance regressions caused by
#141010 (IR diff
https://github.com/dtcxzyw/llvm-opt-benchmark/pull/2365/files). We don't
need to propagate FMF from fcmp into select, since we can infer demanded
properties from the user of select.
llvm-sync bot pushed a commit to arm/arm-toolchain that referenced this pull request May 28, 2025
@dtcxzyw @github-actions
Automerge: [ValueTracking][InstCombine] Generalize ignoreSignBitOfZer…
1345445 
…o/NaN to handle more cases (#141015)

This patch was originally part of
llvm/llvm-project#139861. It generalizes
`ignoreSignBitOfZero/NaN` to handle more instructions/intrinsics.

BTW, I find it mitigates performance regressions caused by
llvm/llvm-project#141010 (IR diff
https://github.com/dtcxzyw/llvm-opt-benchmark/pull/2365/files). We don't
need to propagate FMF from fcmp into select, since we can infer demanded
properties from the user of select.
@dtcxzyw dtcxzyw force-pushed the perf/select-fmf-canon branch from 2a76807 to 53f1175 Compare May 29, 2025 08:22
@dtcxzyw dtcxzyw mentioned this pull request May 29, 2025
Merged
dtcxzyw added a commit that referenced this pull request May 30, 2025
@dtcxzyw
[InstCombine] Use canIgnoreSignBitOfZero in spf->minmax fold (#14…
87fd352 
…1914)

Alive2: https://alive2.llvm.org/ce/z/dCZBB_
Fix remaining regressions caused by #141010.
llvm-sync bot pushed a commit to arm/arm-toolchain that referenced this pull request May 30, 2025
@dtcxzyw @github-actions
Automerge: [InstCombine] Use canIgnoreSignBitOfZero in `spf->minmax…
8d098e6 
…` fold (#141914)

Alive2: https://alive2.llvm.org/ce/z/dCZBB_
Fix remaining regressions caused by llvm/llvm-project#141010.
@dtcxzyw dtcxzyw force-pushed the perf/select-fmf-canon branch from 53f1175 to 2c09b5c Compare May 30, 2025 09:59
@dtcxzyw dtcxzyw mentioned this pull request May 30, 2025
Closed
@dtcxzyw
Copy link
Member Author

dtcxzyw commented May 30, 2025

All real-world regressions have been addressed: dtcxzyw/llvm-opt-benchmark#2380
I am not sure if we should add nsz to address regressions in tests.

@dtcxzyw dtcxzyw requested a review from arsenm May 30, 2025 11:15
@dtcxzyw dtcxzyw marked this pull request as ready for review May 30, 2025 11:15
@dtcxzyw dtcxzyw requested a review from nikic as a code owner May 30, 2025 11:15
@llvmbot
Copy link
Member

llvmbot commented May 30, 2025

@llvm/pr-subscribers-llvm-transforms

Author: Yingwei Zheng (dtcxzyw)

Changes

Previously, 3d6b539 propagates FMF from fcmp to avoid performance regressions. With the help of #139861, #141015, and #141914, we can still convert SPF into fabs/minnum/maxnum intrinsics even if some flags are missing. This patch propagates FMF from select to address the long-standing issue.

Closes #140994.

