[InstCombine] Canonicalize icmp ult (add X, C2), C expressions

[antoniofrighetto]

icmp ult (add X, C2), C can be folded to icmp ne (and X, C), 2C, subject to C == -C2 and C2 being a power of 2.

Proofs: https://alive2.llvm.org/ce/z/P-VVmQ.

Fixes: #75613.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Antonio Frighetto (antoniofrighetto)

Changes

icmp ult (add X, C2), C can be folded to icmp ne (and X, C), 2C, subject to C == -C2 and C2 being a power of 2.

Proofs: https://alive2.llvm.org/ce/z/P-VVmQ.

Fixes: #75613.

---

Full diff: https://github.com/llvm/llvm-project/pull/95649.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp (+7)
• (modified) llvm/test/Transforms/InstCombine/icmp-add.ll (+65)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 34b0f8b860497..51dd3384c1fb5 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -3130,6 +3130,13 @@ Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
     return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
                         ConstantExpr::getNeg(cast<Constant>(Y)));
 
+  // X+C2 <u C -> (X & C) == 2C
+  //   iff C == -(C2)
+  //       C2 is a power of 2
+  if (Pred == ICmpInst::ICMP_ULT && C2->isPowerOf2() && (C == -(*C2)))
+    return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, C),
+                        ConstantInt::get(Ty, C * 2));
+
   // X+C >u C2 -> (X & ~C2) != C
   //   iff C & C2 == 0
   //       C2+1 is a power of 2
diff --git a/llvm/test/Transforms/InstCombine/icmp-add.ll b/llvm/test/Transforms/InstCombine/icmp-add.ll
index 6b4e5a5372c52..f5e8a58580ce8 100644
--- a/llvm/test/Transforms/InstCombine/icmp-add.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-add.ll
@@ -3023,4 +3023,69 @@ define i1 @icmp_addnuw_nonzero_fail_multiuse(i32 %x, i32 %y) {
   ret i1 %c
 }
 
+define i1 @ult_add_C2_pow2_C_neg(i8 %x) {
+; CHECK-LABEL: @ult_add_C2_pow2_C_neg(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i8 [[X:%.*]], -32
+; CHECK-NEXT:    [[C:%.*]] = icmp ne i8 [[TMP1]], -64
+; CHECK-NEXT:    ret i1 [[C]]
+;
+  %i = add i8 %x, 32
+  %c = icmp ult i8 %i, -32
+  ret i1 %c
+}
+
+define i1 @ult_add_nsw_C2_pow2_C_neg(i8 %x) {
+; CHECK-LABEL: @ult_add_nsw_C2_pow2_C_neg(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i8 [[X:%.*]], -32
+; CHECK-NEXT:    [[C:%.*]] = icmp ne i8 [[TMP1]], -64
+; CHECK-NEXT:    ret i1 [[C]]
+;
+  %i = add nsw i8 %x, 32
+  %c = icmp ult i8 %i, -32
+  ret i1 %c
+}
+
+define i1 @ult_add_nuw_nsw_C2_pow2_C_neg(i8 %x) {
+; CHECK-LABEL: @ult_add_nuw_nsw_C2_pow2_C_neg(
+; CHECK-NEXT:    [[C:%.*]] = icmp ult i8 [[X:%.*]], -64
+; CHECK-NEXT:    ret i1 [[C]]
+;
+  %i = add nuw nsw i8 %x, 32
+  %c = icmp ult i8 %i, -32
+  ret i1 %c
+}
+
+define i1 @ult_add_C2_neg_C_pow2(i8 %x) {
+; CHECK-LABEL: @ult_add_C2_neg_C_pow2(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i8 [[X:%.*]], -32
+; CHECK-NEXT:    [[C:%.*]] = icmp eq i8 [[TMP1]], 32
+; CHECK-NEXT:    ret i1 [[C]]
+;
+  %i = add i8 %x, -32
+  %c = icmp ult i8 %i, 32
+  ret i1 %c
+}
+
+define <2 x i1> @ult_add_C2_pow2_C_neg_vec(<2 x i8> %x) {
+; CHECK-LABEL: @ult_add_C2_pow2_C_neg_vec(
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i8> [[X:%.*]], <i8 -32, i8 -32>
+; CHECK-NEXT:    [[C:%.*]] = icmp ne <2 x i8> [[TMP1]], <i8 -64, i8 -64>
+; CHECK-NEXT:    ret <2 x i1> [[C]]
+;
+  %i = add <2 x i8> %x, <i8 32, i8 32>
+  %c = icmp ult <2 x i8> %i, <i8 -32, i8 -32>
+  ret <2 x i1> %c
+}
+
+define i1 @uge_add_C2_pow2_C_neg(i8 %x) {
+; CHECK-LABEL: @uge_add_C2_pow2_C_neg(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i8 [[X:%.*]], -32
+; CHECK-NEXT:    [[C:%.*]] = icmp eq i8 [[TMP1]], -64
+; CHECK-NEXT:    ret i1 [[C]]
+;
+  %i = add i8 %x, 32
+  %c = icmp uge i8 %i, -32
+  ret i1 %c
+}
+
 declare void @llvm.assume(i1)

[goldsteinn]

NB regarding the proofs, you can just do %C2_cnt = call i8 @llvm.ctpop.i8(i8 %C2); %C2_is_pow2 = icmp eq i8 %C2_cnt, 1 for power of 2 lemma.

[goldsteinn]

LGTM, but can you look into CI failures.

[nikic]

The key question here is whether this is the right canonicalization direction -- do we prefer a mask check over a range check? I guess if @dtcxzyw's tests don't show regressions from this, then this direction is fine.

[goldsteinn]

The key question here is whether this is the right canonicalization direction -- do we prefer a mask check over a range check? I guess if @dtcxzyw's tests don't show regressions from this, then this direction is fine.

W.o nuw/nsw would def think mask check is better. W/ the flags could imagine the other way around though.

[dtcxzyw]

The key question here is whether this is the right canonicalization direction -- do we prefer a mask check over a range check? I guess if @dtcxzyw's tests don't show regressions from this, then this direction is fine.

Done.

[antoniofrighetto]

Added a multi-use, thanks. CI failures look unrelated.

[nikic]

LGTM

[dtcxzyw]

LGTM.