[llvm] [InstCombine] fold "icmp eq (X + (V - 1)) & -V, X" to "icmp eq (and X, V - 1), 0" #152851

pskrgag · 2025-08-09T12:09:12Z

This fold optimizes

define i1 @src(i32 %num, i32 %val) {
  %mask = add i32 %val, -1
  %neg = sub nsw i32 0, %val

  %num.biased = add i32 %num, %mask
  %_2.sroa.0.0 = and i32 %num.biased, %neg
  %_0 = icmp eq i32 %_2.sroa.0.0, %num
  ret i1 %_0
}

to

define i1 @tgt(i32 %num, i32 %val) {
  %mask = add i32 %val, -1
  %tmp = and i32 %num, %mask
  %ret = icmp eq i32 %tmp, 0
  ret i1 %ret
}

For power-of-two val.

Observed in real life for following code

pub fn is_aligned(num: usize) -> bool {
    num.next_multiple_of(1 << 12) == num
}

which verifies that num is aligned to 4096.

Alive2 proof https://alive2.llvm.org/ce/z/QisECm

… val - 1)"

llvmbot · 2025-08-09T12:09:46Z

@llvm/pr-subscribers-llvm-transforms

Author: Pavel Skripkin (pskrgag)

Changes

This fold optimizes

define i1 @<!-- -->src(i32 %num, i32 %val) {
  %mask = add i32 %val, -1
  %neg = sub nsw i32 0, %val

  %num.biased = add i32 %num, %mask
  %_2.sroa.0.0 = and i32 %num.biased, %neg
  %_0 = icmp eq i32 %_2.sroa.0.0, %num
  ret i1 %_0
}

to

define i1 @<!-- -->tgt(i32 %num, i32 %val) {
  %mask = add i32 %val, -1
  %tmp = and i32 %num, %mask
  %ret = icmp eq i32 %tmp, 0
  ret i1 %ret
}

For power-of-two val.

Observed in real life for following code

pub fn is_aligned(num: usize) -&gt; bool {
    num.next_multiple_of(1 &lt;&lt; 12) == num
}

which verifies that num is aligned to 4096.

Alive2 proof https://alive2.llvm.org/ce/z/QisECm

Full diff: https://github.com/llvm/llvm-project/pull/152851.diff

3 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp (+64)
(modified) llvm/lib/Transforms/InstCombine/InstCombineInternal.h (+1)
(modified) llvm/test/Transforms/InstCombine/icmp-add.ll (+90)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index cf94d28100488..722b03eb53f06 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -1320,6 +1320,67 @@ Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
   return nullptr;
 }
 
+// Fold icmp eq (num + (val - 1)) & -val, num
+//      to
+//      icmp eq 0, (and num, val - 1)
+// For value being power of two
+Instruction *InstCombinerImpl::foldNextMultiply(ICmpInst &Cmp) {
+  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
+  Value *Neg, *Add, *Num, *Mask, *Value;
+  CmpInst::Predicate Pred = Cmp.getPredicate();
+  const APInt *NegConst, *MaskConst, *NumCost;
+
+  if (Pred != ICmpInst::ICMP_EQ)
+    return nullptr;
+
+  // Match num + neg
+  if (!match(Op0, m_And(m_Value(Add), m_Value(Neg))))
+    return nullptr;
+
+  // Match num & mask
+  if (!match(Add, m_Add(m_Value(Num), m_Value(Mask))))
+    return nullptr;
+
+  // Check the constant case
+  if (match(Neg, m_APInt(NegConst)) && match(Mask, m_APInt(MaskConst))) {
+    // Mask + 1 should be a power-of-two
+    if (!(*MaskConst + 1).isPowerOf2())
+      return nullptr;
+
+    // Neg = -(Mask + 1)
+    if (*NegConst != -(*MaskConst + 1))
+      return nullptr;
+  } else {
+    // Match neg = sub 0, val
+    if (!match(Neg, m_Sub(m_Zero(), m_Value(Value))))
+      return nullptr;
+
+    // mask = %val - 1, which can be represented as sub %val, 1 or add %val, -1
+    if (!match(Mask, m_Add(m_Value(Value), m_AllOnes())) &&
+        !match(Mask, m_Sub(m_Value(Value), m_One())))
+      return nullptr;
+
+    // Value should be a known power-of-two.
+    if (!isKnownToBeAPowerOfTwo(Value, false, &Cmp))
+      return nullptr;
+  }
+
+  // Guard against weird special-case where Op1 gets optimized to constant. Leave it constant
+  // fonder.
+  if (match(Op1, m_APInt(NumCost)))
+    return nullptr;
+
+  if (!match(Op1, m_Value(Num)))
+    return nullptr;
+
+  // Create new icmp eq (num & (val - 1)), 0
+  auto NewAnd = Builder.CreateAnd(Num, Mask);
+  auto Zero = llvm::Constant::getNullValue(Num->getType());
+  auto ICmp = Builder.CreateICmp(CmpInst::ICMP_EQ, NewAnd, Zero);
+
+  return replaceInstUsesWith(Cmp, ICmp);
+}
+
 /// Fold icmp Pred X, C.
 /// TODO: This code structure does not make sense. The saturating add fold
 /// should be moved to some other helper and extended as noted below (it is also
@@ -7644,6 +7705,9 @@ Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
   if (Instruction *Res = foldICmpUsingKnownBits(I))
     return Res;
 