Patch is 36.68 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/141010.diff

7 Files Affected:

  • (modified) llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp (+8-2)
  • (modified) llvm/test/Transforms/InstCombine/clamp-to-minmax.ll (+32-21)
  • (modified) llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll (+20-40)
  • (modified) llvm/test/Transforms/InstCombine/minmax-fold.ll (+4-2)
  • (modified) llvm/test/Transforms/InstCombine/minmax-fp.ll (+6-4)
  • (modified) llvm/test/Transforms/InstCombine/unordered-fcmp-select.ll (+10-11)
  • (modified) llvm/test/Transforms/PhaseOrdering/AArch64/predicated-reduction.ll (+12-6)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index a791fc5db6698..cca5705f6b58d 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -25,6 +25,7 @@
 #include "llvm/IR/ConstantRange.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/FMF.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/InstrTypes.h"
 #include "llvm/IR/Instruction.h"
@@ -3879,11 +3880,16 @@ Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) {
       // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X
       if (FCmp->hasOneUse() && FCmpInst::isUnordered(Pred)) {
         FCmpInst::Predicate InvPred = FCmp->getInversePredicate();
-        // FIXME: The FMF should propagate from the select, not the fcmp.
         Value *NewCond = Builder.CreateFCmpFMF(InvPred, Cmp0, Cmp1, FCmp,
                                                FCmp->getName() + ".inv");
+        // Propagate ninf/nnan from fcmp to select.
+        FastMathFlags FMF = SI.getFastMathFlags();
+        if (FCmp->hasNoNaNs())
+          FMF.setNoNaNs(true);
+        if (FCmp->hasNoInfs())
+          FMF.setNoInfs(true);
         Value *NewSel =
-            Builder.CreateSelectFMF(NewCond, FalseVal, TrueVal, FCmp);
+            Builder.CreateSelectFMF(NewCond, FalseVal, TrueVal, FMF);
         return replaceInstUsesWith(SI, NewSel);
       }
     }
diff --git a/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll b/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
index b557c0dbe2629..7f3276608c5da 100644
--- a/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
+++ b/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
@@ -7,7 +7,8 @@ define float @clamp_float_fast_ordered_strict_maxmin(float %x) {
 ; CHECK-LABEL: @clamp_float_fast_ordered_strict_maxmin(
 ; CHECK-NEXT:    [[CMP2:%.*]] = fcmp fast olt float [[X:%.*]], 2.550000e+02
 ; CHECK-NEXT:    [[MIN:%.*]] = select i1 [[CMP2]], float [[X]], float 2.550000e+02
-; CHECK-NEXT:    [[R1:%.*]] = call fast float @llvm.maxnum.f32(float [[MIN]], float 1.000000e+00)
+; CHECK-NEXT:    [[DOTINV:%.*]] = fcmp fast oge float [[MIN]], 1.000000e+00
+; CHECK-NEXT:    [[R1:%.*]] = select nnan ninf i1 [[DOTINV]], float [[MIN]], float 1.000000e+00
 ; CHECK-NEXT:    ret float [[R1]]
 ;
   %cmp2 = fcmp fast olt float %x, 255.0
@@ -22,7 +23,8 @@ define float @clamp_float_fast_ordered_nonstrict_maxmin(float %x) {
 ; CHECK-LABEL: @clamp_float_fast_ordered_nonstrict_maxmin(
 ; CHECK-NEXT:    [[CMP2:%.*]] = fcmp fast olt float [[X:%.*]], 2.550000e+02
 ; CHECK-NEXT:    [[MIN:%.*]] = select i1 [[CMP2]], float [[X]], float 2.550000e+02
-; CHECK-NEXT:    [[R1:%.*]] = call fast float @llvm.maxnum.f32(float [[MIN]], float 1.000000e+00)
+; CHECK-NEXT:    [[DOTINV:%.*]] = fcmp fast oge float [[MIN]], 1.000000e+00
+; CHECK-NEXT:    [[R1:%.*]] = select nnan ninf i1 [[DOTINV]], float [[MIN]], float 1.000000e+00
 ; CHECK-NEXT:    ret float [[R1]]
 ;
   %cmp2 = fcmp fast olt float %x, 255.0
@@ -37,7 +39,8 @@ define float @clamp_float_fast_ordered_strict_minmax(float %x) {
 ; CHECK-LABEL: @clamp_float_fast_ordered_strict_minmax(
 ; CHECK-NEXT:    [[CMP2:%.*]] = fcmp fast ogt float [[X:%.*]], 1.000000e+00
