[llvm] [SCEV] Prove umin(Ops) u< RHS by checking individual operands (PR #207729)
Aleksandr Popov via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 8 07:07:27 PDT 2026
https://github.com/aleks-tmb updated https://github.com/llvm/llvm-project/pull/207729
>From 7226fd63d47e4b8fe20e2b9de33278c46bef5811 Mon Sep 17 00:00:00 2001
From: Aleksandr Popov <apopov at azul.com>
Date: Sun, 5 Jul 2026 20:22:34 +0000
Subject: [PATCH 1/3] [SCEV] Prove umin(Ops) u< RHS by checking individual
operands
Teach IsKnownPredicateViaMinOrMax to handle ICMP_ULT/UGT for
SCEVUMinExpr. Since umin(Ops) u<= each Op by definition, proving
Op u< RHS for any single operand is sufficient to establish
umin(Ops) u< RHS by transitivity.
This unblocks loop vectorization for early-exit loops whose trip count
is a umin of a buffer length and an iteration bound (e.g. searching a
bounded slice for a sentinel value).
Fixes https://github.com/llvm/llvm-project/issues/196935
---
llvm/lib/Analysis/ScalarEvolution.cpp | 24 +++++++++++
.../early-exit-umin-trip-count.ll | 43 ++++++++++++++++---
2 files changed, 61 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Analysis/ScalarEvolution.cpp b/llvm/lib/Analysis/ScalarEvolution.cpp
index c0cdce982e623..a168e4e8e1c12 100644
--- a/llvm/lib/Analysis/ScalarEvolution.cpp
+++ b/llvm/lib/Analysis/ScalarEvolution.cpp
@@ -12924,6 +12924,30 @@ static bool IsKnownPredicateViaMinOrMax(ScalarEvolution &SE, CmpPredicate Pred,
IsMinMaxConsistingOf<SCEVUMinExpr>(LHS, RHS) ||
// A <= max(A, ...)
IsMinMaxConsistingOf<SCEVUMaxExpr>(RHS, LHS);
+
+ case ICmpInst::ICMP_UGT:
+ std::swap(LHS, RHS);
+ [[fallthrough]];
+ case ICmpInst::ICMP_ULT:
+ // umin(Ops) u<= each Op, so proving Op u< RHS for any Op proves
+ // umin(Ops) u< RHS.
+ //
+ // Use computeConstantDifference instead of the more powerful
+ // isKnownPredicate to keep this check cheap: isKnownPredicateViaMinOrMax
+ // is called from isKnownViaNonRecursiveReasoning, so recursing into
+ // the full predicate prover would be expensive.
+ if (const auto *Min = dyn_cast<SCEVUMinExpr>(LHS)) {
+ for (SCEVUse Op : Min->operands()) {
+ std::optional<APInt> Diff = SE.computeConstantDifference(RHS, Op);
+ // When Op and RHS share a common base differing by a
+ // constant offset D (RHS - Op = D), Op u< RHS holds iff D != 0 and
+ // RHS >= D (unsigned), i.e. the subtraction doesn't underflow.
+ if (Diff && !Diff->isZero() &&
+ SE.getUnsignedRangeMin(RHS).uge(*Diff))
+ return true;
+ }
+ }
+ return false;
}
llvm_unreachable("covered switch fell through?!");
diff --git a/llvm/test/Transforms/LoopVectorize/early-exit-umin-trip-count.ll b/llvm/test/Transforms/LoopVectorize/early-exit-umin-trip-count.ll
index 36936ec9832a4..762a118130a46 100644
--- a/llvm/test/Transforms/LoopVectorize/early-exit-umin-trip-count.ll
+++ b/llvm/test/Transforms/LoopVectorize/early-exit-umin-trip-count.ll
@@ -1,7 +1,7 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -S -passes=loop-vectorize %s | FileCheck %s
-; TODO: The loop should be vectorized. The vectorizer needs to prove
+; The loop should be vectorized. The vectorizer needs to prove
; (A umin B) u<= A to establish memory safety from the dereferenceable
; assumption.