+  if (Instruction *Res = foldNextMultiply(I))
+    return Res;
+
   // Test if the ICmpInst instruction is used exclusively by a select as
   // part of a minimum or maximum operation. If so, refrain from doing
   // any other folding. This helps out other analyses which understand
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
index c67e27e5b3e7c..5f83cb1b9ae28 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
+++ b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
@@ -721,6 +721,7 @@ class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final
   Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp);
   Instruction *foldICmpWithDominatingICmp(ICmpInst &Cmp);
   Instruction *foldICmpWithConstant(ICmpInst &Cmp);
+  Instruction *foldNextMultiply(ICmpInst &Cmp);
   Instruction *foldICmpUsingBoolRange(ICmpInst &I);
   Instruction *foldICmpInstWithConstant(ICmpInst &Cmp);
   Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp);
diff --git a/llvm/test/Transforms/InstCombine/icmp-add.ll b/llvm/test/Transforms/InstCombine/icmp-add.ll
index 1a41c1f3e1045..698619ab8aad1 100644
--- a/llvm/test/Transforms/InstCombine/icmp-add.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-add.ll
@@ -3300,3 +3300,93 @@ entry:
   %cmp = icmp ult i32 %add, 253
   ret i1 %cmp
 }
+
+define i1 @val_is_aligend_sub(i32 %num, i32 %val) {
+; CHECK-LABEL: @val_is_aligend_sub(
+; CHECK-NEXT:    [[TMP1:%.*]] = tail call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 [[NUM:%.*]])
+; CHECK-NEXT:    [[POW:%.*]] = icmp eq i32 [[TMP1]], 1
+; CHECK-NEXT:    call void @llvm.assume(i1 [[POW]])
+; CHECK-NEXT:    [[NEG:%.*]] = add i32 [[NUM]], -1
+; CHECK-NEXT:    [[_2_SROA_0_0:%.*]] = and i32 [[NUM_BIASED:%.*]], [[NEG]]
+; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[_2_SROA_0_0]], 0
+; CHECK-NEXT:    ret i1 [[_0]]
+;
+  %1 = tail call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 %val)
+  %pow = icmp eq i32 %1, 1
+  call void @llvm.assume(i1 %pow)
+
+  %mask = sub i32 %val, 1
+  %neg = sub nsw i32 0, %val
+
+  %num.biased = add i32 %num, %mask
+  %_2.sroa.0.0 = and i32 %num.biased, %neg
+  %_0 = icmp eq i32 %_2.sroa.0.0, %num
+  ret i1 %_0
+}
+
+define i1 @val_is_aligend_add(i32 %num, i32 %val) {
+; CHECK-LABEL: @val_is_aligend_add(
+; CHECK-NEXT:    [[TMP1:%.*]] = tail call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 [[NUM:%.*]])
