[InstCombine] Add an additional nnan constraint for xfrm fcmp + sel => fmax/fmin #117977

rajatbajpai · 2024-11-28T08:11:18Z

Add an additional nnan constraint on the fcmp instruction for the fcmp + select => fmax/fmin xfrm. Without the nnan constraint, this transformation is incorrect.

Alive2 with and without the fix: https://alive2.llvm.org/ce/z/Hpo7WX

llvmbot · 2024-11-28T08:11:52Z

@llvm/pr-subscribers-llvm-transforms

Author: Rajat Bajpai (rajatbajpai)

Changes

Add an additional nnan constraint on the fcmp instruction for the fcmp + select => fmax/fmin xfrm. Without the nnan constraint, this transformation is incorrect.

Alive2 with and without the fix: https://alive2.llvm.org/ce/z/Hpo7WX

Full diff: https://github.com/llvm/llvm-project/pull/117977.diff

6 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp (+4-1)
(modified) llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll (+23-23)
(modified) llvm/test/Transforms/InstCombine/fcmp-select.ll (+2-2)
(modified) llvm/test/Transforms/InstCombine/fneg.ll (+1-1)
(modified) llvm/test/Transforms/InstCombine/minmax-fp.ll (+4-4)
(modified) llvm/test/Transforms/InstCombine/unordered-fcmp-select.ll (+2-2)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index e5525133e5dbb5..46473af3f99dfc 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -3864,9 +3864,12 @@ Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) {
   if (SIFPOp) {
     // TODO: Try to forward-propagate FMF from select arms to the select.
 