;
@@ -22,17 +22,42 @@ define ptr @main(i32 %length, ptr %ptr, i64 %0) #0 {
; CHECK-NEXT: [[NULL_CHECK:%.*]] = icmp eq i32 [[LENGTH]], 0
; CHECK-NEXT: br i1 [[NULL_CHECK]], label %[[DEOPT]], label %[[SCALAR_PH:.*]]
; CHECK: [[SCALAR_PH]]:
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[EXIT]], 4
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH1:.*]], label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[EXIT]], 4
+; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[EXIT]], [[N_MOD_VF]]
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
-; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ], [ 0, %[[SCALAR_PH]] ]
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY_INTERIM:.*]] ]
; CHECK-NEXT: [[ELEMENT_GEP:%.*]] = getelementptr i8, ptr [[PTR]], i64 [[IV]]
-; CHECK-NEXT: [[ELEMENT:%.*]] = load i8, ptr [[ELEMENT_GEP]], align 1
+; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i8>, ptr [[ELEMENT_GEP]], align 1
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq <4 x i8> [[WIDE_LOAD]], zeroinitializer
+; CHECK-NEXT: [[TMP3:%.*]] = freeze <4 x i1> [[TMP2]]
+; CHECK-NEXT: [[TMP4:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP3]])
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[IV]], 4
+; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP4]], label %[[VECTOR_EARLY_EXIT:.*]], label %[[VECTOR_BODY_INTERIM]]
+; CHECK: [[VECTOR_BODY_INTERIM]]:
+; CHECK-NEXT: br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[LOOP]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[EXIT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label %[[DEOPT_LOOPEXIT:.*]], label %[[SCALAR_PH1]]
+; CHECK: [[VECTOR_EARLY_EXIT]]:
+; CHECK-NEXT: br label %[[FOUND:.*]]
+; CHECK: [[SCALAR_PH1]]:
+; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[SCALAR_PH]] ]
+; CHECK-NEXT: br label %[[LOOP1:.*]]
+; CHECK: [[LOOP1]]:
+; CHECK-NEXT: [[IV1:%.*]] = phi i64 [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ], [ [[BC_RESUME_VAL]], %[[SCALAR_PH1]] ]
+; CHECK-NEXT: [[ELEMENT_GEP1:%.*]] = getelementptr i8, ptr [[PTR]], i64 [[IV1]]
+; CHECK-NEXT: [[ELEMENT:%.*]] = load i8, ptr [[ELEMENT_GEP1]], align 1
; CHECK-NEXT: [[FOUND_CHECK:%.*]] = icmp eq i8 [[ELEMENT]], 0
-; CHECK-NEXT: br i1 [[FOUND_CHECK]], label %[[FOUND:.*]], label %[[LATCH]]
+; CHECK-NEXT: br i1 [[FOUND_CHECK]], label %[[FOUND]], label %[[LATCH]]
; CHECK: [[LATCH]]:
-; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
+; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV1]], 1
; CHECK-NEXT: [[RANGE_CHECK:%.*]] = icmp ult i64 [[IV_NEXT]], [[EXIT]]
-; CHECK-NEXT: br i1 [[RANGE_CHECK]], label %[[LOOP]], label %[[DEOPT_LOOPEXIT:.*]]
+; CHECK-NEXT: br i1 [[RANGE_CHECK]], label %[[LOOP1]], label %[[DEOPT_LOOPEXIT]], !llvm.loop [[LOOP3:![0-9]+]]
; CHECK: [[FOUND]]:
; CHECK-NEXT: ret ptr null
; CHECK: [[DEOPT_LOOPEXIT]]:
@@ -75,3 +100,9 @@ declare i64 @llvm.umin.i64(i64, i64)
declare void @llvm.assume(i1 noundef)
attributes #0 = { nofree nosync }
+;.
+; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
+;.
\ No newline at end of file
>From 2f54c7c6e3a8f83ade28d11c318ac0479bb07fb1 Mon Sep 17 00:00:00 2001
From: Aleksandr Popov <apopov at azul.com>
Date: Mon, 6 Jul 2026 14:16:39 +0000
Subject: [PATCH 2/3] Add unit test, clang-format fix
---
llvm/lib/Analysis/ScalarEvolution.cpp | 3 +-
.../Analysis/ScalarEvolutionTest.cpp | 40 +++++++++++++++++++
2 files changed, 41 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Analysis/ScalarEvolution.cpp b/llvm/lib/Analysis/ScalarEvolution.cpp
index a168e4e8e1c12..9bf24fc644e0a 100644
--- a/llvm/lib/Analysis/ScalarEvolution.cpp
+++ b/llvm/lib/Analysis/ScalarEvolution.cpp
@@ -12942,8 +12942,7 @@ static bool IsKnownPredicateViaMinOrMax(ScalarEvolution &SE, CmpPredicate Pred,
// When Op and RHS share a common base differing by a
// constant offset D (RHS - Op = D), Op u< RHS holds iff D != 0 and
// RHS >= D (unsigned), i.e. the subtraction doesn't underflow.