+; CHECK-NEXT:    [[POW:%.*]] = icmp eq i32 [[TMP1]], 1
+; CHECK-NEXT:    call void @llvm.assume(i1 [[POW]])
+; CHECK-NEXT:    [[NEG:%.*]] = add i32 [[NUM]], -1
+; CHECK-NEXT:    [[_2_SROA_0_0:%.*]] = and i32 [[NUM_BIASED:%.*]], [[NEG]]
+; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[_2_SROA_0_0]], 0
+; CHECK-NEXT:    ret i1 [[_0]]
+;
+  %1 = tail call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 %val)
+  %pow = icmp eq i32 %1, 1
+  call void @llvm.assume(i1 %pow)
+
+  %mask = add i32 %val, -1
+  %neg = sub nsw i32 0, %val
+
+  %num.biased = add i32 %num, %mask
+  %_2.sroa.0.0 = and i32 %num.biased, %neg
+  %_0 = icmp eq i32 %_2.sroa.0.0, %num
+  ret i1 %_0
+}
+
+define i1 @val_is_aligend_const_pow2(i32 %num) {
+; CHECK-LABEL: @val_is_aligend_const_pow2(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[NUM:%.*]], 4095
+; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[TMP1]], 0
+; CHECK-NEXT:    ret i1 [[_0]]
+;
+  %num.biased = add i32 %num, 4095
+  %_2.sroa.0.0 = and i32 %num.biased, -4096
+  %_0 = icmp eq i32 %_2.sroa.0.0, %num
+  ret i1 %_0
+}
+
+; Should not work for non-power-of-two cases
+define i1 @val_is_aligend_const_non-pow2(i32 %num) {
+; CHECK-LABEL: @val_is_aligend_const_non-pow2(
+; CHECK-NEXT:    [[NUM_BIASED:%.*]] = add i32 [[NUM:%.*]], 6
+; CHECK-NEXT:    [[_2_SROA_0_0:%.*]] = and i32 [[NUM_BIASED]], -7
+; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[_2_SROA_0_0]], [[NUM]]
+; CHECK-NEXT:    ret i1 [[_0]]
+;
+  %num.biased = add i32 %num, 6
+  %_2.sroa.0.0 = and i32 %num.biased, -7
+  %_0 = icmp eq i32 %_2.sroa.0.0, %num
+  ret i1 %_0
+}
+
+define i1 @val_is_aligend_non_pow(i32 %num, i32 %val) {
+; CHECK-LABEL: @val_is_aligend_non_pow(
+; CHECK-NEXT:    [[MASK:%.*]] = add i32 [[VAL:%.*]], -1
+; CHECK-NEXT:    [[NEG:%.*]] = sub nsw i32 0, [[VAL]]
+; CHECK-NEXT:    [[NUM_BIASED:%.*]] = add i32 [[NUM:%.*]], [[MASK]]
+; CHECK-NEXT:    [[_2_SROA_0_0:%.*]] = and i32 [[NUM_BIASED]], [[NEG]]
+; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[_2_SROA_0_0]], [[NUM]]
+; CHECK-NEXT:    ret i1 [[_0]]
+;
+  %mask = add i32 %val, -1
+  %neg = sub nsw i32 0, %val
+
+  %num.biased = add i32 %num, %mask
+  %_2.sroa.0.0 = and i32 %num.biased, %neg
+  %_0 = icmp eq i32 %_2.sroa.0.0, %num
+  ret i1 %_0
+}

