https://github.com/yxsamliu created 
https://github.com/llvm/llvm-project/pull/244342

When an ninf select chooses the result of an inner select, an
infinity from the inner select makes the outer result poison.
If the inner select has no other uses, give it ninf so InstSimplify
can fold a constant infinity arm. This also propagates ninf through
nested selects in AMDGPU libcalls.

>From 3bc88d5dc6e43fb11e3a49bd1fa32a531b9a0664 Mon Sep 17 00:00:00 2001
From: "Yaxun (Sam) Liu" <[email protected]>
Date: Thu, 19 Nov 2026 18:42:58 +0000
Subject: [PATCH] [InstCombine] Propagate ninf through single-use scalar
 selects

When an ninf select chooses the result of an inner select, an
infinity from the inner select makes the outer result poison.
If the inner select has no other uses, give it ninf so InstSimplify
can fold a constant infinity arm. This also propagates ninf through
nested selects in AMDGPU libcalls.

AI assisted.
---
 .../InstCombine/InstCombineSelect.cpp         | 16 ++++++++
 .../AMDGPU/amdgpu-simplify-libcall-pow.ll     | 14 +++----
 .../AMDGPU/amdgpu-simplify-libcall-rootn.ll   |  6 +--
 llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll |  4 +-
 .../Transforms/InstCombine/select-ninf.ll     | 39 +++++++++++++++++++
 5 files changed, 67 insertions(+), 12 deletions(-)
 create mode 100644 llvm/test/Transforms/InstCombine/select-ninf.ll

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp 
b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index d7bb8de19c6da7..85e87a58ae6d98 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -4742,6 +4742,22 @@ Instruction 
*InstCombinerImpl::visitSelectInst(SelectInst &SI) {
                                     SQ.getWithInstruction(&SI)))
     return replaceInstUsesWith(SI, V);
 
+  if (FMF.noInfs() && SelType->isFloatingPointTy()) {
+    // The constraint applies to the selected result. A nested select can
+    // inherit it when its only use is as an arm of this select.
+    bool Changed = false;
+    for (Value *Arm : {TrueVal, FalseVal}) {
+      auto *Inner = dyn_cast<SelectInst>(Arm);
+      if (!Inner || !Inner->hasOneUse() || Inner->hasNoInfs())
+        continue;
+      Inner->setHasNoInfs(true);
+      addToWorklist(Inner);
+      Changed = true;
+    }
+    if (Changed)
+      return &SI;
+  }
+
   if (Instruction *I = canonicalizeSelectToShuffle(SI))
     return I;
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll 
b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
index d1435fac8d9c69..de4fbfff0bb3d7 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
@@ -52,7 +52,7 @@ define float @test_pow_fast_f32(float %x, float %y) {
 ; NOPRELINK-NEXT:    [[TMP21:%.*]] = select fast i1 [[TMP20]], float +qnan, 
float [[TMP16]]
 ; NOPRELINK-NEXT:    [[TMP22:%.*]] = fcmp fast oeq float [[TMP4]], 0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP23:%.*]] = call fast float @llvm.copysign.f32(float 
0.000000e+00, float [[TMP4]])
-; NOPRELINK-NEXT:    [[TMP24:%.*]] = select i1 [[TMP14]], float [[TMP23]], 
float 0.000000e+00
+; NOPRELINK-NEXT:    [[TMP24:%.*]] = select ninf i1 [[TMP14]], float 
