[InstCombine] Canonicalize (X +/- Y) & Y into ~X & Y when Y is a power of 2

[dtcxzyw]

This patch canonicalizes the pattern (X +/- Y) & Y into ~X & Y when Y is a power of 2 or zero.
It will reduce the patterns to match in #67836 and exploit more optimization opportunities.
Alive2: https://alive2.llvm.org/ce/z/LBpvRF

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Changes

This patch canonicalizes the pattern (X +/- Y) & Y into ~X & Y when Y is a power of 2 or zero.
It will reduce the patterns to match in #67836 and exploit more optimization opportunities.
Alive2: https://alive2.llvm.org/ce/z/LBpvRF

---

Full diff: https://github.com/llvm/llvm-project/pull/67915.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+8)
• (modified) llvm/test/Transforms/InstCombine/and.ll (+40-4)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index cbdab3e9c5fb91d..4322cc96f5a2b6c 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -2250,6 +2250,14 @@ Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {
     return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), Y);
   }
 
+  // Canonicalize:
+  // (X +/- Y) & Y --> ~X & Y when Y is a power of 2.
+  if (match(&I, m_c_And(m_Value(Y), m_OneUse(m_CombineOr(
+                                        m_c_Add(m_Value(X), m_Deferred(Y)),
+                                        m_Sub(m_Value(X), m_Deferred(Y)))))) &&
+      isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, /*Depth*/ 0, &I))
+    return BinaryOperator::CreateAnd(Builder.CreateNot(X), Y);
+
   const APInt *C;
   if (match(Op1, m_APInt(C))) {
     const APInt *XorC;
diff --git a/llvm/test/Transforms/InstCombine/and.ll b/llvm/test/Transforms/InstCombine/and.ll
index 90f027010e2aea6..eb39ff9014ff3a4 100644
--- a/llvm/test/Transforms/InstCombine/and.ll
+++ b/llvm/test/Transforms/InstCombine/and.ll
@@ -1595,8 +1595,8 @@ define <2 x i8> @flip_masked_bit_uniform(<2 x i8> %A) {
 
 define <2 x i8> @flip_masked_bit_undef(<2 x i8> %A) {
 ; CHECK-LABEL: @flip_masked_bit_undef(
-; CHECK-NEXT:    [[B:%.*]] = add <2 x i8> [[A:%.*]], <i8 16, i8 undef>
-; CHECK-NEXT:    [[C:%.*]] = and <2 x i8> [[B]], <i8 16, i8 undef>
+; CHECK-NEXT:    [[TMP1:%.*]] = xor <2 x i8> [[A:%.*]], <i8 -1, i8 -1>
+; CHECK-NEXT:    [[C:%.*]] = and <2 x i8> [[TMP1]], <i8 16, i8 undef>
 ; CHECK-NEXT:    ret <2 x i8> [[C]]
 ;
   %B = add <2 x i8> %A, <i8 16, i8 undef>
@@ -1606,8 +1606,8 @@ define <2 x i8> @flip_masked_bit_undef(<2 x i8> %A) {
 
 define <2 x i8> @flip_masked_bit_nonuniform(<2 x i8> %A) {
 ; CHECK-LABEL: @flip_masked_bit_nonuniform(
-; CHECK-NEXT:    [[B:%.*]] = add <2 x i8> [[A:%.*]], <i8 16, i8 4>
-; CHECK-NEXT:    [[C:%.*]] = and <2 x i8> [[B]], <i8 16, i8 4>
+; CHECK-NEXT:    [[TMP1:%.*]] = xor <2 x i8> [[A:%.*]], <i8 -1, i8 -1>
+; CHECK-NEXT:    [[C:%.*]] = and <2 x i8> [[TMP1]], <i8 16, i8 4>
 ; CHECK-NEXT:    ret <2 x i8> [[C]]
 ;
   %B = add <2 x i8> %A, <i8 16, i8 4>
@@ -2546,3 +2546,39 @@ define i32 @and_zext_eq_zero(i32 %A, i32 %C)  {
   %5 = and i32 %2, %4
   ret i32 %5
 }
+
+define i32 @canonicalize_and_add_power2_or_zero(i32 %x, i32 %y) {
+; CHECK-LABEL: @canonicalize_and_add_power2_or_zero(
+; CHECK-NEXT:    [[NY:%.*]] = sub i32 0, [[Y:%.*]]
+; CHECK-NEXT:    [[P2:%.*]] = and i32 [[NY]], [[Y]]
+; CHECK-NEXT:    call void @use32(i32 [[P2]])
+; CHECK-NEXT:    [[TMP1:%.*]] = xor i32 [[X:%.*]], -1
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[P2]], [[TMP1]]
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %ny = sub i32 0, %y
+  %p2 = and i32 %y, %ny
+  call void @use32(i32 %p2) ; keep p2
+
+  %val = add i32 %x, %p2
+  %and = and i32 %val, %p2
+  ret i32 %and
+}
+
+define i32 @canonicalize_and_sub_power2_or_zero(i32 %x, i32 %y) {
+; CHECK-LABEL: @canonicalize_and_sub_power2_or_zero(
+; CHECK-NEXT:    [[NY:%.*]] = sub i32 0, [[Y:%.*]]
+; CHECK-NEXT:    [[P2:%.*]] = and i32 [[NY]], [[Y]]
+; CHECK-NEXT:    call void @use32(i32 [[P2]])
+; CHECK-NEXT:    [[TMP1:%.*]] = xor i32 [[X:%.*]], -1
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[P2]], [[TMP1]]
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %ny = sub i32 0, %y
+  %p2 = and i32 %y, %ny
+  call void @use32(i32 %p2) ; keep p2
+
+  %val = sub i32 %x, %p2
+  %and = and i32 %val, %p2
+  ret i32 %and
+}

[goldsteinn]

LGTM.

[nikic]

Seems like some test coverage is missing here?

• Commuted add
• Commuted sub (negative test)
• Commuted and
• Multi-use add (negative test)

[nikic]

Can you please pre-commit the new tests as well (and double-check that there is no unexpected operand order canonicalization while doing that)?

[dtcxzyw]

Ping.

[nikic]

LGTM