[InstCombine] Fold select(X >s 0, 0, -X) | smax(X, 0) to abs(X) #165200

wenju-he · 2025-10-27T04:55:03Z

The IR pattern is compiled from OpenCL code:
__builtin_astype(x > (uchar2)(0) ? x : -x, uchar2);
where smax is created by foldSelectInstWithICmp + canonicalizeSPF.

smax could also come from direct elementwise max call:
int c = b > (int)(0) ? (int)(0) : -b;
int d = __builtin_elementwise_max(b, (int)(0));
*a = c | d;

https://alive2.llvm.org/ce/z/2-brvr
https://alive2.llvm.org/ce/z/Dowjzk
https://alive2.llvm.org/ce/z/kathwZ

The IR pattern is compiled from OpenCL code: __builtin_astype(x > (uchar2)(0) ? x : -x, uchar2); where smax is created by foldSelectInstWithICmp + canonicalizeSPF. smax could also come from direct elementwise max call: int c = b > (int)(0) ? (int)(0) : -b; int d = __builtin_elementwise_max(b, (int)(0)); *a = c | d;

llvmbot · 2025-10-27T04:55:46Z

@llvm/pr-subscribers-clang
@llvm/pr-subscribers-backend-systemz

@llvm/pr-subscribers-llvm-transforms

Author: Wenju He (wenju-he)

Changes

The IR pattern is compiled from OpenCL code:
__builtin_astype(x > (uchar2)(0) ? x : -x, uchar2);
where smax is created by foldSelectInstWithICmp + canonicalizeSPF.

smax could also come from direct elementwise max call:
int c = b > (int)(0) ? (int)(0) : -b;
int d = __builtin_elementwise_max(b, (int)(0));
*a = c | d;

Full diff: https://github.com/llvm/llvm-project/pull/165200.diff

2 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+18)
(modified) llvm/test/Transforms/InstCombine/or.ll (+28)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index 3ddf182149e57..4e863ca2c6dfd 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -3997,6 +3997,20 @@ static Value *foldOrUnsignedUMulOverflowICmp(BinaryOperator &I,
   return nullptr;
 }
 
+// Fold select(X >s 0, 0, -X) | smax(X, 0) --> abs(X)
+static Value *FoldOrOfSelectSmaxToAbs(BinaryOperator &I,
+                                      InstCombiner::BuilderTy &Builder) {
+  CmpPredicate Pred;
+  Value *X;
+  if (match(&I, m_c_Or(m_Select(m_ICmp(Pred, m_Value(X), m_ZeroInt()),
+                                m_ZeroInt(), m_Sub(m_ZeroInt(), m_Deferred(X))),
+                       m_OneUse(m_Intrinsic<Intrinsic::smax>(m_Deferred(X),
+                                                             m_ZeroInt())))) &&
+      Pred == ICmpInst::ICMP_SGT)
+    return Builder.CreateBinaryIntrinsic(Intrinsic::abs, X, Builder.getFalse());
+  return nullptr;
+}
+
 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
 // here. We should standardize that construct where it is needed or choose some
 // other way to ensure that commutated variants of patterns are not missed.
@@ -4545,6 +4559,10 @@ Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
     if (Value *V = SimplifyAddWithRemainder(I))
       return replaceInstUsesWith(I, V);
 
+  // select(X >s 0, 0, -X) | smax(X, 0) -> abs(X)
+  if (Value *Res = FoldOrOfSelectSmaxToAbs(I, Builder))
+    return replaceInstUsesWith(I, Res);
+
   return nullptr;
 }
 
diff --git a/llvm/test/Transforms/InstCombine/or.ll b/llvm/test/Transforms/InstCombine/or.ll
index 6b090e982af0a..bbc79e8c16a56 100644
--- a/llvm/test/Transforms/InstCombine/or.ll
+++ b/llvm/test/Transforms/InstCombine/or.ll
@@ -2113,3 +2113,31 @@ define <4 x i32> @or_zext_nneg_minus_constant_splat(<4 x i8> %a) {
   %or = or <4 x i32> %zext, splat (i32 -9)
   ret <4 x i32> %or
 }
+
+define i8 @or_positive_minus_non_positive_to_abs(i8 noundef %0){
+; CHECK-LABEL: @or_positive_minus_non_positive_to_abs(
+; CHECK-NEXT:    [[TMP2:%.*]] = call i8 @llvm.abs.i8(i8 [[TMP0:%.*]], i1 false)
+; CHECK-NEXT:    ret i8 [[TMP2]]
+;
+  %2 = icmp sgt i8 %0, zeroinitializer
+  %3 = sext i1 %2 to i8
+  %4 = sub i8 zeroinitializer, %0
+  %5 = xor i8 %3, -1
+  %6 = and i8 %4, %5
+  %7 = and i8 %0, %3
+  %8 = or i8 %6, %7
+  ret i8 %8
+}
+
+define <2 x i8> @or_select_smax_to_abs(<2 x i8> %0){
+; CHECK-LABEL: @or_select_smax_to_abs(
+; CHECK-NEXT:    [[TMP2:%.*]] = call <2 x i8> @llvm.abs.v2i8(<2 x i8> [[TMP0:%.*]], i1 false)
+; CHECK-NEXT:    ret <2 x i8> [[TMP2]]
+;
+  %2 = icmp sgt <2 x i8> %0, zeroinitializer
+  %3 = sub <2 x i8> zeroinitializer, %0
+  %4 = select <2 x i1> %2, <2 x i8> zeroinitializer, <2 x i8> %3
+  %5 = tail call <2 x i8> @llvm.smax.v2i8(<2 x i8> %0, <2 x i8> zeroinitializer)
+  %6 = or <2 x i8> %4, %5
+  ret <2 x i8> %6
+}

