[llvm] [InstCombine] Simplify conditional round-up to power-of-2 alignment (PR #184989)

via llvm-commits llvm-commits at lists.llvm.org
Fri Mar 6 03:11:00 PST 2026


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Antonio Frighetto (antoniofrighetto)

<details>
<summary>Changes</summary>

The pattern `(X + (X % P == 0 ? 0 : P)) & -P`, that may arise when computing `div_ceil(V, P) * P`, can be canonicalized to `(X + P-1) & -P`, with P being a power of 2.

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

Fixes: https://github.com/llvm/llvm-project/issues/184762.

---
Full diff: https://github.com/llvm/llvm-project/pull/184989.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+41) 
- (added) llvm/test/Transforms/InstCombine/simple-round-up-pow2-alignment.ll (+229) 


``````````diff
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index 073f094639fa0..a21e0a1b83421 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -2382,6 +2382,44 @@ static Value *simplifyAndOrWithOpReplaced(Value *V, Value *Op, Value *RepOp,
   return IC.Builder.CreateBinOp(I->getOpcode(), NewOp0, NewOp1);
 }
 
+// The pattern div_ceil(X, P) * P, where P is a power of 2, lowers to a
+// conditional round-up of the following kind: (X + select(X % Pow2 == 0, 0,
+// Pow2)) & -Pow2. This may be simplified to (X + (Pow2-1)) & -Pow2.
+static Instruction *
+foldRoundUpToPow2Alignment(BinaryOperator &I,
+                           InstCombiner::BuilderTy &Builder) {
+  const APInt *NegP;
+  Value *Add;
+  if (!match(&I, m_And(m_Value(Add), m_APInt(NegP))) ||
+      !NegP->isNegatedPowerOf2())
+    return nullptr;
+
+  Value *X, *Cond;
+  const APInt *T, *F;
+  if (!match(Add, m_c_Add(m_Value(X),
+                          m_Select(m_Value(Cond), m_APInt(T), m_APInt(F)))) ||
+      !Add->hasOneUse())
+    return nullptr;
+
+  CmpPredicate Pred;
+  const APInt &Pow2 = -*NegP;
+  const APInt &Mask = Pow2 - 1;
+  // icmp ne should have already been canonicalized to the eq form for this
+  // pattern.
+  if (!match(Cond, m_ICmp(Pred, m_And(m_Specific(X), m_SpecificInt(Mask)),
+                          m_Zero())) ||
+      Pred != ICmpInst::ICMP_EQ)
+    return nullptr;
+
+  // Ensure the true arm of the select is zero, and the false one is the Pow2.
+  if (!T->isZero() || *F != Pow2)
+    return nullptr;
+
+  Type *Ty = I.getType();
+  Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(Ty, Mask));
+  return BinaryOperator::CreateAnd(NewAdd, ConstantInt::get(Ty, *NegP));
+}
+
 /// Reassociate and/or expressions to see if we can fold the inner and/or ops.
 /// TODO: Make this recursive; it's a little tricky because an arbitrary
 /// number of and/or instructions might have to be created.
@@ -2898,6 +2936,9 @@ Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {
                                       /*SimplifyOnly*/ false, *this))
     return BinaryOperator::CreateAnd(Op0, V);
 
+  if (Instruction *Res = foldRoundUpToPow2Alignment(I, Builder))
+    return Res;
+
   return nullptr;
 }
 
diff --git a/llvm/test/Transforms/InstCombine/simple-round-up-pow2-alignment.ll b/llvm/test/Transforms/InstCombine/simple-round-up-pow2-alignment.ll
new file mode 100644
index 0000000000000..66dd468800197
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/simple-round-up-pow2-alignment.ll
@@ -0,0 +1,229 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+; Simple round up to next multiple of power-of-2 pattern that may arise
+; from div_ceil(X, P) * P.
+
+define i32 @round_up_pow2_2(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_2(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[AND:%.*]] = shl i32 [[X]], 1
+; CHECK-NEXT:    [[SEL:%.*]] = and i32 [[AND]], 2
+; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[SEL]], [[X]]
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[ADD]], -2
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 1
