[InstCombine] Preserve signbit semantics of NaN with fold to fabs #136648

dtcxzyw · 2025-04-22T03:24:12Z

As per the LangRef and IEEE 754-2008 standard, the sign bit of NaN is preserved if there is no floating-point operation being performed.
See also 862e35e for reference.

Alive2: https://alive2.llvm.org/ce/z/QYtEGj
Closes #136646

llvmbot · 2025-04-22T03:24:45Z

@llvm/pr-subscribers-llvm-transforms

Author: Yingwei Zheng (dtcxzyw)

Changes

As per the LangRef and IEEE 754-2008 standard, the sign bit of NaN is preserved if there is no floating-point operation being performed.
See also 862e35e for reference.

Alive2: https://alive2.llvm.org/ce/z/4mBcJx
Closes #136646

Full diff: https://github.com/llvm/llvm-project/pull/136648.diff

2 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp (+5-1)
(modified) llvm/test/Transforms/InstCombine/fabs.ll (+24-11)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index 4bba2f406b4c1..06792e32de3e1 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -2793,7 +2793,11 @@ static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI,
 
     // fold (X <= +/-0.0) ? (0.0 - X) : X to fabs(X), when 'Swap' is false
     // fold (X >  +/-0.0) ? X : (0.0 - X) to fabs(X), when 'Swap' is true
-    if (match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(X)))) {
+    // Note: We require "nnan" for this fold because fcmp ignores the signbit
+    //       of NAN, but IEEE-754 specifies the signbit of NAN values with
+    //       fneg/fabs operations.
+    if (cast<FPMathOperator>(CondVal)->hasNoNaNs() &&
+        match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(X)))) {
       if (!Swap && (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) {
         Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI);
         return IC.replaceInstUsesWith(SI, Fabs);
diff --git a/llvm/test/Transforms/InstCombine/fabs.ll b/llvm/test/Transforms/InstCombine/fabs.ll
index 7b9a672f188ca..827315b184085 100644
--- a/llvm/test/Transforms/InstCombine/fabs.ll
+++ b/llvm/test/Transforms/InstCombine/fabs.ll
@@ -256,6 +256,19 @@ define double @select_fcmp_ole_zero(double %x) {
 ; CHECK-LABEL: @select_fcmp_ole_zero(
 ; CHECK-NEXT:    [[FABS:%.*]] = call double @llvm.fabs.f64(double [[X:%.*]])
 ; CHECK-NEXT:    ret double [[FABS]]
+;
+  %lezero = fcmp nnan ole double %x, 0.0
+  %negx = fsub double 0.0, %x
+  %fabs = select i1 %lezero, double %negx, double %x
+  ret double %fabs
+}
+
+define double @select_fcmp_ole_zero_no_nnan(double %x) {
+; CHECK-LABEL: @select_fcmp_ole_zero_no_nnan(
+; CHECK-NEXT:    [[LEZERO:%.*]] = fcmp ole double [[X:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[NEGX:%.*]] = fsub double 0.000000e+00, [[X]]
+; CHECK-NEXT:    [[FABS:%.*]] = select i1 [[LEZERO]], double [[NEGX]], double [[X]]
+; CHECK-NEXT:    ret double [[FABS]]
 ;
   %lezero = fcmp ole double %x, 0.0
   %negx = fsub double 0.0, %x
@@ -268,7 +281,7 @@ define double @select_fcmp_nnan_ole_zero(double %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call double @llvm.fabs.f64(double [[X:%.*]])
 ; CHECK-NEXT:    ret double [[FABS]]
 ;
-  %lezero = fcmp ole double %x, 0.0
+  %lezero = fcmp nnan ole double %x, 0.0
   %negx = fsub nnan double 0.0, %x
   %fabs = select i1 %lezero, double %negx, double %x
   ret double %fabs
@@ -279,7 +292,7 @@ define double @select_nnan_fcmp_nnan_ole_zero(double %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call nnan double @llvm.fabs.f64(double [[X:%.*]])
 ; CHECK-NEXT:    ret double [[FABS]]
 ;
-  %lezero = fcmp ole double %x, 0.0
+  %lezero = fcmp nnan ole double %x, 0.0
   %negx = fsub nnan double 0.0, %x
   %fabs = select nnan i1 %lezero, double %negx, double %x
   ret double %fabs
@@ -292,7 +305,7 @@ define double @select_fcmp_nnan_ule_zero(double %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call double @llvm.fabs.f64(double [[X:%.*]])