- if (Diff && !Diff->isZero() &&
- SE.getUnsignedRangeMin(RHS).uge(*Diff))
+ if (Diff && !Diff->isZero() && SE.getUnsignedRangeMin(RHS).uge(*Diff))
return true;
}
}
diff --git a/llvm/unittests/Analysis/ScalarEvolutionTest.cpp b/llvm/unittests/Analysis/ScalarEvolutionTest.cpp
index dcffed39d9381..64254c2c3980f 100644
--- a/llvm/unittests/Analysis/ScalarEvolutionTest.cpp
+++ b/llvm/unittests/Analysis/ScalarEvolutionTest.cpp
@@ -1278,6 +1278,46 @@ TEST_F(ScalarEvolutionsTest, SCEVAddNUW) {
});
}
+TEST_F(ScalarEvolutionsTest, ProveUMinULT) {
+ LLVMContext C;
+ SMDiagnostic Err;
+ std::unique_ptr<Module> M =
+ parseAssemblyString("define void @foo(i32 %x, i32 %y) { "
+ " ret void "
+ "} ",
+ Err, C);
+
+ ASSERT_TRUE(M && "Could not parse module?");
+ ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
+
+ runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
+ auto *Ty = Type::getInt32Ty(F.getContext());
+ const SCEV *X = SE.getSCEV(getArgByName(F, "x"));
+ const SCEV *Y = SE.getSCEV(getArgByName(F, "y"));
+ const SCEV *UMin = SE.getUMinExpr(X, Y);
+
+ // umin(X, Y) u< X + 1 when add is NUW.
+ const SCEV *XPlusOneNUW = SE.getAddExpr(X, SE.getOne(Ty), SCEV::FlagNUW);
+ EXPECT_TRUE(SE.isKnownPredicate(ICmpInst::ICMP_ULT, UMin, XPlusOneNUW));
+
+ // Same via ICMP_UGT (swapped operands).
+ EXPECT_TRUE(SE.isKnownPredicate(ICmpInst::ICMP_UGT, XPlusOneNUW, UMin));
+
+ // Check the second operand: umin(X, Y) u< Y + 5 with NUW.
+ const SCEV *Five = SE.getConstant(APInt(32, 5));
+ const SCEV *YPlus5NUW = SE.getAddExpr(Y, Five, SCEV::FlagNUW);
+ EXPECT_TRUE(SE.isKnownPredicate(ICmpInst::ICMP_ULT, UMin, YPlus5NUW));
+
+ // Negative: without NUW, X + 2 might wrap to 0.
+ const SCEV *Two = SE.getConstant(APInt(32, 2));
+ const SCEV *XPlus2NoFlags = SE.getAddExpr(X, Two);
+ EXPECT_FALSE(SE.isKnownPredicate(ICmpInst::ICMP_ULT, UMin, XPlus2NoFlags));
+
+ // Negative: umin(X, Y) u< X is not provable (equal when X <= Y).
+ EXPECT_FALSE(SE.isKnownPredicate(ICmpInst::ICMP_ULT, UMin, X));
+ });
+}
+
TEST_F(ScalarEvolutionsTest, SCEVgetRanges) {
LLVMContext C;
SMDiagnostic Err;
>From 539bd3bbdbf66a3f28f22d5c91c4a57b3ffe384f Mon Sep 17 00:00:00 2001
From: Aleksandr Popov <apopov at azul.com>
Date: Wed, 8 Jul 2026 14:02:50 +0000
Subject: [PATCH 3/3] Regenerate the lit test with --check-globals=none (NFC)
---
.../early-exit-umin-trip-count.ll | 49 +++++++++----------
1 file changed, 22 insertions(+), 27 deletions(-)
diff --git a/llvm/test/Transforms/LoopVectorize/early-exit-umin-trip-count.ll b/llvm/test/Transforms/LoopVectorize/early-exit-umin-trip-count.ll
index 762a118130a46..202c839e44aad 100644
--- a/llvm/test/Transforms/LoopVectorize/early-exit-umin-trip-count.ll
+++ b/llvm/test/Transforms/LoopVectorize/early-exit-umin-trip-count.ll
@@ -1,4 +1,4 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
; RUN: opt -S -passes=loop-vectorize %s | FileCheck %s
; The loop should be vectorized. The vectorizer needs to prove
@@ -20,44 +20,44 @@ define ptr @main(i32 %length, ptr %ptr, i64 %0) #0 {
; CHECK-NEXT: [[EXIT:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP0]], i64 [[LENGTH_64]])
; CHECK-NEXT: call void @llvm.assume(i1 true) [ "align"(ptr [[PTR]], i64 8), "dereferenceable"(ptr [[PTR]], i64 [[LENGTH_64]]) ]
; CHECK-NEXT: [[NULL_CHECK:%.*]] = icmp eq i32 [[LENGTH]], 0
-; CHECK-NEXT: br i1 [[NULL_CHECK]], label %[[DEOPT]], label %[[SCALAR_PH:.*]]
-; CHECK: [[SCALAR_PH]]:
+; CHECK-NEXT: br i1 [[NULL_CHECK]], label %[[DEOPT]], label %[[SCALAR_PH1:.*]]
+; CHECK: [[SCALAR_PH1]]:
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[EXIT]], 4
-; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH1:.*]], label %[[VECTOR_PH:.*]]
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
; CHECK: [[VECTOR_PH]]:
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[EXIT]], 4
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[EXIT]], [[N_MOD_VF]]
-; CHECK-NEXT: br label %[[LOOP:.*]]
-; CHECK: [[LOOP]]:
-; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY_INTERIM:.*]] ]
-; CHECK-NEXT: [[ELEMENT_GEP:%.*]] = getelementptr i8, ptr [[PTR]], i64 [[IV]]
-; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i8>, ptr [[ELEMENT_GEP]], align 1
+; CHECK-NEXT: br label %[[LOOP1:.*]]
+; CHECK: [[LOOP1]]:
+; CHECK-NEXT: [[IV1:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY_INTERIM:.*]] ]
+; CHECK-NEXT: [[ELEMENT_GEP1:%.*]] = getelementptr i8, ptr [[PTR]], i64 [[IV1]]
+; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i8>, ptr [[ELEMENT_GEP1]], align 1
; CHECK-NEXT: [[TMP2:%.*]] = icmp eq <4 x i8> [[WIDE_LOAD]], zeroinitializer
; CHECK-NEXT: [[TMP3:%.*]] = freeze <4 x i1> [[TMP2]]
; CHECK-NEXT: [[TMP4:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP3]])
-; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[IV]], 4
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[IV1]], 4
; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP4]], label %[[VECTOR_EARLY_EXIT:.*]], label %[[VECTOR_BODY_INTERIM]]
; CHECK: [[VECTOR_BODY_INTERIM]]:
-; CHECK-NEXT: br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[LOOP]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK-NEXT: br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[LOOP1]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: [[MIDDLE_BLOCK]]:
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[EXIT]], [[N_VEC]]
-; CHECK-NEXT: br i1 [[CMP_N]], label %[[DEOPT_LOOPEXIT:.*]], label %[[SCALAR_PH1]]
+; CHECK-NEXT: br i1 [[CMP_N]], label %[[DEOPT_LOOPEXIT:.*]], label %[[SCALAR_PH]]
; CHECK: [[VECTOR_EARLY_EXIT]]:
; CHECK-NEXT: br label %[[FOUND:.*]]
-; CHECK: [[SCALAR_PH1]]:
-; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[SCALAR_PH]] ]
-; CHECK-NEXT: br label %[[LOOP1:.*]]
-; CHECK: [[LOOP1]]:
-; CHECK-NEXT: [[IV1:%.*]] = phi i64 [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ], [ [[BC_RESUME_VAL]], %[[SCALAR_PH1]] ]
-; CHECK-NEXT: [[ELEMENT_GEP1:%.*]] = getelementptr i8, ptr [[PTR]], i64 [[IV1]]
-; CHECK-NEXT: [[ELEMENT:%.*]] = load i8, ptr [[ELEMENT_GEP1]], align 1
+; CHECK: [[SCALAR_PH]]:
+; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[SCALAR_PH1]] ]
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ], [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ]
+; CHECK-NEXT: [[ELEMENT_GEP:%.*]] = getelementptr i8, ptr [[PTR]], i64 [[IV]]
+; CHECK-NEXT: [[ELEMENT:%.*]] = load i8, ptr [[ELEMENT_GEP]], align 1
; CHECK-NEXT: [[FOUND_CHECK:%.*]] = icmp eq i8 [[ELEMENT]], 0
; CHECK-NEXT: br i1 [[FOUND_CHECK]], label %[[FOUND]], label %[[LATCH]]
; CHECK: [[LATCH]]:
-; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV1]], 1
+; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
; CHECK-NEXT: [[RANGE_CHECK:%.*]] = icmp ult i64 [[IV_NEXT]], [[EXIT]]
-; CHECK-NEXT: br i1 [[RANGE_CHECK]], label %[[LOOP1]], label %[[DEOPT_LOOPEXIT]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK-NEXT: br i1 [[RANGE_CHECK]], label %[[LOOP]], label %[[DEOPT_LOOPEXIT]], !llvm.loop [[LOOP3:![0-9]+]]
; CHECK: [[FOUND]]:
; CHECK-NEXT: ret ptr null
; CHECK: [[DEOPT_LOOPEXIT]]:
@@ -100,9 +100,4 @@ declare i64 @llvm.umin.i64(i64, i64)
declare void @llvm.assume(i1 noundef)
attributes #0 = { nofree nosync }
-;.
-; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
-; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
-; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
-; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
-;.
\ No newline at end of file
+
More information about the llvm-commits
mailing list