[InstCombine] Fold xor of bittests into bittest of xor'd value #125676

dtcxzyw · 2025-02-04T12:00:10Z

Motivating case:

llvm-project/llvm/lib/Analysis/ValueTracking.cpp

Lines 8600 to 8602 in 64927af

    
           // They must both have samesign flag or not. 
        
           if (cast<ICmpInst>(X)->hasSameSign() != cast<ICmpInst>(Y)->hasSameSign()) 
        
             return false;

It is translated into xor (X & 2) != 0, (Y & 2) != 0.
Alive2: https://alive2.llvm.org/ce/z/dJehZ8

llvmbot · 2025-02-04T12:00:44Z

@llvm/pr-subscribers-llvm-transforms

Author: Yingwei Zheng (dtcxzyw)

Changes

Motivating case:

llvm-project/llvm/lib/Analysis/ValueTracking.cpp

Lines 8600 to 8602 in 64927af

    
           // They must both have samesign flag or not. 
        
           if (cast<ICmpInst>(X)->hasSameSign() != cast<ICmpInst>(Y)->hasSameSign()) 
        
             return false;

It is translated into xor (X & 2) != 0, (Y & 2) != 0.
Alive2: https://alive2.llvm.org/ce/z/dJehZ8

Full diff: https://github.com/llvm/llvm-project/pull/125676.diff

2 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+15-3)
(modified) llvm/test/Transforms/InstCombine/xor-icmps.ll (+155)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index ca8a20b4b7312d..8701f7c28a39fc 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -4150,9 +4150,6 @@ Value *InstCombinerImpl::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS,
     }
   }
 
-  // TODO: This can be generalized to compares of non-signbits using
-  // decomposeBitTestICmp(). It could be enhanced more by using (something like)
-  // foldLogOpOfMaskedICmps().
   const APInt *LC, *RC;
   if (match(LHS1, m_APInt(LC)) && match(RHS1, m_APInt(RC)) &&
       LHS0->getType() == RHS0->getType() &&
@@ -4200,6 +4197,21 @@ Value *InstCombinerImpl::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS,
           }
         }
     }
+
+    // Fold (icmp eq/ne (X & Pow2), 0) ^ (icmp eq/ne (Y & Pow2), 0) into
+    // (icmp eq/ne ((X ^ Y) & Pow2), 0)
+    Value *X, *Y;
+    const APInt *Mask;
+    if (ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
+        LC->isZero() && RC->isZero() && LHS->hasOneUse() && RHS->hasOneUse() &&
+        match(LHS0, m_And(m_Value(X), m_Power2(Mask))) &&
+        match(RHS0, m_And(m_Value(Y), m_SpecificInt(*Mask)))) {
+      Value *Xor = Builder.CreateXor(X, Y);
+      Value *And = Builder.CreateAnd(Xor, *Mask);
+      return Builder.CreateICmp(PredL == PredR ? ICmpInst::ICMP_NE
+                                               : ICmpInst::ICMP_EQ,
+                                And, ConstantInt::getNullValue(Xor->getType()));
+    }
   }
 