[[TMP23]], float 0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP25:%.*]] = select fast i1 [[TMP22]], float 
[[TMP24]], float [[TMP21]]
 ; NOPRELINK-NEXT:    ret float [[TMP25]]
 ;
@@ -189,7 +189,7 @@ define float @test_pow_afn_f32_nnan_ninf(float %x, float 
%y) {
 ; NOPRELINK-NEXT:    [[TMP21:%.*]] = select nnan ninf afn i1 [[TMP20]], float 
+qnan, float [[TMP16]]
 ; NOPRELINK-NEXT:    [[TMP22:%.*]] = fcmp nnan ninf afn oeq float [[TMP4]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP23:%.*]] = call nnan ninf afn float 
@llvm.copysign.f32(float 0.000000e+00, float [[TMP4]])
-; NOPRELINK-NEXT:    [[TMP24:%.*]] = select i1 [[TMP14]], float [[TMP23]], 
float 0.000000e+00
+; NOPRELINK-NEXT:    [[TMP24:%.*]] = select ninf i1 [[TMP14]], float 
[[TMP23]], float 0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP25:%.*]] = select nnan ninf afn i1 [[TMP22]], float 
[[TMP24]], float [[TMP21]]
 ; NOPRELINK-NEXT:    ret float [[TMP25]]
 ;
@@ -1383,7 +1383,7 @@ define float @test_pow_afn_f32_0.5(float %x) {
 ; NOPRELINK-NEXT:    [[TMP19:%.*]] = select ninf nsz afn i1 [[TMP18]], float 
+qnan, float [[TMP14]]
 ; NOPRELINK-NEXT:    [[TMP20:%.*]] = fcmp ninf nsz afn oeq float [[X]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP21:%.*]] = call ninf nsz afn float 
@llvm.copysign.f32(float 0.000000e+00, float [[X]])
-; NOPRELINK-NEXT:    [[TMP22:%.*]] = select i1 [[TMP12]], float [[TMP21]], 
float 0.000000e+00
+; NOPRELINK-NEXT:    [[TMP22:%.*]] = select ninf i1 [[TMP12]], float 
[[TMP21]], float 0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP23:%.*]] = select ninf nsz afn i1 [[TMP20]], float 
[[TMP22]], float [[TMP19]]
 ; NOPRELINK-NEXT:    [[TMP24:%.*]] = fcmp ninf nsz afn uno float [[X]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP25:%.*]] = select ninf nsz afn i1 [[TMP24]], float 
+qnan, float [[TMP23]]
@@ -1422,7 +1422,7 @@ define float @test_pow_afn_f32_neg0.5(float %x) {
 ; NOPRELINK-NEXT:    [[TMP19:%.*]] = select ninf nsz afn i1 [[TMP18]], float 
+qnan, float [[TMP14]]
 ; NOPRELINK-NEXT:    [[TMP20:%.*]] = fcmp ninf nsz afn oeq float [[X]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP21:%.*]] = call ninf nsz afn float 
@llvm.copysign.f32(float 0.000000e+00, float [[X]])
-; NOPRELINK-NEXT:    [[TMP22:%.*]] = select i1 [[TMP12]], float [[TMP21]], 
float 0.000000e+00
+; NOPRELINK-NEXT:    [[TMP22:%.*]] = select ninf i1 [[TMP12]], float 
[[TMP21]], float 0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP23:%.*]] = select ninf nsz afn i1 [[TMP20]], float 
[[TMP22]], float [[TMP19]]
 ; NOPRELINK-NEXT:    [[TMP24:%.*]] = fcmp ninf nsz afn uno float [[X]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP25:%.*]] = select ninf nsz afn i1 [[TMP24]], float 
+qnan, float [[TMP23]]
@@ -3250,7 +3250,7 @@ define float 
@test_pow_afn_f32_nnan_ninf_x_known_positive(float nofpclass(ninf n
 ; NOPRELINK-NEXT:    [[TMP15:%.*]] = call nnan ninf afn float 
@llvm.copysign.f32(float [[TMP7]], float [[TMP14]])
 ; NOPRELINK-NEXT:    [[TMP16:%.*]] = fcmp nnan ninf afn oeq float [[TMP4]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP17:%.*]] = call nnan ninf afn float 