 ; CHECK-NEXT:    [[MAX:%.*]] = select i1 [[CMP2]], float [[X]], float 1.000000e+00
-; CHECK-NEXT:    [[R1:%.*]] = call fast float @llvm.minnum.f32(float [[MAX]], float 2.550000e+02)
+; CHECK-NEXT:    [[DOTINV:%.*]] = fcmp fast ole float [[MAX]], 2.550000e+02
+; CHECK-NEXT:    [[R1:%.*]] = select nnan ninf i1 [[DOTINV]], float [[MAX]], float 2.550000e+02
 ; CHECK-NEXT:    ret float [[R1]]
 ;
   %cmp2 = fcmp fast ogt float %x, 1.0
@@ -52,7 +55,8 @@ define float @clamp_float_fast_ordered_nonstrict_minmax(float %x) {
 ; CHECK-LABEL: @clamp_float_fast_ordered_nonstrict_minmax(
 ; CHECK-NEXT:    [[CMP2:%.*]] = fcmp fast ogt float [[X:%.*]], 1.000000e+00
 ; CHECK-NEXT:    [[MAX:%.*]] = select i1 [[CMP2]], float [[X]], float 1.000000e+00
-; CHECK-NEXT:    [[R1:%.*]] = call fast float @llvm.minnum.f32(float [[MAX]], float 2.550000e+02)
+; CHECK-NEXT:    [[DOTINV:%.*]] = fcmp fast ole float [[MAX]], 2.550000e+02
+; CHECK-NEXT:    [[R1:%.*]] = select nnan ninf i1 [[DOTINV]], float [[MAX]], float 2.550000e+02
 ; CHECK-NEXT:    ret float [[R1]]
 ;
   %cmp2 = fcmp fast ogt float %x, 1.0
@@ -68,9 +72,10 @@ define float @clamp_float_fast_ordered_nonstrict_minmax(float %x) {
 ; (X < C1) ? C1 : MIN(X, C2)
 define float @clamp_float_fast_unordered_strict_maxmin(float %x) {
 ; CHECK-LABEL: @clamp_float_fast_unordered_strict_maxmin(
-; CHECK-NEXT:    [[MIN:%.*]] = call fast float @llvm.minnum.f32(float [[X:%.*]], float 2.550000e+02)
-; CHECK-NEXT:    [[CMP1:%.*]] = fcmp fast ult float [[X]], 1.000000e+00
-; CHECK-NEXT:    [[R:%.*]] = select i1 [[CMP1]], float 1.000000e+00, float [[MIN]]
+; CHECK-NEXT:    [[CMP2_INV:%.*]] = fcmp fast oge float [[X:%.*]], 2.550000e+02
+; CHECK-NEXT:    [[MIN:%.*]] = select nnan ninf i1 [[CMP2_INV]], float 2.550000e+02, float [[X]]
+; CHECK-NEXT:    [[DOTINV:%.*]] = fcmp fast oge float [[MIN]], 1.000000e+00
+; CHECK-NEXT:    [[R:%.*]] = select nnan ninf i1 [[DOTINV]], float [[MIN]], float 1.000000e+00
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %cmp2 = fcmp fast ult float %x, 255.0
@@ -83,9 +88,10 @@ define float @clamp_float_fast_unordered_strict_maxmin(float %x) {
 ; (X <= C1) ? C1 : MIN(X, C2)
 define float @clamp_float_fast_unordered_nonstrict_maxmin(float %x) {
 ; CHECK-LABEL: @clamp_float_fast_unordered_nonstrict_maxmin(
-; CHECK-NEXT:    [[MIN:%.*]] = call fast float @llvm.minnum.f32(float [[X:%.*]], float 2.550000e+02)
-; CHECK-NEXT:    [[CMP1:%.*]] = fcmp fast ule float [[X]], 1.000000e+00
-; CHECK-NEXT:    [[R:%.*]] = select i1 [[CMP1]], float 1.000000e+00, float [[MIN]]
+; CHECK-NEXT:    [[CMP2_INV:%.*]] = fcmp fast oge float [[X:%.*]], 2.550000e+02
+; CHECK-NEXT:    [[MIN:%.*]] = select nnan ninf i1 [[CMP2_INV]], float 2.550000e+02, float [[X]]
+; CHECK-NEXT:    [[DOTINV:%.*]] = fcmp fast oge float [[MIN]], 1.000000e+00
+; CHECK-NEXT:    [[R:%.*]] = select nnan ninf i1 [[DOTINV]], float [[MIN]], float 1.000000e+00
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %cmp2 = fcmp fast ult float %x, 255.0
@@ -98,9 +104,10 @@ define float @clamp_float_fast_unordered_nonstrict_maxmin(float %x) {
 ; (X > C1) ? C1 : MAX(X, C2)
 define float @clamp_float_fast_unordered_strict_minmax(float %x) {
 ; CHECK-LABEL: @clamp_float_fast_unordered_strict_minmax(
-; CHECK-NEXT:    [[MAX:%.*]] = call fast float @llvm.maxnum.f32(float [[X:%.*]], float 1.000000e+00)
-; CHECK-NEXT:    [[CMP1:%.*]] = fcmp fast ugt float [[X]], 2.550000e+02
-; CHECK-NEXT:    [[R:%.*]] = select i1 [[CMP1]], float 2.550000e+02, float [[MAX]]
+; CHECK-NEXT:    [[CMP2_INV:%.*]] = fcmp fast ole float [[X:%.*]], 1.000000e+00
+; CHECK-NEXT:    [[MAX:%.*]] = select nnan ninf i1 [[CMP2_INV]], float 1.000000e+00, float [[X]]
+; CHECK-NEXT:    [[DOTINV:%.*]] = fcmp fast ole float [[MAX]], 2.550000e+02
+; CHECK-NEXT:    [[R:%.*]] = select nnan ninf i1 [[DOTINV]], float [[MAX]], float 2.550000e+02