arsenm

Can you add alive2 link

arsenm · 2025-10-27T05:23:27Z

llvm/test/Transforms/InstCombine/or.ll

+; CHECK-NEXT:    [[TMP2:%.*]] = call i8 @llvm.abs.i8(i8 [[TMP0:%.*]], i1 false)
+; CHECK-NEXT:    ret i8 [[TMP2]]
+;
+  %2 = icmp sgt i8 %0, zeroinitializer


Use named values in tests

Use named values in tests

done

arsenm · 2025-10-27T05:23:47Z

llvm/test/Transforms/InstCombine/or.ll

+  %5 = tail call <2 x i8> @llvm.smax.v2i8(<2 x i8> %0, <2 x i8> zeroinitializer)
+  %6 = or <2 x i8> %4, %5
+  ret <2 x i8> %6
+}


Missing negative test for the multiple use case

Missing negative test for the multiple use case

done

wenju-he · 2025-10-27T06:34:09Z

Can you add alive2 link

done. Updated commit message:
https://alive2.llvm.org/ce/z/2-brvr
https://alive2.llvm.org/ce/z/Dowjzk

dtcxzyw

When we know X is never INT_MIN, select(X >s 0, 0, -X) may be folded into smax(-X, 0): https://alive2.llvm.org/ce/z/wDiDh2
Currently we don't do this fold. Can you please also add a test with smax(-X, 0) | smax(X, 0) and leave it as a todo?

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

dtcxzyw · 2025-10-27T15:31:39Z

llvm/test/Transforms/InstCombine/or.ll

+  ret i8 %or
+}
+
+define <2 x i8> @or_select_smax_to_abs(<2 x i8> %a){


Please add a test with select(X <s 0, -X, 0) | smax(X, 0) --> abs(X).

Please add a test with select(X <s 0, -X, 0) | smax(X, 0) --> abs(X).

thanks, done in bd8b791

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

Co-authored-by: Yingwei Zheng <[email protected]>

github-actions · 2025-10-28T02:03:14Z

✅ With the latest revision this PR passed the C/C++ code formatter.

wenju-he · 2025-10-28T02:38:24Z

When we know X is never INT_MIN, select(X >s 0, 0, -X) may be folded into smax(-X, 0): https://alive2.llvm.org/ce/z/wDiDh2 Currently we don't do this fold. Can you please also add a test with smax(-X, 0) | smax(X, 0) and leave it as a todo?

thanks, added 2 TODO tests in 1e9a8ce

uweigand · 2025-10-28T09:09:57Z

clang/test/CodeGen/SystemZ/builtins-systemz-zvector.c


  vsc = vec_abs(vsc);
-  // CHECK-ASM: vlcb
+  // CHECK-ASM: vlpb


The SystemZ changes LGTM. In fact, this fixes a regression I hadn't even been aware of, which was introduced here: 1a60ae0

The SystemZ changes LGTM. In fact, this fixes a regression I hadn't even been aware of, which was introduced here: 1a60ae0

thanks @uweigand for review

dtcxzyw · 2025-10-30T17:15:12Z

llvm/test/Transforms/InstCombine/or.ll

+  ret <2 x i8> %or
+}
+
+define <2 x i8> @or_lgt_select_smax_to_abs(<2 x i8> %a){


Suggested change

define <2 x i8> @or_lgt_select_smax_to_abs(<2 x i8> %a){

define <2 x i8> @or_slt_select_smax_to_abs(<2 x i8> %a){

Suggested change
define <2 x i8> @or_lgt_select_smax_to_abs(<2 x i8> %a){
define <2 x i8> @or_slt_select_smax_to_abs(<2 x i8> %a){

done, thanks @dtcxzyw

arsenm · 2025-10-31T00:37:12Z

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

+                                      InstCombiner::BuilderTy &Builder) {
+  Value *X;
+  Value *Sel;
+  if (match(&I, m_c_Or(m_Value(Sel), m_OneUse(m_Intrinsic<Intrinsic::smax>(


Use m_SMax

done

llvm/test/Transforms/InstCombine/or.ll

Co-authored-by: Matt Arsenault <[email protected]>

dtcxzyw

LGTM.