   // Instead of trying to imitate the folds for and/or, decompose this 'xor'
diff --git a/llvm/test/Transforms/InstCombine/xor-icmps.ll b/llvm/test/Transforms/InstCombine/xor-icmps.ll
index 0384c1aa184b82..382355cae7694d 100644
--- a/llvm/test/Transforms/InstCombine/xor-icmps.ll
+++ b/llvm/test/Transforms/InstCombine/xor-icmps.ll
@@ -319,3 +319,158 @@ define i1 @xor_icmp_to_icmp_add_multiuse2(i32 %a) {
   %cmp3 = xor i1 %cmp, %cmp1
   ret i1 %cmp3
 }
+
+define i1 @test_xor_of_bittest_ne_ne(i8 %x, i8 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_ne_ne(
+; CHECK-NEXT:    [[Y:%.*]] = xor i8 [[X:%.*]], [[Y1:%.*]]
+; CHECK-NEXT:    [[MASK2:%.*]] = and i8 [[Y]], 2
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp ne i8 [[MASK2]], 0
+; CHECK-NEXT:    ret i1 [[CMP2]]
+;
+  %mask1 = and i8 %x, 2
+  %cmp1 = icmp ne i8 %mask1, 0
+  %mask2 = and i8 %y, 2
+  %cmp2 = icmp ne i8 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+define i1 @test_xor_of_bittest_eq_eq(i8 %x, i8 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_eq_eq(
+; CHECK-NEXT:    [[Y:%.*]] = xor i8 [[X:%.*]], [[Y1:%.*]]
+; CHECK-NEXT:    [[MASK2:%.*]] = and i8 [[Y]], 2
+; CHECK-NEXT:    [[XOR:%.*]] = icmp ne i8 [[MASK2]], 0
+; CHECK-NEXT:    ret i1 [[XOR]]
+;
+  %mask1 = and i8 %x, 2
+  %cmp1 = icmp eq i8 %mask1, 0
+  %mask2 = and i8 %y, 2
+  %cmp2 = icmp eq i8 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+define i1 @test_xor_of_bittest_ne_eq(i8 %x, i8 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_ne_eq(
+; CHECK-NEXT:    [[Y:%.*]] = xor i8 [[X:%.*]], [[Y1:%.*]]
+; CHECK-NEXT:    [[MASK2:%.*]] = and i8 [[Y]], 2
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp eq i8 [[MASK2]], 0
+; CHECK-NEXT:    ret i1 [[CMP2]]
+;
+  %mask1 = and i8 %x, 2
+  %cmp1 = icmp ne i8 %mask1, 0
+  %mask2 = and i8 %y, 2
+  %cmp2 = icmp eq i8 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+define i1 @test_xor_of_bittest_eq_ne(i8 %x, i8 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_eq_ne(
+; CHECK-NEXT:    [[X:%.*]] = xor i8 [[X1:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[MASK1:%.*]] = and i8 [[X]], 2
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i8 [[MASK1]], 0
+; CHECK-NEXT:    ret i1 [[CMP1]]
+;
+  %mask1 = and i8 %x, 2
+  %cmp1 = icmp eq i8 %mask1, 0
+  %mask2 = and i8 %y, 2
+  %cmp2 = icmp ne i8 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+define i1 @test_xor_of_bittest_ne_ne_multiuse1(i8 %x, i8 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_ne_ne_multiuse1(
+; CHECK-NEXT:    [[MASK1:%.*]] = and i8 [[X:%.*]], 2
+; CHECK-NEXT:    call void @usei8(i8 [[MASK1]])
+; CHECK-NEXT:    [[MASK2:%.*]] = and i8 [[Y:%.*]], 2
+; CHECK-NEXT:    call void @usei8(i8 [[MASK2]])
+; CHECK-NEXT:    [[TMP1:%.*]] = xor i8 [[X]], [[Y]]
+; CHECK-NEXT:    [[TMP2:%.*]] = and i8 [[TMP1]], 2
+; CHECK-NEXT:    [[XOR:%.*]] = icmp ne i8 [[TMP2]], 0
+; CHECK-NEXT:    ret i1 [[XOR]]
+;
+  %mask1 = and i8 %x, 2