 ; CHECK-NEXT:    ret double [[FABS]]
 ;
-  %lezero = fcmp ule double %x, 0.0
+  %lezero = fcmp nnan ule double %x, 0.0
   %negx = fsub nnan double 0.0, %x
   %fabs = select i1 %lezero, double %negx, double %x
   ret double %fabs
@@ -320,7 +333,7 @@ define <2 x float> @select_fcmp_nnan_ole_negzero(<2 x float> %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call <2 x float> @llvm.fabs.v2f32(<2 x float> [[X:%.*]])
 ; CHECK-NEXT:    ret <2 x float> [[FABS]]
 ;
-  %lezero = fcmp ole <2 x float> %x, <float -0.0, float -0.0>
+  %lezero = fcmp nnan ole <2 x float> %x, <float -0.0, float -0.0>
   %negx = fsub nnan <2 x float> <float 0.0, float poison>, %x
   %fabs = select <2 x i1> %lezero, <2 x float> %negx, <2 x float> %x
   ret <2 x float> %fabs
@@ -331,7 +344,7 @@ define <2 x float> @select_nnan_fcmp_nnan_ole_negzero(<2 x float> %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call nnan <2 x float> @llvm.fabs.v2f32(<2 x float> [[X:%.*]])
 ; CHECK-NEXT:    ret <2 x float> [[FABS]]
 ;
-  %lezero = fcmp ole <2 x float> %x, <float -0.0, float -0.0>
+  %lezero = fcmp nnan ole <2 x float> %x, <float -0.0, float -0.0>
   %negx = fsub nnan <2 x float> <float 0.0, float poison>, %x
   %fabs = select nnan <2 x i1> %lezero, <2 x float> %negx, <2 x float> %x
   ret <2 x float> %fabs
@@ -344,7 +357,7 @@ define fp128 @select_fcmp_ogt_zero(fp128 %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call fp128 @llvm.fabs.f128(fp128 [[X:%.*]])
 ; CHECK-NEXT:    ret fp128 [[FABS]]
 ;
-  %gtzero = fcmp ogt fp128 %x, zeroinitializer
+  %gtzero = fcmp nnan ogt fp128 %x, zeroinitializer
   %negx = fsub fp128 zeroinitializer, %x
   %fabs = select i1 %gtzero, fp128 %x, fp128 %negx
   ret fp128 %fabs
@@ -382,7 +395,7 @@ define fp128 @select_fcmp_nnan_ogt_zero(fp128 %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call fp128 @llvm.fabs.f128(fp128 [[X:%.*]])
 ; CHECK-NEXT:    ret fp128 [[FABS]]
 ;
-  %gtzero = fcmp ogt fp128 %x, zeroinitializer
+  %gtzero = fcmp nnan ogt fp128 %x, zeroinitializer
   %negx = fsub nnan fp128 zeroinitializer, %x
   %fabs = select i1 %gtzero, fp128 %x, fp128 %negx
   ret fp128 %fabs
@@ -393,7 +406,7 @@ define fp128 @select_nnan_fcmp_nnan_ogt_zero(fp128 %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call nnan fp128 @llvm.fabs.f128(fp128 [[X:%.*]])
 ; CHECK-NEXT:    ret fp128 [[FABS]]
 ;
-  %gtzero = fcmp ogt fp128 %x, zeroinitializer
+  %gtzero = fcmp nnan ogt fp128 %x, zeroinitializer
   %negx = fsub nnan fp128 zeroinitializer, %x
   %fabs = select nnan i1 %gtzero, fp128 %x, fp128 %negx
   ret fp128 %fabs
@@ -406,7 +419,7 @@ define half @select_fcmp_nnan_ogt_negzero(half %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call half @llvm.fabs.f16(half [[X:%.*]])
 ; CHECK-NEXT:    ret half [[FABS]]
 ;
-  %gtzero = fcmp ogt half %x, -0.0
+  %gtzero = fcmp nnan ogt half %x, -0.0
   %negx = fsub nnan half 0.0, %x
   %fabs = select i1 %gtzero, half %x, half %negx
   ret half %fabs
@@ -417,7 +430,7 @@ define half @select_nnan_fcmp_nnan_ogt_negzero(half %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call nnan half @llvm.fabs.f16(half [[X:%.*]])
 ; CHECK-NEXT:    ret half [[FABS]]
 ;
-  %gtzero = fcmp ogt half %x, -0.0
+  %gtzero = fcmp nnan ogt half %x, -0.0
   %negx = fsub nnan half 0.0, %x
   %fabs = select nnan i1 %gtzero, half %x, half %negx
   ret half %fabs
@@ -430,7 +443,7 @@ define half @select_fcmp_nnan_ugt_negzero(half %x) {
 ; CHECK-NEXT:    [[FABS:%.*]] = call half @llvm.fabs.f16(half [[X:%.*]])
 ; CHECK-NEXT:    ret half [[FABS]]
 ;
-  %gtzero = fcmp ugt half %x, -0.0
+  %gtzero = fcmp nnan ugt half %x, -0.0
   %negx = fsub nnan half 0.0, %x
   %fabs = select i1 %gtzero, half %x, half %negx
   ret half %fabs

dtcxzyw · 2025-04-22T04:04:28Z

No real-world impact.