@llvm.copysign.f32(float 0.000000e+00, float [[TMP4]])
-; NOPRELINK-NEXT:    [[TMP18:%.*]] = select i1 [[TMP13]], float [[TMP17]], 
float 0.000000e+00
+; NOPRELINK-NEXT:    [[TMP18:%.*]] = select ninf i1 [[TMP13]], float 
[[TMP17]], float 0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP19:%.*]] = select nnan ninf afn i1 [[TMP16]], float 
[[TMP18]], float [[TMP15]]
 ; NOPRELINK-NEXT:    ret float [[TMP19]]
 ;
@@ -4309,7 +4309,7 @@ define float @test_pow_afn_f32_nnan_ninf__y_4_5(float %x) 
{
 ; NOPRELINK-NEXT:    [[TMP19:%.*]] = select nnan ninf afn i1 [[TMP18]], float 
+qnan, float [[TMP14]]
 ; NOPRELINK-NEXT:    [[TMP20:%.*]] = fcmp nnan ninf afn oeq float [[X]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP21:%.*]] = call nnan ninf afn float 
@llvm.copysign.f32(float 0.000000e+00, float [[X]])
-; NOPRELINK-NEXT:    [[TMP22:%.*]] = select i1 [[TMP12]], float [[TMP21]], 
float 0.000000e+00
+; NOPRELINK-NEXT:    [[TMP22:%.*]] = select ninf i1 [[TMP12]], float 
[[TMP21]], float 0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP23:%.*]] = select nnan ninf afn i1 [[TMP20]], float 
[[TMP22]], float [[TMP19]]
 ; NOPRELINK-NEXT:    ret float [[TMP23]]
 ;
@@ -5123,7 +5123,7 @@ define float 
@test_pow_afn_ninf_f32_known_integral_sitofp(float %x, i32 %y) {
 ; NOPRELINK-NEXT:    [[TMP21:%.*]] = select ninf afn i1 [[TMP20]], float 
+qnan, float [[TMP16]]
 ; NOPRELINK-NEXT:    [[TMP22:%.*]] = fcmp ninf afn oeq float [[TMP4]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP23:%.*]] = call ninf afn float 
@llvm.copysign.f32(float 0.000000e+00, float [[TMP4]])
-; NOPRELINK-NEXT:    [[TMP24:%.*]] = select i1 [[TMP14]], float [[TMP23]], 
float 0.000000e+00
+; NOPRELINK-NEXT:    [[TMP24:%.*]] = select ninf i1 [[TMP14]], float 
[[TMP23]], float 0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP25:%.*]] = select ninf afn i1 [[TMP22]], float 
[[TMP24]], float [[TMP21]]
 ; NOPRELINK-NEXT:    [[TMP26:%.*]] = fcmp ninf afn uno float [[TMP4]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP27:%.*]] = select ninf afn i1 [[TMP26]], float 
+qnan, float [[TMP25]]
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-rootn.ll 
b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-rootn.ll
index ff8ad993d7cfa0..64bde03039c3ea 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-rootn.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-rootn.ll
@@ -1176,7 +1176,7 @@ define float @test_rootn_afn_nnan_ninf_f32(float %x, i32 
%y) {
 ; NOPRELINK-NEXT:    [[TMP8:%.*]] = call nnan ninf afn float 
@llvm.copysign.f32(float [[TMP5]], float [[TMP7]])
 ; NOPRELINK-NEXT:    [[TMP9:%.*]] = fcmp nnan ninf afn oeq float [[X]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP10:%.*]] = call nnan ninf afn float 
@llvm.copysign.f32(float 0.000000e+00, float [[X]])
-; NOPRELINK-NEXT:    [[TMP14:%.*]] = select i1 [[DOTNOT]], float 0.000000e+00, 
float [[TMP10]]
+; NOPRELINK-NEXT:    [[TMP14:%.*]] = select ninf i1 [[DOTNOT]], float 
0.000000e+00, float [[TMP10]]
 ; NOPRELINK-NEXT:    [[TMP15:%.*]] = select nnan ninf afn i1 [[TMP9]], float 