+  call void @usei8(i8 %mask1)
+  %cmp1 = icmp ne i8 %mask1, 0
+  %mask2 = and i8 %y, 2
+  call void @usei8(i8 %mask2)
+  %cmp2 = icmp ne i8 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+; Negative tests
+
+define i1 @test_xor_of_bittest_ne_ne_type_mismatch(i8 %x, i16 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_ne_ne_type_mismatch(
+; CHECK-NEXT:    [[MASK1:%.*]] = and i8 [[X:%.*]], 2
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i8 [[MASK1]], 0
+; CHECK-NEXT:    [[MASK2:%.*]] = and i16 [[Y:%.*]], 2
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp ne i16 [[MASK2]], 0
+; CHECK-NEXT:    [[XOR:%.*]] = xor i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT:    ret i1 [[XOR]]
+;
+  %mask1 = and i8 %x, 2
+  %cmp1 = icmp ne i8 %mask1, 0
+  %mask2 = and i16 %y, 2
+  %cmp2 = icmp ne i16 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+define i1 @test_xor_of_bittest_ne_ne_mask_mismatch(i8 %x, i8 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_ne_ne_mask_mismatch(
+; CHECK-NEXT:    [[MASK1:%.*]] = and i8 [[X:%.*]], 4
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i8 [[MASK1]], 0
+; CHECK-NEXT:    [[MASK2:%.*]] = and i8 [[Y:%.*]], 2
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp ne i8 [[MASK2]], 0
+; CHECK-NEXT:    [[XOR:%.*]] = xor i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT:    ret i1 [[XOR]]
+;
+  %mask1 = and i8 %x, 4
+  %cmp1 = icmp ne i8 %mask1, 0
+  %mask2 = and i8 %y, 2
+  %cmp2 = icmp ne i8 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+define i1 @test_xor_of_bittest_ne_ne_nonpower2(i8 %x, i8 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_ne_ne_nonpower2(
+; CHECK-NEXT:    [[MASK1:%.*]] = and i8 [[X:%.*]], 3
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i8 [[MASK1]], 0
+; CHECK-NEXT:    [[MASK2:%.*]] = and i8 [[Y:%.*]], 3
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp ne i8 [[MASK2]], 0
+; CHECK-NEXT:    [[XOR:%.*]] = xor i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT:    ret i1 [[XOR]]
+;
+  %mask1 = and i8 %x, 3
+  %cmp1 = icmp ne i8 %mask1, 0
+  %mask2 = and i8 %y, 3
+  %cmp2 = icmp ne i8 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+define i1 @test_xor_of_bittest_ne_ne_multiuse2(i8 %x, i8 %y) {
+; CHECK-LABEL: @test_xor_of_bittest_ne_ne_multiuse2(
+; CHECK-NEXT:    [[MASK1:%.*]] = and i8 [[X:%.*]], 2
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i8 [[MASK1]], 0
+; CHECK-NEXT:    call void @use(i1 [[CMP1]])
+; CHECK-NEXT:    [[MASK2:%.*]] = and i8 [[Y:%.*]], 2
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp ne i8 [[MASK2]], 0
+; CHECK-NEXT:    [[XOR:%.*]] = xor i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT:    ret i1 [[XOR]]
+;
+  %mask1 = and i8 %x, 2
+  %cmp1 = icmp ne i8 %mask1, 0
+  call void @use(i1 %cmp1)
+  %mask2 = and i8 %y, 2
+  %cmp2 = icmp ne i8 %mask2, 0
+  %xor = xor i1 %cmp1, %cmp2
+  ret i1 %xor
+}
+
+declare void @usei8(i8)