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

dtcxzyw · 2025-04-28T09:27:14Z

/cherry-pick 3e1e406

llvmbot · 2025-04-28T09:34:34Z

/cherry-pick 3e1e406

Error: Command failed due to missing milestone.

dtcxzyw · 2025-04-28T09:36:17Z

/cherry-pick 3e1e406

llvmbot · 2025-04-28T09:43:59Z

/pull-request #137608

…vm#136648) As per the LangRef and IEEE 754-2008 standard, the sign bit of NaN is preserved if there is no floating-point operation being performed. See also llvm@862e35e for reference. Alive2: https://alive2.llvm.org/ce/z/QYtEGj Closes llvm#136646

…vm#136648) As per the LangRef and IEEE 754-2008 standard, the sign bit of NaN is preserved if there is no floating-point operation being performed. See also llvm@862e35e for reference. Alive2: https://alive2.llvm.org/ce/z/QYtEGj Closes llvm#136646 (cherry picked from commit 3e1e406)

…vm#136648) As per the LangRef and IEEE 754-2008 standard, the sign bit of NaN is preserved if there is no floating-point operation being performed. See also llvm@862e35e for reference. Alive2: https://alive2.llvm.org/ce/z/QYtEGj Closes llvm#136646

[InstCombine] Preserve signbit semantics of NaN with fold to fabs

98992d3

dtcxzyw requested review from arsenm and jcranmer-intel April 22, 2025 03:24

dtcxzyw requested a review from nikic as a code owner April 22, 2025 03:24

llvmbot added llvm:instcombine llvm:transforms labels Apr 22, 2025

This was referenced Apr 22, 2025

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

Fuzz PR136648 dtcxzyw/llvm-mutation-based-fuzz-service#36

Closed

pre-commit: PR136648 dtcxzyw/llvm-opt-benchmark#2291

Closed

dtcxzyw added the floating-point Floating-point math label Apr 22, 2025

arsenm reviewed Apr 23, 2025

View reviewed changes

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp Outdated Show resolved Hide resolved

dtcxzyw added 2 commits April 23, 2025 20:08

[InstCombine] Use isKnownNeverNaN

fc2ed05

[InstCombine] Check nnan on select

322bb13

arsenm approved these changes Apr 25, 2025

View reviewed changes

dtcxzyw merged commit 3e1e406 into llvm:main Apr 26, 2025
11 checks passed

dtcxzyw deleted the perf/fabs-nnan branch April 26, 2025 06:03

dtcxzyw added this to the LLVM 20.X Release milestone Apr 28, 2025

github-project-automation bot added this to LLVM Release Status Apr 28, 2025

github-project-automation bot moved this to Needs Triage in LLVM Release Status Apr 28, 2025

llvmbot moved this from Needs Triage to Done in LLVM Release Status Apr 28, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[InstCombine] Preserve signbit semantics of NaN with fold to fabs #136648

[InstCombine] Preserve signbit semantics of NaN with fold to fabs #136648

dtcxzyw commented Apr 22, 2025 •

edited

Loading

llvmbot commented Apr 22, 2025

dtcxzyw commented Apr 22, 2025

dtcxzyw commented Apr 28, 2025

llvmbot commented Apr 28, 2025

dtcxzyw commented Apr 28, 2025

llvmbot commented Apr 28, 2025

[InstCombine] Preserve signbit semantics of NaN with fold to fabs #136648

[InstCombine] Preserve signbit semantics of NaN with fold to fabs #136648

Conversation

dtcxzyw commented Apr 22, 2025 • edited Loading

llvmbot commented Apr 22, 2025

dtcxzyw commented Apr 22, 2025

dtcxzyw commented Apr 28, 2025

llvmbot commented Apr 28, 2025

dtcxzyw commented Apr 28, 2025

llvmbot commented Apr 28, 2025

dtcxzyw commented Apr 22, 2025 •

edited

Loading