[[TMP14]], float [[TMP8]]
 ; NOPRELINK-NEXT:    [[TMP16:%.*]] = fcmp nnan ninf afn olt float [[X]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP17:%.*]] = and i1 [[TMP16]], [[DOTNOT]]
@@ -1314,7 +1314,7 @@ define float @test_rootn_fast_f32_strictfp(float %x, i32 
%y) #1 {
 ; NOPRELINK-NEXT:    [[TMP11:%.*]] = call i1 
@llvm.experimental.constrained.fcmp.f32(float [[X]], float 0.000000e+00, 
metadata !"oeq", metadata !"fpexcept.strict") #[[ATTR0]]
 ; NOPRELINK-NEXT:    [[TMP12:%.*]] = or i1 [[TMP10]], [[TMP11]]
 ; NOPRELINK-NEXT:    [[TMP13:%.*]] = call fast float @llvm.copysign.f32(float 
0.000000e+00, float [[X]]) #[[ATTR0]]
-; NOPRELINK-NEXT:    [[TMP17:%.*]] = select i1 [[DOTNOT]], float 0.000000e+00, 
float [[TMP13]]
+; NOPRELINK-NEXT:    [[TMP17:%.*]] = select ninf i1 [[DOTNOT]], float 
0.000000e+00, float [[TMP13]]
 ; NOPRELINK-NEXT:    [[TMP18:%.*]] = select fast i1 [[TMP12]], float 
[[TMP17]], float [[TMP8]]
 ; NOPRELINK-NEXT:    [[TMP19:%.*]] = call i1 
@llvm.experimental.constrained.fcmp.f32(float [[X]], float 0.000000e+00, 
metadata !"olt", metadata !"fpexcept.strict") #[[ATTR0]]
 ; NOPRELINK-NEXT:    [[TMP20:%.*]] = and i1 [[TMP19]], [[DOTNOT]]
@@ -1750,7 +1750,7 @@ define float 
@test_rootn_afn_ninf_nnan_f32__x_known_positive(float nofpclass(nin
 ; NOPRELINK-NEXT:    [[TMP7:%.*]] = call nnan ninf afn float 
@llvm.copysign.f32(float [[TMP4]], float [[TMP6]])
 ; NOPRELINK-NEXT:    [[TMP8:%.*]] = fcmp nnan ninf afn oeq float [[X]], 
0.000000e+00
 ; NOPRELINK-NEXT:    [[TMP9:%.*]] = call nnan ninf afn float 
@llvm.copysign.f32(float 0.000000e+00, float [[X]])
-; NOPRELINK-NEXT:    [[TMP13:%.*]] = select i1 [[DOTNOT]], float 0.000000e+00, 
float [[TMP9]]
+; NOPRELINK-NEXT:    [[TMP13:%.*]] = select ninf i1 [[DOTNOT]], float 
0.000000e+00, float [[TMP9]]
 ; NOPRELINK-NEXT:    [[TMP14:%.*]] = select nnan ninf afn i1 [[TMP8]], float 
[[TMP13]], float [[TMP7]]
 ; NOPRELINK-NEXT:    [[TMP15:%.*]] = icmp eq i32 [[Y]], 0
 ; NOPRELINK-NEXT:    [[TMP16:%.*]] = select nnan ninf afn i1 [[TMP15]], float 
+qnan, float [[TMP14]]
diff --git a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll 
b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
index 9d5ab6de3e0625..469fa1404d1795 100644
--- a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
+++ b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
@@ -668,7 +668,7 @@ define amdgpu_kernel void @test_pow_half(ptr addrspace(1) 
nocapture %a) {
 ; GCN-POSTLINK-NEXT:    [[TMP15:%.*]] = select fast i1 [[DOTNOT1]], float 
[[TMP13]], float +qnan
 ; GCN-POSTLINK-NEXT:    [[TMP16:%.*]] = fcmp fast oeq float [[TMP]], 
0.000000e+00
 ; GCN-POSTLINK-NEXT:    [[TMP17:%.*]] = tail call fast float 
@llvm.copysign.f32(float 0.000000e+00, float [[TMP]])
-; GCN-POSTLINK-NEXT:    [[TMP19:%.*]] = select i1 [[TMP11]], float [[TMP17]], 
float 0.000000e+00
+; GCN-POSTLINK-NEXT:    [[TMP19:%.*]] = select ninf i1 [[TMP11]], float 
[[TMP17]], float 0.000000e+00
 ; GCN-POSTLINK-NEXT:    [[TMP20:%.*]] = select fast i1 [[TMP16]], float 
[[TMP19]], float [[TMP15]]
