From 632ee84d2d21ec83f585ea5ac591ec7183cebc42 Mon Sep 17 00:00:00 2001 From: Florian Hahn Date: Thu, 7 Mar 2024 16:18:19 +0000 Subject: [PATCH] [ValueTracking] Treat phi as underlying obj when not decomposing further At the moment, getUnderlyingObjects simply continues for phis that do not refer to the same underlying object in loops, without adding them to the list of underlying objects, effectively ignoring those phis. Instead of ignoring those phis, add them to the list of underlying objects. This fixes a miscompile where LoopAccessAnalysis fails to identify a memory dependence, because no underlying objects can be found for a set of memory accesses. Fixes https://github.com/llvm/llvm-project/issues/82665. --- llvm/lib/Analysis/ValueTracking.cpp | 2 ++ .../underlying-object-loop-varying-phi.ll | 7 ++++++- 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp index 52ae9f034e5d3..b9c046ea9fe83 100644 --- a/llvm/lib/Analysis/ValueTracking.cpp +++ b/llvm/lib/Analysis/ValueTracking.cpp @@ -6123,6 +6123,8 @@ void llvm::getUnderlyingObjects(const Value *V, if (!LI || !LI->isLoopHeader(PN->getParent()) || isSameUnderlyingObjectInLoop(PN, LI)) append_range(Worklist, PN->incoming_values()); + else + Objects.push_back(P); continue; } diff --git a/llvm/test/Analysis/LoopAccessAnalysis/underlying-object-loop-varying-phi.ll b/llvm/test/Analysis/LoopAccessAnalysis/underlying-object-loop-varying-phi.ll index 1a5a6ac08d404..106dc8c13a49f 100644 --- a/llvm/test/Analysis/LoopAccessAnalysis/underlying-object-loop-varying-phi.ll +++ b/llvm/test/Analysis/LoopAccessAnalysis/underlying-object-loop-varying-phi.ll @@ -7,8 +7,13 @@ target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128" define void @indirect_ptr_recurrences_read_write(ptr %A, ptr %B) { ; CHECK-LABEL: 'indirect_ptr_recurrences_read_write' ; CHECK-NEXT: loop: -; CHECK-NEXT: Memory dependences are safe +; CHECK-NEXT: Report: unsafe dependent memory operations in loop. Use #pragma clang loop distribute(enable) to allow loop distribution to attempt to isolate the offending operations into a separate loop +; CHECK-NEXT: Unsafe indirect dependence. ; CHECK-NEXT: Dependences: +; CHECK-NEXT: IndidrectUnsafe: +; CHECK-NEXT: %l = load i32, ptr %ptr.recur, align 4, !tbaa !4 -> +; CHECK-NEXT: store i32 %xor, ptr %ptr.recur, align 4, !tbaa !4 +; CHECK-EMPTY: ; CHECK-NEXT: Run-time memory checks: ; CHECK-NEXT: Grouped accesses: ; CHECK-EMPTY: