[InstCombine] Change (add x, c) to (xor x, c)

[Peter9606]

Change (add x, c) to (xor x, c) iff c is constant and c equals the top bit of the demanded bits.
Alive2: https://alive2.llvm.org/ce/z/DKmkwF

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Fujun Han (Peter9606)

Changes

Simplify a complex OR to XOR

Following pattern:
  (or (and x, half_c1), c3), (and x, c2))
  IFF
    c1, c2, c3 is constant
    c1 is pow2
    c2 < c1
    c3 == (c1 - 1) ^ c2
    half_c1 = (lshr c1, 1)
    (c1 >> 1) & c3 == (c1 >> 1)
    x is known to be less than c1
can be simplified into:
  (xor x, half_c)

Proof: https://alive2.llvm.org/ce/z/Lfax3w

---

Full diff: https://github.com/llvm/llvm-project/pull/75129.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+21-1)
• (added) llvm/test/Transforms/InstCombine/or-and-add-constant-constant-and-constant.ll (+36)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index 6002f599ca71ab..67ec4a1608169a 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -3343,6 +3343,27 @@ Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
   if (Instruction *X = foldComplexAndOrPatterns(I, Builder))
     return X;
 
+  // Change (or (and (add x, half_c1), c3), (and x, c2)) to (xor x, half_c1),
+  // iff c1, c2, c3 is constant, and half_c1 = (lshr c1, 1),  and c1 is pow2,
+  // and c2 < c1, and c3 == (c1 - 1) ^ c2, and (c1 >> 1) & c3 == (c1 >> 1) and x
+  // is known to be less than c1.
+  Type *Ty = I.getType();
+  if (Ty->isIntegerTy()) {
+    Value *X = nullptr;
+    const APInt *HalfC1 = nullptr, *C2 = nullptr, *C3 = nullptr;
+    if (match(&I,
+              m_c_Or(m_c_And(m_c_Add(m_Value(X), m_APInt(HalfC1)), m_APInt(C3)),
+                     m_c_And(m_Value(X), m_APInt(C2))))) {
+      const APInt C1 = HalfC1->shl(1);
+      KnownBits KnownX = computeKnownBits(X, 0, nullptr);
+      if (C1.isPowerOf2() && C2->ult(C1) && (*C3 == (*C2 ^ (C1 - 1))) &&
+          ((*HalfC1 & *C3) == *HalfC1) && KnownX.getMaxValue().ult(C1)) {
+        Value *Xor = Builder.CreateXor(X, ConstantInt::get(Ty, *HalfC1));
+        return replaceInstUsesWith(I, Xor);
+      }
+    }
+  }
+
   // (A&B)|(A&C) -> A&(B|C) etc
   if (Value *V = foldUsingDistributiveLaws(I))
     return replaceInstUsesWith(I, V);
@@ -3351,7 +3372,6 @@ Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
     return replaceInstUsesWith(I, V);
 
   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
-  Type *Ty = I.getType();
   if (Ty->isIntOrIntVectorTy(1)) {
     if (auto *SI0 = dyn_cast<SelectInst>(Op0)) {
       if (auto *R =
diff --git a/llvm/test/Transforms/InstCombine/or-and-add-constant-constant-and-constant.ll b/llvm/test/Transforms/InstCombine/or-and-add-constant-constant-and-constant.ll
new file mode 100644
index 00000000000000..3e48f3c3f331b9
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/or-and-add-constant-constant-and-constant.ll
@@ -0,0 +1,36 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+; Pattern:
+;   (or (and (add x, half_c1), c3), (and x, c2))
+; IFF:
+;   c1, c2, c3 is constant
+;   c1 is pow2
+;   c2 < c1
+;   c3 == (c1 - 1) ^ c2
+;   half_c1 == (lshr c1, 1)
+;   (c1 >> 1) & c3 == (c1 >> 1)
+;   x is known to be less than c1
+; Could be transformed into:
+;   (xor x, half_c1)
+; Proof: https://alive2.llvm.org/ce/z/Lfax3w
+
+define i16 @or_and_add_and() {
+; CHECK-LABEL: @or_and_add_and(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[X:%.*]] = call i16 @dummy(), !range [[RNG0:![0-9]+]]
+; CHECK-NEXT:    [[OR:%.*]] = xor i16 [[X]], 32
+; CHECK-NEXT:    ret i16 [[OR]]
+;
+entry:
+  %x = call i16 @dummy(), !range !0
+  %add = add i16 32, %x
+  %and1 = and i16 %add, 48
+  %and2 = and i16 %x, 15
+  %or = or i16 %and1, %and2
+  ret i16 %or
+}
+
+declare i16 @dummy()
+
+!0 = !{i16 0, i16 64}

[dtcxzyw]

This pattern looks complex. Does it come from a real-world application?

[Peter9606]

This pattern looks complex. Does it come from a real-world application?

Yes, it does.

[bcl5980]

Precommit test First, then second commit is code change + test diff.
It looks your test is the second commit.

[Peter9606]

Precommit test First, then second commit is code change + test diff. It looks your test is the second commit.

You are right, fixed! Thanks!

[bcl5980]

We may need to rewrite the conditions.
https://alive2.llvm.org/ce/z/-ct9Wt

Not all cases can be catched for now.

[topperc]

I think (and (add x, 32), 48) can be replaced with (and (xor x, 32), 48). Is that optimization alone enough to trigger the rest?

[Peter9606]

I think (and (add x, 32), 48) can be replaced with (and (xor x, 32), 48). Is that optimization alone enough to trigger the rest?

Just did a quick verify, your sugguestion works. At least for the case pointed by @bcl5980 in comment, it can be simplified to XOR now.

[Peter9606]

I think (and (add x, 32), 48) can be replaced with (and (xor x, 32), 48). Is that optimization alone enough to trigger the rest?

@topperc If I understand correctlly, your suggestion can be formalized as https://alive2.llvm.org/ce/z/aoJPsn, right?

[nikic]

@Peter9606 We'd probably approach that as a demanded bits transform (

llvm-project/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

Line 492 in d293a35

case Instruction::Add: {

). Convert X + C to X ^ C if C is equal to the top bit of the DemandedMask.

[Peter9606]

@Peter9606 We'd probably approach that as a demanded bits transform (

llvm-project/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

Line 492 in d293a35

case Instruction::Add: {

). Convert X + C to X ^ C if C is equal to the top bit of the DemandedMask.

Thanks for your detail guidance. @nikic
PR is updated!

[nikic]

Looks reasonable to me.

[bcl5980]

Add a vector case?

[nikic]

LGTM

[dtcxzyw]

LGTM. I will merge this PR after the CI is complete.

[jplehr]

I see errors in our buildbot (https://lab.llvm.org/buildbot/#/builders/193/builds/43753) after this patch landed.

Reverting this patch locally resolves the errors.
If you need additional info / help for reproducing, please let me know.

[nikic]

Should be fixed by 9100228.