+  %cond = icmp eq i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 2
+  %add = add i32 %sel, %x
+  %res = and i32 %add, -2
+  ret i32 %res
+}
+
+define i32 @round_up_pow2_64(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_64(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[X]], 63
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[TMP1]], -64
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 63
+  %cond = icmp eq i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 64
+  %add = add i32 %sel, %x
+  %res = and i32 %add, -64
+  ret i32 %res
+}
+
+define i32 @round_up_pow2_add_commuted(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_add_commuted(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[X]], 63
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[TMP1]], -64
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 63
+  %cond = icmp eq i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 64
+  %add = add i32 %x, %sel
+  %res = and i32 %add, -64
+  ret i32 %res
+}
+
+define i32 @round_up_pow2_icmp_ne(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_icmp_ne(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[X]], 63
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[TMP1]], -64
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 63
+  %cond = icmp ne i32 %and, 0
+  %sel = select i1 %cond, i32 64, i32 0
+  %add = add i32 %sel, %x
+  %res = and i32 %add, -64
+  ret i32 %res
+}
+
+define <2 x i32> @round_up_pow2_vector(<2 x i32> %x) {
+; CHECK-LABEL: define <2 x i32> @round_up_pow2_vector(
+; CHECK-SAME: <2 x i32> [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = add <2 x i32> [[X]], splat (i32 63)
+; CHECK-NEXT:    [[RES:%.*]] = and <2 x i32> [[TMP1]], splat (i32 -64)
+; CHECK-NEXT:    ret <2 x i32> [[RES]]
+;
+  %and = and <2 x i32> %x, <i32 63, i32 63>
+  %cond = icmp eq <2 x i32> %and, zeroinitializer
+  %sel = select <2 x i1> %cond, <2 x i32> zeroinitializer, <2 x i32> <i32 64, i32 64>
+  %add = add <2 x i32> %sel, %x
+  %res = and <2 x i32> %add, <i32 -64, i32 -64>
+  ret <2 x i32> %res
+}
+
+define <2 x i32> @round_up_pow2_vector_icmp_ne(<2 x i32> %x) {
+; CHECK-LABEL: define <2 x i32> @round_up_pow2_vector_icmp_ne(
+; CHECK-SAME: <2 x i32> [[X:%.*]]) {
+; CHECK-NEXT:    [[ADD:%.*]] = add <2 x i32> [[X]], splat (i32 63)
+; CHECK-NEXT:    [[RES:%.*]] = and <2 x i32> [[ADD]], splat (i32 -64)
+; CHECK-NEXT:    ret <2 x i32> [[RES]]
+;
+  %and = and <2 x i32> %x, <i32 63, i32 63>
+  %cond = icmp ne <2 x i32> %and, zeroinitializer
+  %sel = select <2 x i1> %cond, <2 x i32> <i32 64, i32 64>, <2 x i32> zeroinitializer
+  %add = add <2 x i32> %sel, %x
+  %res = and <2 x i32> %add, <i32 -64, i32 -64>
+  ret <2 x i32> %res
+}
+
+define i32 @round_up_pow2_multiuse_and(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_multiuse_and(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[X]], 63
+; CHECK-NEXT:    call void @use(i32 [[AND]])
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[X]], 63
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[TMP1]], -64
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 63
+  call void @use(i32 %and)
+  %cond = icmp eq i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 64
+  %add = add i32 %sel, %x
+  %res = and i32 %add, -64
+  ret i32 %res
+}
+
+; Negative tests.
+
+define i32 @round_up_pow2_multiuse_add(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_multiuse_add(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[X]], 63
+; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[AND]], 0
+; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[COND]], i32 0, i32 64
+; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[SEL]], [[X]]
+; CHECK-NEXT:    call void @use(i32 [[ADD]])
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[ADD]], -64
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 63
