[llvm] [SCEV] Sharpen BTC computations with minus-one overflow trick (PR #222939)
Ramkumar Ramachandra via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 11 05:44:17 PDT 2026
https://github.com/artagnon created https://github.com/llvm/llvm-project/pull/222939
Add a special case for LHS >= RHS in isKnownPredicateWithNoOverflow in order to sharpen the result for users doing BTC-computations. It should be non-functional for other users.
>From 023fa64a7f0cc13f9ab4e02ed92d9837e090d536 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <artagnon at tenstorrent.com>
Date: Fri, 11 Sep 2026 13:16:07 +0100
Subject: [PATCH] [SCEV] Sharpen BTC computations with minus-one overflow trick
Add a special case for LHS >= RHS in isKnownPredicateWithNoOverflow in
order to sharpen the result for users doing BTC-computations. It should
be non-functional for other users.
---
llvm/lib/Analysis/ScalarEvolution.cpp | 15 ++++++++++++---
.../max-backedge-taken-count-guard-info.ll | 4 ++--
.../test/Analysis/ScalarEvolution/trip-count13.ll | 8 ++++----
.../Transforms/IndVarSimplify/trivial-checks.ll | 10 +++-------
4 files changed, 21 insertions(+), 16 deletions(-)
diff --git a/llvm/lib/Analysis/ScalarEvolution.cpp b/llvm/lib/Analysis/ScalarEvolution.cpp
index 06f350fd2179b..f418114c62fd9 100644
--- a/llvm/lib/Analysis/ScalarEvolution.cpp
+++ b/llvm/lib/Analysis/ScalarEvolution.cpp
@@ -11703,9 +11703,9 @@ bool ScalarEvolution::isKnownPredicateViaNoOverflow(CmpPredicate Pred,
// C1 and C2 are constant integers. If either X or Y are not add expressions,
// consider them as X + 0 and Y + 0 respectively. C1 and C2 are returned via
// OutC1 and OutC2.
- auto MatchBinaryAddToConst = [this](SCEVUse X, SCEVUse Y, APInt &OutC1,
- APInt &OutC2,
- SCEV::NoWrapFlags ExpectedFlags) {
+ auto MatchBinaryAddToConst = [this, &Pred](SCEVUse X, SCEVUse Y, APInt &OutC1,
+ APInt &OutC2,
+ SCEV::NoWrapFlags ExpectedFlags) {
SCEVUse XNonConstOp, XConstOp;
SCEVUse YNonConstOp, YConstOp;
SCEV::NoWrapFlags XFlagsPresent;
@@ -11740,6 +11740,15 @@ bool ScalarEvolution::isKnownPredicateViaNoOverflow(CmpPredicate Pred,
return false;
}
+ // X >= Y is equivalent to X > Y - 1, and this holds trivially when Y - 1
+ // does not overflow. When it does overflow, it must be INT_MAX, and the
+ // comparison would evaluate to false anyway.
+ // This simplification is useful when computing backedge-taken-counts.
+ if (YConstOp->isZero() && ICmpInst::isGE(Pred)) {
+ Pred = ICmpInst::getStrictPredicate(Pred);
+ YConstOp = getMinusOne(Y->getType());
+ }
+
OutC1 = cast<SCEVConstant>(XConstOp)->getAPInt();
OutC2 = cast<SCEVConstant>(YConstOp)->getAPInt();
diff --git a/llvm/test/Analysis/ScalarEvolution/max-backedge-taken-count-guard-info.ll b/llvm/test/Analysis/ScalarEvolution/max-backedge-taken-count-guard-info.ll
index 8f7d2007cdcc0..f9c7f041bff46 100644
--- a/llvm/test/Analysis/ScalarEvolution/max-backedge-taken-count-guard-info.ll
+++ b/llvm/test/Analysis/ScalarEvolution/max-backedge-taken-count-guard-info.ll
@@ -1751,7 +1751,7 @@ define void @range_check_idiom_through_zext_ugt(i32 %len) {
; CHECK-NEXT: --> {1,+,1}<nuw><nsw><%loop> U: [1,-2147483648) S: [1,-2147483648) Exits: %len LoopDispositions: { %loop: Computable }
; CHECK-NEXT: Determining loop execution counts for: @range_check_idiom_through_zext_ugt
; CHECK-NEXT: Loop %loop: backedge-taken count is (-1 + %len)
-; CHECK-NEXT: Loop %loop: constant max backedge-taken count is i32 -1