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

pskrgag · 2025-08-09T19:16:26Z

@dtcxzyw Big thanks you for review! Addressed your comments

github-actions · 2025-08-09T19:17:06Z

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

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

pskrgag · 2025-08-11T18:05:59Z

CI looks unrelated

  ERROR: test_modulelist_deadlock (TestStatusline.TestStatusline.test_modulelist_deadlock)
     Regression test for a deadlock that occurs when the status line is enabled before connecting

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

llvm/test/Transforms/InstCombine/icmp-add.ll

dtcxzyw

LGTM w/ comments resolved.

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

pskrgag · 2025-08-13T20:05:19Z

CI complains about CodeGen/Thumb2/mve-vcvt-fixed-to-float.ll:479. IR looks unrelated to my change

define arm_aapcs_vfpcc <8 x half> @vcvt_i16_15(<8 x i16> %0) {
; CHECK-LABEL: vcvt_i16_15:
; CHECK:       @ %bb.0:
; CHECK-NEXT:    vmov.i16 q1, #0x200
; CHECK-NEXT:    vcvt.f16.s16 q0, q0
; CHECK-NEXT:    vmul.f16 q0, q0, q1
; CHECK-NEXT:    bx lr
  %2 = sitofp <8 x i16> %0 to <8 x half>
  %3 = fmul ninf <8 x half> %2, <half 0xH0200, half 0xH0200, half 0xH0200, half 0xH0200, half 0xH0200, half 0xH0200, half 0xH0200, half 0xH0200>
  ret <8 x half> %3
}

pskrgag · 2025-08-14T06:30:51Z

Should i wait for CI to become green, or I can merge now?

nikic · 2025-08-14T07:21:45Z

You can go ahead and merge.

fold "icmp eq (num + (val - 1)) & -val, num" to "icmp eq 0, (and num,…

d8b5d15

… val - 1)"

pskrgag requested a review from nikic as a code owner August 9, 2025 12:09

pskrgag requested review from dtcxzyw and removed request for nikic August 9, 2025 12:09

llvmbot added llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:transforms labels Aug 9, 2025

coding style...

93fd0c6

This was referenced Aug 9, 2025

Fuzz PR152851 dtcxzyw/llvm-mutation-based-fuzz-service#76

Closed

Fuzz PR152851 dtcxzyw/llvm-fuzz-service#113

Closed

dtcxzyw reviewed Aug 9, 2025

View reviewed changes

dtcxzyw requested a review from nikic August 9, 2025 18:23

address review comments

914dd1f

coding style....

fa89f88

dtcxzyw mentioned this pull request Aug 10, 2025

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

zyw-bot mentioned this pull request Aug 10, 2025

pre-commit: PR152851 dtcxzyw/llvm-opt-benchmark#2656

Closed

dtcxzyw reviewed Aug 10, 2025

View reviewed changes

review

25aec6b

dtcxzyw reviewed Aug 10, 2025

View reviewed changes

pskrgag added 2 commits August 10, 2025 21:12

learning llvm...

b66f128

use m_Defereded instead of m_Specific

db1316b

dtcxzyw reviewed Aug 11, 2025

View reviewed changes

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

dtcxzyw mentioned this pull request Aug 11, 2025

Fuzz PR152851 dtcxzyw/llvm-fuzz-service#114

Closed

zyw-bot mentioned this pull request Aug 11, 2025

pre-commit: PR152851 dtcxzyw/llvm-opt-benchmark#2657

Closed

don't open code bitwise not

c72f7bb

drop non-constant path as it happens to be non-profitable

0e9e8e7

nikic reviewed Aug 11, 2025

View reviewed changes

review

327aaed

dtcxzyw approved these changes Aug 13, 2025

View reviewed changes

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

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

dtcxzyw changed the title ~~[llvm] [InstCombine] fold "icmp eq (X + (V - 1)) & -V, X" to "icmp eq 0, (and X, V - 1)"~~ [llvm] [InstCombine] fold "icmp eq (X + (V - 1)) & -V, X" to "icmp eq (and X, V - 1), 0" Aug 13, 2025

pskrgag added 3 commits August 13, 2025 21:38

style fixes

ee4c1a2

Merge branch 'main' into foldIcmp

98e655c

swap operands in comment

b58e388

nikic approved these changes Aug 13, 2025

View reviewed changes

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

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

comment fixes

52ef6a6

dtcxzyw mentioned this pull request Aug 14, 2025

[InstCombine] Propagate invariant.load metadata across unpacked loads #152186

Merged

pskrgag merged commit 3014422 into llvm:main Aug 14, 2025
7 of 9 checks passed

[llvm] [InstCombine] fold "icmp eq (X + (V - 1)) & -V, X" to "icmp eq (and X, V - 1), 0" #152851

[llvm] [InstCombine] fold "icmp eq (X + (V - 1)) & -V, X" to "icmp eq (and X, V - 1), 0" #152851

Uh oh!

Conversation

pskrgag commented Aug 9, 2025

Uh oh!

llvmbot commented Aug 9, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

pskrgag commented Aug 9, 2025

Uh oh!

github-actions bot commented Aug 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

pskrgag commented Aug 11, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dtcxzyw left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

pskrgag commented Aug 13, 2025

Uh oh!

pskrgag commented Aug 14, 2025

Uh oh!

nikic commented Aug 14, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions bot commented Aug 9, 2025 •

edited

Loading