 ; GCN-POSTLINK-NEXT:    store float [[TMP20]], ptr addrspace(1) [[A]], align 4
 ; GCN-POSTLINK-NEXT:    ret void
@@ -724,7 +724,7 @@ define amdgpu_kernel void @test_pow_mhalf(ptr addrspace(1) 
nocapture %a) {
 ; GCN-POSTLINK-NEXT:    [[TMP15:%.*]] = select fast i1 [[DOTNOT1]], float 
[[TMP13]], float +qnan
 ; GCN-POSTLINK-NEXT:    [[TMP16:%.*]] = fcmp fast oeq float [[TMP]], 
0.000000e+00
 ; GCN-POSTLINK-NEXT:    [[TMP17:%.*]] = tail call fast float 
@llvm.copysign.f32(float 0.000000e+00, float [[TMP]])
-; GCN-POSTLINK-NEXT:    [[TMP21:%.*]] = select i1 [[TMP11]], float [[TMP17]], 
float 0.000000e+00
+; GCN-POSTLINK-NEXT:    [[TMP21:%.*]] = select ninf i1 [[TMP11]], float 
[[TMP17]], float 0.000000e+00
 ; GCN-POSTLINK-NEXT:    [[TMP22:%.*]] = select fast i1 [[TMP16]], float 
[[TMP21]], float [[TMP15]]
 ; GCN-POSTLINK-NEXT:    store float [[TMP22]], ptr addrspace(1) [[A]], align 4
 ; GCN-POSTLINK-NEXT:    ret void
diff --git a/llvm/test/Transforms/InstCombine/select-ninf.ll 
b/llvm/test/Transforms/InstCombine/select-ninf.ll
new file mode 100644
index 00000000000000..0e2e5a914e9fb1
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/select-ninf.ll
@@ -0,0 +1,39 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 
UTC_ARGS: --version 6
+; RUN: opt -passes=instcombine -S %s | FileCheck %s
+
+define float @nested(i1 %overflow, i1 %lower, float %normal) {
+; CHECK-LABEL: define float @nested(
+; CHECK-SAME: i1 [[OVERFLOW:%.*]], i1 [[LOWER:%.*]], float [[NORMAL:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = select ninf i1 [[LOWER]], float 
-1.000000e+00, float [[NORMAL]]
+; CHECK-NEXT:    ret float [[RESULT]]
+;
+  %upper = select i1 %overflow, float +inf, float %normal
+  %result = select ninf i1 %lower, float -1.0, float %upper
+  ret float %result
+}
+
+define float @nested_unrestricted(i1 %overflow, i1 %lower, float %normal) {
+; CHECK-LABEL: define float @nested_unrestricted(
+; CHECK-SAME: i1 [[OVERFLOW:%.*]], i1 [[LOWER:%.*]], float [[NORMAL:%.*]]) {
+; CHECK-NEXT:    [[UPPER:%.*]] = select i1 [[OVERFLOW]], float +inf, float 
[[NORMAL]]
+; CHECK-NEXT:    [[RESULT:%.*]] = select i1 [[LOWER]], float -1.000000e+00, 
float [[UPPER]]
+; CHECK-NEXT:    ret float [[RESULT]]
+;
+  %upper = select i1 %overflow, float +inf, float %normal
+  %result = select i1 %lower, float -1.0, float %upper
+  ret float %result
+}
+
+define float @nested_shared(i1 %overflow, i1 %lower, float %normal, ptr %out) {
+; CHECK-LABEL: define float @nested_shared(
+; CHECK-SAME: i1 [[OVERFLOW:%.*]], i1 [[LOWER:%.*]], float [[NORMAL:%.*]], ptr 
[[OUT:%.*]]) {
+; CHECK-NEXT:    [[UPPER:%.*]] = select i1 [[OVERFLOW]], float +inf, float 
[[NORMAL]]
+; CHECK-NEXT:    store float [[UPPER]], ptr [[OUT]], align 4
+; CHECK-NEXT:    [[RESULT:%.*]] = select ninf i1 [[LOWER]], float 
-1.000000e+00, float [[UPPER]]
+; CHECK-NEXT:    ret float [[RESULT]]
+;
+  %upper = select i1 %overflow, float +inf, float %normal
+  store float %upper, ptr %out
+  %result = select ninf i1 %lower, float -1.0, float %upper
+  ret float %result
+}

_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits

Reply via email to