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %cmp2 = fcmp fast ugt float %x, 1.0
@@ -113,9 +120,10 @@ define float @clamp_float_fast_unordered_strict_minmax(float %x) {
 ; (X >= C1) ? C1 : MAX(X, C2)
 define float @clamp_float_fast_unordered_nonstrict_minmax(float %x) {
 ; CHECK-LABEL: @clamp_float_fast_unordered_nonstrict_minmax(
-; CHECK-NEXT:    [[MAX:%.*]] = call fast float @llvm.maxnum.f32(float [[X:%.*]], float 1.000000e+00)
-; CHECK-NEXT:    [[CMP1:%.*]] = fcmp fast uge float [[X]], 2.550000e+02
-; CHECK-NEXT:    [[R:%.*]] = select i1 [[CMP1]], float 2.550000e+02, float [[MAX]]
+; CHECK-NEXT:    [[CMP2_INV:%.*]] = fcmp fast ole float [[X:%.*]], 1.000000e+00
+; CHECK-NEXT:    [[MAX:%.*]] = select nnan ninf i1 [[CMP2_INV]], float 1.000000e+00, float [[X]]
+; CHECK-NEXT:    [[DOTINV:%.*]] = fcmp fast ole float [[MAX]], 2.550000e+02
+; CHECK-NEXT:    [[R:%.*]] = select nnan ninf i1 [[DOTINV]], float [[MAX]], float 2.550000e+02
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %cmp2 = fcmp fast ugt float %x, 1.0
@@ -130,9 +138,10 @@ define float @clamp_float_fast_unordered_nonstrict_minmax(float %x) {
 ; (X > 1.0) ? min(x, 255.0) : 1.0
 define float @clamp_test_1(float %x) {
 ; CHECK-LABEL: @clamp_test_1(
-; CHECK-NEXT:    [[INNER_SEL:%.*]] = call fast float @llvm.minnum.f32(float [[X:%.*]], float 2.550000e+02)
-; CHECK-NEXT:    [[OUTER_CMP:%.*]] = fcmp fast ugt float [[X]], 1.000000e+00
-; CHECK-NEXT:    [[R:%.*]] = select i1 [[OUTER_CMP]], float [[INNER_SEL]], float 1.000000e+00
+; CHECK-NEXT:    [[INNER_CMP_INV:%.*]] = fcmp fast oge float [[X:%.*]], 2.550000e+02
+; CHECK-NEXT:    [[INNER_SEL:%.*]] = select nnan ninf i1 [[INNER_CMP_INV]], float 2.550000e+02, float [[X]]
+; CHECK-NEXT:    [[OUTER_CMP:%.*]] = fcmp fast oge float [[INNER_SEL]], 1.000000e+00
+; CHECK-NEXT:    [[R:%.*]] = select nnan ninf i1 [[OUTER_CMP]], float [[INNER_SEL]], float 1.000000e+00
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %inner_cmp = fcmp fast ult float %x, 255.0
@@ -147,7 +156,8 @@ define float @clamp_test_1(float %x) {
 ; Like @clamp_test_1 but HighConst < LowConst
 define float @clamp_negative_wrong_const(float %x) {
 ; CHECK-LABEL: @clamp_negative_wrong_const(
-; CHECK-NEXT:    [[INNER_SEL:%.*]] = call fast float @llvm.minnum.f32(float [[X:%.*]], float 2.550000e+02)
+; CHECK-NEXT:    [[INNER_CMP_INV:%.*]] = fcmp fast oge float [[X:%.*]], 2.550000e+02
+; CHECK-NEXT:    [[INNER_SEL:%.*]] = select nnan ninf i1 [[INNER_CMP_INV]], float 2.550000e+02, float [[X]]
 ; CHECK-NEXT:    [[OUTER_CMP:%.*]] = fcmp fast ugt float [[X]], 5.120000e+02
 ; CHECK-NEXT:    [[R:%.*]] = select i1 [[OUTER_CMP]], float [[INNER_SEL]], float 5.120000e+02
 ; CHECK-NEXT:    ret float [[R]]
@@ -162,7 +172,8 @@ define float @clamp_negative_wrong_const(float %x) {
 ; Like @clamp_test_1 but both are min
 define float @clamp_negative_same_op(float %x) {
 ; CHECK-LABEL: @clamp_negative_same_op(
-; CHECK-NEXT:    [[INNER_SEL:%.*]] = call fast float @llvm.minnum.f32(float [[X:%.*]], float 2.550000e+02)
+; CHECK-NEXT:    [[INNER_CMP_INV:%.*]] = fcmp fast oge float [[X:%.*]], 2.550000e+02
+; CHECK-NEXT:    [[INNER_SEL:%.*]] = select nnan ninf i1 [[INNER_CMP_INV]], float 2.550000e+02, float [[X]]
 ; CHECK-NEXT:    [[OUTER_CMP:%.*]] = fcmp fast ult float [[X]], 1.000000e+00
 ; CHECK-NEXT:    [[R:%.*]] = select i1 [[OUTER_CMP]], float [[INNER_SEL]], float 1.000000e+00
 ; CHECK-NEXT:    ret float [[R]]
diff --git a/llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll b/llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll
index e05ef6df1d41b..ffc8183d07b2d 100644
--- a/llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll
+++ b/llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll
@@ -278,8 +278,7 @@ define <2 x float> @test_fcmp_ole_fadd_select_constant_vectors(<2 x float> %in)