nikic

LGTM

goldsteinn

LGTM

goldsteinn · 2025-02-04T15:44:42Z

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

+    if (ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
+        LC->isZero() && RC->isZero() && LHS->hasOneUse() && RHS->hasOneUse() &&
+        match(LHS0, m_And(m_Value(X), m_Power2(Mask))) &&
+        match(RHS0, m_And(m_Value(Y), m_SpecificInt(*Mask)))) {


Any value to handling any Value with IsKnownPow2?

I will try it after landing this patch.

andjo403 · 2025-02-04T19:15:33Z

is there a benefit to handle the trunc to i1 or other bittest patterns?

dtcxzyw · 2025-02-04T21:30:29Z

is there a benefit to handle the trunc to i1 or other bittest patterns?

trunc to i1: https://godbolt.org/z/3r93zKvbr
Xor of sign bit tests has been handled by the logic above.

Alive2: https://alive2.llvm.org/ce/z/dJehZ8

…125676) Motivating case: https://github.com/llvm/llvm-project/blob/64927af52a3bedf2b20d6cdd98bb47d9bba630f9/llvm/lib/Analysis/ValueTracking.cpp#L8600-L8602 It is translated into `xor (X & 2) != 0, (Y & 2) != 0`. Alive2: https://alive2.llvm.org/ce/z/dJehZ8

…125855) Alive2: https://alive2.llvm.org/ce/z/dJehZ8

dtcxzyw added 2 commits February 4, 2025 19:28

[InstCombine] Add pre-commit tests. NFC.

dceb898

[InstCombine] Fold xor of bittests intto bittest of xor'd values

7c18e17

dtcxzyw requested a review from goldsteinn February 4, 2025 12:00

dtcxzyw requested a review from nikic as a code owner February 4, 2025 12:00

llvmbot added llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:transforms labels Feb 4, 2025

This was referenced Feb 4, 2025

Fuzz PR125676 dtcxzyw/llvm-mutation-based-fuzz-service#26

Closed

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

pre-commit: PR125676 dtcxzyw/llvm-opt-benchmark#2082

Closed

nikic approved these changes Feb 4, 2025

View reviewed changes

goldsteinn approved these changes Feb 4, 2025

View reviewed changes

goldsteinn reviewed Feb 4, 2025

View reviewed changes

dtcxzyw merged commit fad6375 into llvm:main Feb 4, 2025
11 checks passed

dtcxzyw deleted the perf/xor-bittest branch February 4, 2025 21:38

dtcxzyw added a commit to dtcxzyw/llvm-project that referenced this pull request Feb 5, 2025

[InstCombine] Extend llvm#125676 to handle variable power of 2

0867d3d

dtcxzyw added a commit that referenced this pull request Feb 6, 2025

[InstCombine] Extend #125676 to handle variable power of 2 (#125855)

6410bdd

Alive2: https://alive2.llvm.org/ce/z/dJehZ8

Icohedron pushed a commit to Icohedron/llvm-project that referenced this pull request Feb 11, 2025

[InstCombine] Extend llvm#125676 to handle variable power of 2 (llvm#…

3e7d572

…125855) Alive2: https://alive2.llvm.org/ce/z/dJehZ8

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[InstCombine] Fold xor of bittests into bittest of xor'd value #125676

[InstCombine] Fold xor of bittests into bittest of xor'd value #125676

Uh oh!

dtcxzyw commented Feb 4, 2025

Uh oh!

llvmbot commented Feb 4, 2025

Uh oh!

nikic left a comment

Uh oh!

goldsteinn left a comment

Uh oh!

goldsteinn Feb 4, 2025

Uh oh!

dtcxzyw Feb 4, 2025

Uh oh!

andjo403 commented Feb 4, 2025

Uh oh!

dtcxzyw commented Feb 4, 2025

Uh oh!

Uh oh!

Uh oh!

	// They must both have samesign flag or not.
	if (cast<ICmpInst>(X)->hasSameSign() != cast<ICmpInst>(Y)->hasSameSign())
	return false;

[InstCombine] Fold xor of bittests into bittest of xor'd value #125676

[InstCombine] Fold xor of bittests into bittest of xor'd value #125676

Uh oh!

Conversation

dtcxzyw commented Feb 4, 2025

Uh oh!

llvmbot commented Feb 4, 2025

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

goldsteinn left a comment

Choose a reason for hiding this comment

Uh oh!

goldsteinn Feb 4, 2025

Choose a reason for hiding this comment

Uh oh!

dtcxzyw Feb 4, 2025

Choose a reason for hiding this comment

Uh oh!

andjo403 commented Feb 4, 2025

Uh oh!

dtcxzyw commented Feb 4, 2025

Uh oh!

Uh oh!

Uh oh!