+  %cond = icmp eq i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 64
+  %add = add i32 %sel, %x
+  call void @use(i32 %add)
+  %res = and i32 %add, -64
+  ret i32 %res
+}
+
+define i32 @round_up_pow2_wrong_mask(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_wrong_mask(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[X]], 31
+; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[AND]], 0
+; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[COND]], i32 0, i32 64
+; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[SEL]], [[X]]
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[ADD]], -64
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 31
+  %cond = icmp eq i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 64
+  %add = add i32 %sel, %x
+  %res = and i32 %add, -64
+  ret i32 %res
+}
+
+define i32 @round_up_pow2_wrong_negated_pow2(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_wrong_negated_pow2(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[X]], 63
+; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[AND]], 0
+; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[COND]], i32 0, i32 64
+; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[SEL]], [[X]]
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[ADD]], -63
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 63
+  %cond = icmp eq i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 64
+  %add = add i32 %sel, %x
+  %res = and i32 %add, -63
+  ret i32 %res
+}
+
+define i32 @round_up_pow2_wrong_select_arms(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_wrong_select_arms(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[X]], 63
+; CHECK-NEXT:    [[COND_NOT:%.*]] = icmp eq i32 [[AND]], 0
+; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[COND_NOT]], i32 64, i32 0
+; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[SEL]], [[X]]
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[ADD]], -64
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 63
+  %cond = icmp ne i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 64
+  %add = add i32 %sel, %x
+  %res = and i32 %add, -64
+  ret i32 %res
+}
+
+define i32 @round_up_pow2_wrong_predicate(i32 %x) {
+; CHECK-LABEL: define i32 @round_up_pow2_wrong_predicate(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[X]], 63
+; CHECK-NEXT:    [[COND_NOT:%.*]] = icmp eq i32 [[AND]], 0
+; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[COND_NOT]], i32 64, i32 0
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[SEL]], [[X]]
+; CHECK-NEXT:    [[RES:%.*]] = and i32 [[TMP1]], -64
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %x, 63
+  %cond = icmp sgt i32 %and, 0
+  %sel = select i1 %cond, i32 0, i32 64
+  %add = add i32 %sel, %x
+  %res = and i32 %add, -64
+  ret i32 %res
+}
+
+define <2 x i32> @round_up_pow2_vector_poison_lane(<2 x i32> %x) {
+; CHECK-LABEL: define <2 x i32> @round_up_pow2_vector_poison_lane(
+; CHECK-SAME: <2 x i32> [[X:%.*]]) {
+; CHECK-NEXT:    [[AND:%.*]] = and <2 x i32> [[X]], splat (i32 63)
+; CHECK-NEXT:    [[COND:%.*]] = icmp eq <2 x i32> [[AND]], zeroinitializer
+; CHECK-NEXT:    [[SEL:%.*]] = select <2 x i1> [[COND]], <2 x i32> zeroinitializer, <2 x i32> <i32 64, i32 poison>
+; CHECK-NEXT:    [[ADD:%.*]] = add <2 x i32> [[SEL]], [[X]]
+; CHECK-NEXT:    [[RES:%.*]] = and <2 x i32> [[ADD]], splat (i32 -64)
+; CHECK-NEXT:    ret <2 x i32> [[RES]]
+;
+  %and = and <2 x i32> %x, <i32 63, i32 63>
+  %cond = icmp eq <2 x i32> %and, zeroinitializer
+  %sel = select <2 x i1> %cond, <2 x i32> zeroinitializer, <2 x i32> <i32 64, i32 poison>
+  %add = add <2 x i32> %sel, %x
+  %res = and <2 x i32> %add, <i32 -64, i32 -64>
+  ret <2 x i32> %res
+}
+
+declare void @use(i32)

``````````

</details>


https://github.com/llvm/llvm-project/pull/184989


More information about the llvm-commits mailing list