 define float @test_fcmp_ugt_fadd_select_constant(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_ugt_fadd_select_constant(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp ole float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.maxnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -292,8 +291,7 @@ define float @test_fcmp_ugt_fadd_select_constant(float %in) {
 define float @test_fcmp_ugt_fadd_select_constant_swapped(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_ugt_fadd_select_constant_swapped(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp ole float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float [[IN]], float 0.000000e+00
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.minnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -306,8 +304,7 @@ define float @test_fcmp_ugt_fadd_select_constant_swapped(float %in) {
 define float @test_fcmp_ugt_fadd_select_neg_constant(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_ugt_fadd_select_neg_constant(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp ole float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.maxnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -320,8 +317,7 @@ define float @test_fcmp_ugt_fadd_select_neg_constant(float %in) {
 define float @test_fcmp_ugt_fadd_select_fastmath_preserve(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_ugt_fadd_select_fastmath_preserve(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp ole float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.maxnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -334,8 +330,7 @@ define float @test_fcmp_ugt_fadd_select_fastmath_preserve(float %in) {
 define <2 x float> @test_fcmp_ugt_fadd_select_constant_vectors(<2 x float> %in) {
 ; CHECK-LABEL: define <2 x float> @test_fcmp_ugt_fadd_select_constant_vectors(
 ; CHECK-SAME: <2 x float> [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp ole <2 x float> [[IN]], zeroinitializer
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select <2 x i1> [[CMP1_INV]], <2 x float> zeroinitializer, <2 x float> [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz <2 x float> @llvm.maxnum.v2f32(<2 x float> [[IN]], <2 x float> zeroinitializer)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz <2 x float> [[SEL_NEW]], splat (float 1.000000e+00)
 ; CHECK-NEXT:    ret <2 x float> [[ADD_NEW]]
 ;
@@ -351,8 +346,7 @@ define <2 x float> @test_fcmp_ugt_fadd_select_constant_vectors(<2 x float> %in)
 define float @test_fcmp_uge_fadd_select_constant(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_uge_fadd_select_constant(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp olt float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.maxnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -365,8 +359,7 @@ define float @test_fcmp_uge_fadd_select_constant(float %in) {
 define float @test_fcmp_uge_fadd_select_constant_swapped(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_uge_fadd_select_constant_swapped(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp olt float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float [[IN]], float 0.000000e+00
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.minnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -379,8 +372,7 @@ define float @test_fcmp_uge_fadd_select_constant_swapped(float %in) {
 define float @test_fcmp_uge_fadd_select_neg_constant(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_uge_fadd_select_neg_constant(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp olt float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.maxnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -393,8 +385,7 @@ define float @test_fcmp_uge_fadd_select_neg_constant(float %in) {