+    auto *FCmp = dyn_cast<FCmpInst>(CondVal);
+
     // Canonicalize select of FP values where NaN and -0.0 are not valid as
     // minnum/maxnum intrinsics.
-    if (SIFPOp->hasNoNaNs() && SIFPOp->hasNoSignedZeros()) {
+    if (SIFPOp->hasNoNaNs() && SIFPOp->hasNoSignedZeros() && FCmp &&
+        FCmp->hasNoNaNs()) {
       Value *X, *Y;
       if (match(&SI, m_OrdOrUnordFMax(m_Value(X), m_Value(Y))))
         return replaceInstUsesWith(
diff --git a/llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll b/llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll
index c49aea3e82b562..0019bb1023d565 100644
--- a/llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll
+++ b/llvm/test/Transforms/InstCombine/fcmp-fadd-select.ll
@@ -10,7 +10,7 @@ define float @test_fcmp_ogt_fadd_select_constant(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ogt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ogt float %in, 0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -23,7 +23,7 @@ define float @test_fcmp_ogt_fadd_select_constant_swapped(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ogt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ogt float %in, 0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float 1.000000e+00, float %add
   ret float %sel
@@ -36,7 +36,7 @@ define float @test_fcmp_ogt_fadd_select_neg_constant(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ogt float %in, -0.000000e+00
+  %cmp1 = fcmp nnan ogt float %in, -0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -49,7 +49,7 @@ define float @test_fcmp_ogt_fadd_select_fastmath_preserve(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ogt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ogt float %in, 0.000000e+00
   %add = fadd nnan float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -62,7 +62,7 @@ define <2 x float> @test_fcmp_ogt_fadd_select_constant_vectors(<2 x float> %in)
 ; 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]]
 ;
-  %cmp1 = fcmp ogt <2 x float> %in, <float 0.000000e+00, float 0.000000e+00>
+  %cmp1 = fcmp nnan ogt <2 x float> %in, <float 0.000000e+00, float 0.000000e+00>
   %add = fadd <2 x float> %in, <float 1.000000e+00, float 1.000000e+00>
   %sel = select nnan nsz <2 x i1> %cmp1, <2 x float> %add, <2 x float> <float 1.000000e+00, float 1.000000e+00>
   ret <2 x float> %sel
@@ -78,7 +78,7 @@ define float @test_fcmp_olt_fadd_select_constant(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp olt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan olt float %in, 0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -91,7 +91,7 @@ define float @test_fcmp_olt_fadd_select_constant_swapped(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp olt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan olt float %in, 0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float 1.000000e+00, float %add
   ret float %sel
@@ -104,7 +104,7 @@ define float @test_fcmp_olt_fadd_select_neg_constant(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp olt float %in, -0.000000e+00
+  %cmp1 = fcmp nnan olt float %in, -0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -117,7 +117,7 @@ define float @test_fcmp_olt_fadd_select_fastmath_preserve(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp olt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan olt float %in, 0.000000e+00
   %add = fadd nnan float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -130,7 +130,7 @@ define <2 x float> @test_fcmp_olt_fadd_select_constant_vectors(<2 x float> %in)
 ; 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]]
 ;
-  %cmp1 = fcmp olt <2 x float> %in, <float 0.000000e+00, float 0.000000e+00>
+  %cmp1 = fcmp nnan olt <2 x float> %in, <float 0.000000e+00, float 0.000000e+00>
   %add = fadd <2 x float> %in, <float 1.000000e+00, float 1.000000e+00>
   %sel = select nnan nsz <2 x i1> %cmp1, <2 x float> %add, <2 x float> <float 1.000000e+00, float 1.000000e+00>
   ret <2 x float> %sel
@@ -146,7 +146,7 @@ define float @test_fcmp_oge_fadd_select_constant(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp oge float %in, 0.000000e+00
+  %cmp1 = fcmp nnan oge float %in, 0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -159,7 +159,7 @@ define float @test_fcmp_oge_fadd_select_constant_swapped(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp oge float %in, 0.000000e+00
+  %cmp1 = fcmp nnan oge float %in, 0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float 1.000000e+00, float %add
   ret float %sel
@@ -172,7 +172,7 @@ define float @test_fcmp_oge_fadd_select_neg_constant(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp oge float %in, -0.000000e+00
+  %cmp1 = fcmp nnan oge float %in, -0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -185,7 +185,7 @@ define float @test_fcmp_oge_fadd_select_fastmath_preserve(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp oge float %in, 0.000000e+00
+  %cmp1 = fcmp nnan oge float %in, 0.000000e+00
   %add = fadd nnan float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -198,7 +198,7 @@ define <2 x float> @test_fcmp_oge_fadd_select_constant_vectors(<2 x float> %in)
 ; 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]]
 ;
-  %cmp1 = fcmp oge <2 x float> %in, <float 0.000000e+00, float 0.000000e+00>
+  %cmp1 = fcmp nnan oge <2 x float> %in, <float 0.000000e+00, float 0.000000e+00>
   %add = fadd <2 x float> %in, <float 1.000000e+00, float 1.000000e+00>
   %sel = select nnan nsz <2 x i1> %cmp1, <2 x float> %add, <2 x float> <float 1.000000e+00, float 1.000000e+00>
   ret <2 x float> %sel
@@ -214,7 +214,7 @@ define float @test_fcmp_ole_fadd_select_constant(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ole float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ole float %in, 0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -227,7 +227,7 @@ define float @test_fcmp_ole_fadd_select_constant_swapped(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ole float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ole float %in, 0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float 1.000000e+00, float %add