+; CHECK-NEXT: Loop %loop: constant max backedge-taken count is i32 -2
; CHECK-NEXT: Loop %loop: symbolic max backedge-taken count is (-1 + %len)
; CHECK-NEXT: Loop %loop: Trip multiple is 1
;
@@ -1982,7 +1982,7 @@ define void @one_sided_guard_does_not_bound_difference(ptr %dst, i64 %n, i64 %st
; CHECK-NEXT: --> {(1 + %start),+,1}<nuw><%loop> U: full-set S: full-set Exits: %n LoopDispositions: { %loop: Computable }
; CHECK-NEXT: Determining loop execution counts for: @one_sided_guard_does_not_bound_difference
; CHECK-NEXT: Loop %loop: backedge-taken count is (-1 + (-1 * %start) + %n)
-; CHECK-NEXT: Loop %loop: constant max backedge-taken count is i64 -1
+; CHECK-NEXT: Loop %loop: constant max backedge-taken count is i64 -2
; CHECK-NEXT: Loop %loop: symbolic max backedge-taken count is (-1 + (-1 * %start) + %n)
; CHECK-NEXT: Loop %loop: Trip multiple is 1
;
diff --git a/llvm/test/Analysis/ScalarEvolution/trip-count13.ll b/llvm/test/Analysis/ScalarEvolution/trip-count13.ll
index 720b2ae49c090..e36dd45d3ba03 100644
--- a/llvm/test/Analysis/ScalarEvolution/trip-count13.ll
+++ b/llvm/test/Analysis/ScalarEvolution/trip-count13.ll
@@ -6,10 +6,10 @@ define void @u_0(i8 %rhs) {
;
; CHECK-LABEL: 'u_0'
; CHECK-NEXT: Determining loop execution counts for: @u_0
-; CHECK-NEXT: Loop %loop: backedge-taken count is (-100 + (-1 * %rhs) + ((100 + %rhs) umax %rhs))
-; CHECK-NEXT: Loop %loop: constant max backedge-taken count is i8 -100, actual taken count either this or zero.
-; CHECK-NEXT: Loop %loop: symbolic max backedge-taken count is (-100 + (-1 * %rhs) + ((100 + %rhs) umax %rhs)), actual taken count either this or zero.
-; CHECK-NEXT: Loop %loop: Trip multiple is 1
+; CHECK-NEXT: Loop %loop: backedge-taken count is i8 -100
+; CHECK-NEXT: Loop %loop: constant max backedge-taken count is i8 -100
+; CHECK-NEXT: Loop %loop: symbolic max backedge-taken count is i8 -100
+; CHECK-NEXT: Loop %loop: Trip multiple is 157
;
entry:
%start = add i8 %rhs, 100
diff --git a/llvm/test/Transforms/IndVarSimplify/trivial-checks.ll b/llvm/test/Transforms/IndVarSimplify/trivial-checks.ll
index 05aba4be5d54f..ce69b5f4fbed5 100644
--- a/llvm/test/Transforms/IndVarSimplify/trivial-checks.ll
+++ b/llvm/test/Transforms/IndVarSimplify/trivial-checks.ll
@@ -2,6 +2,7 @@
; RUN: opt -passes=indvars -S < %s | FileCheck %s
; FIXME: In all cases, x is from [0; 1000) and we cannot prove that x + 1 > x.
+; We can prove it for the ugt case though.
define void @test_sgt(i32 %x) {
; CHECK-LABEL: @test_sgt(
@@ -101,14 +102,9 @@ define void @test_ugt(i32 %x) {
; CHECK: loop.preheader:
; CHECK-NEXT: br label [[LOOP:%.*]]
; CHECK: loop:
-; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], [[GUARDED:%.*]] ], [ [[X]], [[LOOP_PREHEADER]] ]
-; CHECK-NEXT: [[TMP:%.*]] = add nsw i32 [[IV]], 1
-; CHECK-NEXT: [[GUARD:%.*]] = icmp ugt i32 [[TMP]], [[IV]]
-; CHECK-NEXT: br i1 [[GUARD]], label [[GUARDED]], label [[FAIL:%.*]]
+; CHECK-NEXT: br i1 false, label [[GUARDED:%.*]], label [[FAIL:%.*]]
; CHECK: guarded:
-; CHECK-NEXT: [[IV_NEXT]] = add nsw i32 [[IV]], -1
-; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV]], 0
-; CHECK-NEXT: br i1 [[COND]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
+; CHECK-NEXT: br i1 true, label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
; CHECK: exit.loopexit:
; CHECK-NEXT: br label [[EXIT]]
; CHECK: exit:
More information about the llvm-commits
mailing list