 define float @test_fcmp_uge_fadd_select_fastmath_preserve(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_uge_fadd_select_fastmath_preserve(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp olt float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.maxnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -407,8 +398,7 @@ define float @test_fcmp_uge_fadd_select_fastmath_preserve(float %in) {
 define <2 x float> @test_fcmp_uge_fadd_select_constant_vectors(<2 x float> %in) {
 ; CHECK-LABEL: define <2 x float> @test_fcmp_uge_fadd_select_constant_vectors(
 ; CHECK-SAME: <2 x float> [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp olt <2 x float> [[IN]], zeroinitializer
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select <2 x i1> [[CMP1_INV]], <2 x float> zeroinitializer, <2 x float> [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz <2 x float> @llvm.maxnum.v2f32(<2 x float> [[IN]], <2 x float> zeroinitializer)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz <2 x float> [[SEL_NEW]], splat (float 1.000000e+00)
 ; CHECK-NEXT:    ret <2 x float> [[ADD_NEW]]
 ;
@@ -424,8 +414,7 @@ define <2 x float> @test_fcmp_uge_fadd_select_constant_vectors(<2 x float> %in)
 define float @test_fcmp_ult_fadd_select_constant(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_ult_fadd_select_constant(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp oge float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.minnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -438,8 +427,7 @@ define float @test_fcmp_ult_fadd_select_constant(float %in) {
 define float @test_fcmp_ult_fadd_select_constant_swapped(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_ult_fadd_select_constant_swapped(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp oge float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float [[IN]], float 0.000000e+00
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.maxnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -452,8 +440,7 @@ define float @test_fcmp_ult_fadd_select_constant_swapped(float %in) {
 define float @test_fcmp_ult_fadd_select_neg_constant(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_ult_fadd_select_neg_constant(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp oge float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.minnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -466,8 +453,7 @@ define float @test_fcmp_ult_fadd_select_neg_constant(float %in) {
 define float @test_fcmp_ult_fadd_select_fastmath_preserve(float %in) {
 ; CHECK-LABEL: define float @test_fcmp_ult_fadd_select_fastmath_preserve(
 ; CHECK-SAME: float [[IN:%.*]]) {
-; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp oge float [[IN]], 0.000000e+00
-; CHECK-NEXT:    [[SEL_NEW:%.*]] = select i1 [[CMP1_INV]], float 0.000000e+00, float [[IN]]
+; CHECK-NEXT:    [[SEL_NEW:%.*]] = call nsz float @llvm.minnum.f32(float [[IN]], float 0.000000e+00)
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
@@ -480,8 +466,7 @@ define float @test_fcmp_ult_fadd_select_fastmath_preserve(float %in) {
 define <2 x float> @test_fcmp_ult_fadd_select_constant_vectors(<2 x float> %in) {
 ; CHECK-LABEL: define <2 x float> @test_fcmp_ult_fadd_select_constant_vectors(
 ; CHECK-SAME: <2 x f...
[truncated]

@dtcxzyw dtcxzyw merged commit 3ec0c5c into llvm:main May 31, 2025
15 checks passed
@dtcxzyw dtcxzyw deleted the perf/select-fmf-canon branch May 31, 2025 08:25
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@arsenm arsenm arsenm approved these changes

@nikic nikic Awaiting requested review from nikic nikic is a code owner

Assignees
No one assigned
Labels
llvm:instcombine llvm:transforms
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

[InstCombine] (fcmp une T, F) ? T : F --> T should not rely on the nsz flag on fcmp
3 participants
@dtcxzyw @llvmbot @arsenm