   ret float %sel
@@ -240,7 +240,7 @@ define float @test_fcmp_ole_fadd_select_neg_constant(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ole float %in, -0.000000e+00
+  %cmp1 = fcmp nnan ole float %in, -0.000000e+00
   %add = fadd float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -253,7 +253,7 @@ define float @test_fcmp_ole_fadd_select_fastmath_preserve(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ole float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ole float %in, 0.000000e+00
   %add = fadd nnan float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -266,7 +266,7 @@ define <2 x float> @test_fcmp_ole_fadd_select_constant_vectors(<2 x float> %in)
 ; 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]]
 ;
-  %cmp1 = fcmp ole <2 x float> %in, <float 0.000000e+00, float 0.000000e+00>
+  %cmp1 = fcmp nnan ole <2 x float> %in, <float 0.000000e+00, float 0.000000e+00>
   %add = fadd <2 x float> %in, <float 1.000000e+00, float 1.000000e+00>
   %sel = select nnan nsz <2 x i1> %cmp1, <2 x float> %add, <2 x float> <float 1.000000e+00, float 1.000000e+00>
   ret <2 x float> %sel
@@ -641,7 +641,7 @@ define float @test_fcmp_ogt_fadd_select_rewrite_flags1(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd reassoc nnan nsz arcp contract afn float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ogt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ogt float %in, 0.000000e+00
   %add = fadd reassoc afn arcp contract float %in, 1.000000e+00
   %sel = select nnan nsz reassoc afn arcp contract i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -654,7 +654,7 @@ define float @test_fcmp_ogt_fadd_select_rewrite_flags2(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd nnan nsz float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ogt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ogt float %in, 0.000000e+00
   %add = fadd reassoc float %in, 1.000000e+00
   %sel = select nnan nsz i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
@@ -667,7 +667,7 @@ define float @test_fcmp_ogt_fadd_select_rewrite_and_fastmath(float %in) {
 ; CHECK-NEXT:    [[ADD_NEW:%.*]] = fadd fast float [[SEL_NEW]], 1.000000e+00
 ; CHECK-NEXT:    ret float [[ADD_NEW]]
 ;
-  %cmp1 = fcmp ogt float %in, 0.000000e+00
+  %cmp1 = fcmp nnan ogt float %in, 0.000000e+00
   %add = fadd fast reassoc float %in, 1.000000e+00
   %sel = select fast i1 %cmp1, float %add, float 1.000000e+00
   ret float %sel
diff --git a/llvm/test/Transforms/InstCombine/fcmp-select.ll b/llvm/test/Transforms/InstCombine/fcmp-select.ll
index 028de1ff8a99fa..24f9fab05f1ae8 100644
--- a/llvm/test/Transforms/InstCombine/fcmp-select.ll
+++ b/llvm/test/Transforms/InstCombine/fcmp-select.ll
@@ -223,9 +223,9 @@ define double @test_fcmp_select_maxnum(double %x) {
 ; CHECK-NEXT:    [[SEL2:%.*]] = call nnan nsz double @llvm.minnum.f64(double [[SEL1]], double 2.550000e+02)
 ; CHECK-NEXT:    ret double [[SEL2]]
 ;
-  %cmp1 = fcmp ogt double %x, 1.0
+  %cmp1 = fcmp nnan ogt double %x, 1.0
   %sel1 = select nnan nsz i1 %cmp1, double %x, double 1.0
-  %cmp2 = fcmp olt double %sel1, 255.0
+  %cmp2 = fcmp nnan olt double %sel1, 255.0
   %sel2 = select nnan nsz i1 %cmp2, double %sel1, double 255.0
   ret double %sel2
 }
diff --git a/llvm/test/Transforms/InstCombine/fneg.ll b/llvm/test/Transforms/InstCombine/fneg.ll
index 9692005edf2b6a..c531ee52842708 100644
--- a/llvm/test/Transforms/InstCombine/fneg.ll
+++ b/llvm/test/Transforms/InstCombine/fneg.ll
@@ -1103,7 +1103,7 @@ define float @test_fneg_select_maxnum(float %x) {
 ; CHECK-NEXT:    [[NEG:%.*]] = fneg float [[SEL1]]
 ; CHECK-NEXT:    ret float [[NEG]]
 ;
-  %cmp1 = fcmp ogt float %x, 1.0
+  %cmp1 = fcmp nnan ogt float %x, 1.0
   %sel1 = select nnan nsz i1 %cmp1, float %x, float 1.0
   %neg = fneg float %sel1
   ret float %neg
diff --git a/llvm/test/Transforms/InstCombine/minmax-fp.ll b/llvm/test/Transforms/InstCombine/minmax-fp.ll
index 1276b7b3e3867d..6625458b575b5c 100644
--- a/llvm/test/Transforms/InstCombine/minmax-fp.ll
+++ b/llvm/test/Transforms/InstCombine/minmax-fp.ll
@@ -324,7 +324,7 @@ define float @maxnum_ogt_fmf_on_select(float %a, float %b) {
 ; CHECK-NEXT:    [[F:%.*]] = call nnan nsz float @llvm.maxnum.f32(float [[A:%.*]], float [[B:%.*]])
 ; CHECK-NEXT:    ret float [[F]]
 ;
-  %cond = fcmp ogt float %a, %b
+  %cond = fcmp nnan ogt float %a, %b
   %f = select nnan nsz i1 %cond, float %a, float %b
   ret float %f
 }
@@ -334,7 +334,7 @@ define <2 x float> @maxnum_oge_fmf_on_select(<2 x float> %a, <2 x float> %b) {
 ; CHECK-NEXT:    [[F:%.*]] = call nnan ninf nsz <2 x float> @llvm.maxnum.v2f32(<2 x float> [[A:%.*]], <2 x float> [[B:%.*]])
 ; CHECK-NEXT:    ret <2 x float> [[F]]
 ;
-  %cond = fcmp oge <2 x float> %a, %b
+  %cond = fcmp nnan oge <2 x float> %a, %b
   %f = select ninf nnan nsz <2 x i1> %cond, <2 x float> %a, <2 x float> %b
   ret <2 x float> %f
 }
@@ -388,7 +388,7 @@ define float @minnum_olt_fmf_on_select(float %a, float %b) {
 ; CHECK-NEXT:    [[F:%.*]] = call nnan nsz float @llvm.minnum.f32(float [[A:%.*]], float [[B:%.*]])
 ; CHECK-NEXT:    ret float [[F]]
 ;
-  %cond = fcmp olt float %a, %b
+  %cond = fcmp nnan olt float %a, %b
   %f = select nnan nsz i1 %cond, float %a, float %b
   ret float %f
 }
@@ -398,7 +398,7 @@ define <2 x float> @minnum_ole_fmf_on_select(<2 x float> %a, <2 x float> %b) {
 ; CHECK-NEXT:    [[F:%.*]] = call nnan ninf nsz <2 x float> @llvm.minnum.v2f32(<2 x float> [[A:%.*]], <2 x float> [[B:%.*]])
 ; CHECK-NEXT:    ret <2 x float> [[F]]
 ;
-  %cond = fcmp ole <2 x float> %a, %b
+  %cond = fcmp nnan ole <2 x float> %a, %b
   %f = select ninf nnan nsz <2 x i1> %cond, <2 x float> %a, <2 x float> %b
   ret <2 x float> %f
 }
diff --git a/llvm/test/Transforms/InstCombine/unordered-fcmp-select.ll b/llvm/test/Transforms/InstCombine/unordered-fcmp-select.ll
index b164dd983a8925..9042c3c1fa346d 100644
--- a/llvm/test/Transforms/InstCombine/unordered-fcmp-select.ll
+++ b/llvm/test/Transforms/InstCombine/unordered-fcmp-select.ll
@@ -113,12 +113,12 @@ declare void @foo(i1)
 