…#165200) The IR pattern is compiled from OpenCL code: __builtin_astype(x > (uchar2)(0) ? x : -x, uchar2); where smax is created by foldSelectInstWithICmp + canonicalizeSPF. smax could also come from direct elementwise max call: int c = b > (int)(0) ? (int)(0) : -b; int d = __builtin_elementwise_max(b, (int)(0)); *a = c | d; https://alive2.llvm.org/ce/z/2-brvr https://alive2.llvm.org/ce/z/Dowjzk https://alive2.llvm.org/ce/z/kathwZ --------- Co-authored-by: Yingwei Zheng <[email protected]> Co-authored-by: Matt Arsenault <[email protected]>

wenju-he requested a review from nikic as a code owner October 27, 2025 04:55

llvmbot added llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:transforms labels Oct 27, 2025

wenju-he requested a review from dtcxzyw October 27, 2025 04:55

wenju-he mentioned this pull request Oct 27, 2025

[libclc] Implement integer __clc_abs using __builtin_elementwise_abs #164957

Merged

wenju-he requested a review from arsenm October 27, 2025 04:57

arsenm reviewed Oct 27, 2025

View reviewed changes

named value, add negative test with multiple uses of @llvm.smax

af62083

wenju-he requested a review from arsenm October 27, 2025 06:34

rename %e -> %add

21ca09a

dtcxzyw reviewed Oct 27, 2025

View reviewed changes

wenju-he and others added 4 commits October 28, 2025 02:58

add TODO: fold to smax(neg, fold two smax

1e9a8ce

Update llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

e5d284b

Co-authored-by: Yingwei Zheng <[email protected]>

Update llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

e19950b

Co-authored-by: Yingwei Zheng <[email protected]>

clang-format

28f1987

select(X <s 0, -X, 0) | smax(X, 0) --> abs(X)

bd8b791

wenju-he requested a review from dtcxzyw October 28, 2025 02:38

update clang/test/CodeGen/SystemZ/builtins-systemz-zvector.c

154c2c1

llvmbot added clang Clang issues not falling into any other category backend:SystemZ labels Oct 28, 2025

uweigand reviewed Oct 28, 2025

View reviewed changes

Merge branch 'main' into instcombine-or-select-smax-to-abs

3aaa55a

dtcxzyw mentioned this pull request Oct 30, 2025

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

dtcxzyw reviewed Oct 30, 2025

View reviewed changes

dtcxzyw mentioned this pull request Oct 30, 2025

Fuzz PR165200 dtcxzyw/llvm-mutation-based-fuzz-service#115

Closed

zyw-bot mentioned this pull request Oct 30, 2025

pre-commit: PR165200 dtcxzyw/llvm-opt-benchmark#3004

Closed

rename or_lgt_select_smax_to_abs -> or_slt_select_smax_to_abs

c01eb5a

wenju-he requested a review from dtcxzyw October 30, 2025 23:51

arsenm reviewed Oct 31, 2025

View reviewed changes

wenju-he and others added 2 commits October 31, 2025 15:32

Update llvm/test/Transforms/InstCombine/or.ll

4c92c7c

Co-authored-by: Matt Arsenault <[email protected]>

use m_SMax

d6eecb3

wenju-he requested a review from arsenm October 31, 2025 07:34

dtcxzyw approved these changes Oct 31, 2025

View reviewed changes

wenju-he merged commit 79bf8c0 into llvm:main Nov 2, 2025
13 of 14 checks passed

wenju-he deleted the instcombine-or-select-smax-to-abs branch November 2, 2025 23:38

	define <2 x i8> @or_lgt_select_smax_to_abs(<2 x i8> %a){
	define <2 x i8> @or_slt_select_smax_to_abs(<2 x i8> %a){

[InstCombine] Fold select(X >s 0, 0, -X) | smax(X, 0) to abs(X) #165200

[InstCombine] Fold select(X >s 0, 0, -X) | smax(X, 0) to abs(X) #165200

Uh oh!

Conversation

wenju-he commented Oct 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Oct 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

arsenm left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wenju-he commented Oct 27, 2025

Uh oh!

dtcxzyw left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions bot commented Oct 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wenju-he commented Oct 28, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

dtcxzyw left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

wenju-he commented Oct 27, 2025 •

edited

Loading

llvmbot commented Oct 27, 2025 •

edited

Loading

github-actions bot commented Oct 28, 2025 •

edited

Loading