 define float @select_max_ugt_2_use_cmp(float %a, float %b) {
 ; CHECK-LABEL: @select_max_ugt_2_use_cmp(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp reassoc ugt float [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp reassoc nnan ugt float [[A:%.*]], [[B:%.*]]
 ; CHECK-NEXT:    call void @foo(i1 [[CMP]])
 ; CHECK-NEXT:    [[SEL:%.*]] = call fast float @llvm.maxnum.f32(float [[A]], float [[B]])
 ; CHECK-NEXT:    ret float [[SEL]]
 ;
-  %cmp = fcmp reassoc ugt float %a, %b
+  %cmp = fcmp nnan reassoc ugt float %a, %b
   call void @foo(i1 %cmp)
   %sel = select fast i1 %cmp, float %a, float %b
   ret float %sel

rajatbajpai · 2024-11-28T08:16:10Z

@arsenm Could you please review this fix?
Original discussion about fixing this issue: #106492 (comment)

nikic

It doesn't look like the nnan is required, you just can't always preserve it? https://alive2.llvm.org/ce/z/4Dj45b

rajatbajpai · 2024-11-28T08:42:10Z

Right now we try to preserve it even if fcmp doesn't have that flag. https://godbolt.org/z/5vPjrWT95

dtcxzyw

https://alive2.llvm.org/ce/z/4PKqnZ

You should drop nnan in the new maxnum/minnum instead of blocking this transform.

dtcxzyw · 2024-11-28T09:18:00Z

https://alive2.llvm.org/ce/z/4PKqnZ

You should drop nnan in the new maxnum/minnum instead of blocking this transform.

Ignore me. @nikic already pointed out this.

Preserve `nnan` constraint only if present on both the instructions: `fcmp` and `select`. Alive2: https://alive2.llvm.org/ce/z/ZNDjzt

rajatbajpai requested a review from nikic as a code owner November 28, 2024 08:11

llvmbot added llvm:instcombine llvm:transforms labels Nov 28, 2024

nikic reviewed Nov 28, 2024

View reviewed changes

nikic added the floating-point Floating-point math label Nov 28, 2024

nikic assigned arsenm, dtcxzyw and jcranmer-intel Nov 28, 2024

dtcxzyw requested changes Nov 28, 2024

View reviewed changes

[InstCombine][FP] Fix nnan preserve for xfrm fcmp + sel => fmax/fmin

36ade16

Preserve `nnan` constraint only if present on both the instructions: `fcmp` and `select`. Alive2: https://alive2.llvm.org/ce/z/ZNDjzt

rajatbajpai force-pushed the dev/rbajpai/upstream-fix-fmax-fmin-fp-xfrm branch from fe41e3e to 36ade16 Compare November 28, 2024 11:58

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[InstCombine] Add an additional nnan constraint for xfrm fcmp + sel => fmax/fmin #117977

[InstCombine] Add an additional nnan constraint for xfrm fcmp + sel => fmax/fmin #117977

rajatbajpai commented Nov 28, 2024

llvmbot commented Nov 28, 2024

rajatbajpai commented Nov 28, 2024

nikic left a comment

rajatbajpai commented Nov 28, 2024

dtcxzyw left a comment

dtcxzyw commented Nov 28, 2024

[InstCombine] Add an additional nnan constraint for xfrm fcmp + sel => fmax/fmin #117977

Are you sure you want to change the base?

[InstCombine] Add an additional nnan constraint for xfrm fcmp + sel => fmax/fmin #117977

Conversation

rajatbajpai commented Nov 28, 2024

llvmbot commented Nov 28, 2024

rajatbajpai commented Nov 28, 2024

nikic left a comment

Choose a reason for hiding this comment

rajatbajpai commented Nov 28, 2024

dtcxzyw left a comment

Choose a reason for hiding this comment

dtcxzyw commented Nov 28, 2024