[llvm] Revert "LSV: forbid load-cycles when vectorizing; fix bug (#104815)" (PR #106245)
Danial Klimkin via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 27 09:44:34 PDT 2024
https://github.com/dklimkin created https://github.com/llvm/llvm-project/pull/106245
This reverts commit c46b41aaa6eaa787f808738d14c61a2f8b6d839f.
Multiple tests time out, either due to performance hit (see comment) or a cycle.
>From 7782963bc509f539bce764182227760cdaee8710 Mon Sep 17 00:00:00 2001
From: Danial Klimkin <dklimkin at google.com>
Date: Tue, 27 Aug 2024 18:42:00 +0200
Subject: [PATCH] Revert "LSV: forbid load-cycles when vectorizing; fix bug
(#104815)"
This reverts commit c46b41aaa6eaa787f808738d14c61a2f8b6d839f. Multiple tests timeout, either due to performance hit (see comment) or a cycle.
---
.../Vectorize/LoadStoreVectorizer.cpp | 24 +-----
.../LoadStoreVectorizer/AArch64/pr37865.ll | 79 +------------------
2 files changed, 5 insertions(+), 98 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp b/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
index dd37c95eca61a3..02ec1d5c259cd6 100644
--- a/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
@@ -998,32 +998,10 @@ bool Vectorizer::isSafeToMove(
LLVM_DEBUG(dbgs() << "LSV: isSafeToMove(" << *ChainElem << " -> "
<< *ChainBegin << ")\n");
- assert(isa<LoadInst>(ChainElem) == IsLoadChain &&
- isa<LoadInst>(ChainBegin) == IsLoadChain);
-
+ assert(isa<LoadInst>(ChainElem) == IsLoadChain);
if (ChainElem == ChainBegin)
return true;
- if constexpr (IsLoadChain) {
- // If ChainElem depends on ChainBegin, they're not safe to reorder.
- SmallVector<Instruction *, 8> Worklist;
- Worklist.emplace_back(ChainElem);
- while (!Worklist.empty()) {
- Instruction *I = Worklist.pop_back_val();
- for (Use &O : I->operands()) {
- if (isa<PHINode>(O))
- continue;
- if (auto *J = dyn_cast<Instruction>(O)) {
- if (J == ChainBegin) {
- LLVM_DEBUG(dbgs() << "LSV: dependent loads; not safe to reorder\n");
- return false;
- }
- Worklist.emplace_back(J);
- }
- }
- }
- }
-
// Invariant loads can always be reordered; by definition they are not
// clobbered by stores.
if (isInvariantLoad(ChainElem))
diff --git a/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll b/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
index 0beca8c15305f6..833e70814c2917 100644
--- a/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
+++ b/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
@@ -1,22 +1,9 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
-; RUN: opt -mtriple=aarch64 -passes=load-store-vectorizer -S %s | FileCheck %s
-
-; LSV was attempting to vectorize this earlier, but crashed while re-ordering
-; instructions due to the load-load cycle. Now, the candidate loads are no
-; longer considered safe for reordering.
+; REQUIRES: asserts
+; RUN: not --crash opt -mtriple=aarch64 -passes=load-store-vectorizer \
+; RUN: -disable-output %s 2>&1 | FileCheck %s
define i32 @load_cycle(ptr %x) {
-; CHECK-LABEL: define i32 @load_cycle(
-; CHECK-SAME: ptr [[X:%.*]]) {
-; CHECK-NEXT: [[ENTRY:.*:]]
-; CHECK-NEXT: [[GEP_X_1:%.*]] = getelementptr inbounds [2 x i32], ptr [[X]], i32 0, i32 1
-; CHECK-NEXT: [[LOAD_X_1:%.*]] = load i32, ptr [[GEP_X_1]], align 4
-; CHECK-NEXT: [[REM:%.*]] = urem i32 [[LOAD_X_1]], 1
-; CHECK-NEXT: [[GEP_X_2:%.*]] = getelementptr inbounds [2 x i32], ptr [[X]], i32 [[REM]], i32 0
-; CHECK-NEXT: [[LOAD_X_2:%.*]] = load i32, ptr [[GEP_X_2]], align 4
-; CHECK-NEXT: [[RET:%.*]] = add i32 [[LOAD_X_2]], [[LOAD_X_1]]
-; CHECK-NEXT: ret i32 [[RET]]
-;
+; CHECK: Unexpected cycle while re-ordering instructions
entry:
%gep.x.1 = getelementptr inbounds [2 x i32], ptr %x, i32 0, i32 1
%load.x.1 = load i32, ptr %gep.x.1
@@ -26,61 +13,3 @@ entry:
%ret = add i32 %load.x.2, %load.x.1
ret i32 %ret
}
-
-define i32 @load_cycle2(ptr %x, i32 %y) {
-; CHECK-LABEL: define i32 @load_cycle2(
-; CHECK-SAME: ptr [[X:%.*]], i32 [[Y:%.*]]) {
-; CHECK-NEXT: [[ENTRY:.*:]]
-; CHECK-NEXT: [[GEP_X_1:%.*]] = getelementptr inbounds [2 x i32], ptr [[X]], i32 [[Y]], i32 1
-; CHECK-NEXT: [[LOAD_X_1:%.*]] = load i32, ptr [[GEP_X_1]], align 4
-; CHECK-NEXT: [[MUL:%.*]] = mul i32 [[LOAD_X_1]], 2
-; CHECK-NEXT: [[ADD:%.*]] = add i32 [[Y]], [[MUL]]
-; CHECK-NEXT: [[SUB_1:%.*]] = sub i32 [[ADD]], [[LOAD_X_1]]
-; CHECK-NEXT: [[SUB_2:%.*]] = sub i32 [[SUB_1]], [[LOAD_X_1]]
-; CHECK-NEXT: [[GEP_X_2:%.*]] = getelementptr inbounds [2 x i32], ptr [[X]], i32 [[SUB_2]], i32 0
-; CHECK-NEXT: [[LOAD_X_2:%.*]] = load i32, ptr [[GEP_X_2]], align 4
-; CHECK-NEXT: [[RET:%.*]] = add i32 [[LOAD_X_2]], [[LOAD_X_1]]
-; CHECK-NEXT: ret i32 [[RET]]
-;
-entry:
- %gep.x.1 = getelementptr inbounds [2 x i32], ptr %x, i32 %y, i32 1
- %load.x.1 = load i32, ptr %gep.x.1
- %mul = mul i32 %load.x.1, 2
- %add = add i32 %y, %mul
- %sub.1 = sub i32 %add, %load.x.1
- %sub.2 = sub i32 %sub.1, %load.x.1
- %gep.x.2 = getelementptr inbounds [2 x i32], ptr %x, i32 %sub.2, i32 0
- %load.x.2 = load i32, ptr %gep.x.2
- %ret = add i32 %load.x.2, %load.x.1
- ret i32 %ret
-}
-
- at global.1 = global i32 0
- at global.2 = global [1 x [3 x i32]] zeroinitializer
-
-define i16 @load_cycle3() {
-; CHECK-LABEL: define i16 @load_cycle3() {
-; CHECK-NEXT: [[ENTRY:.*:]]
-; CHECK-NEXT: [[LOAD_1:%.*]] = load i32, ptr @global.1, align 4
-; CHECK-NEXT: [[UREM_1:%.*]] = urem i32 [[LOAD_1]], 1
-; CHECK-NEXT: [[GEP_1:%.*]] = getelementptr inbounds [1 x [3 x i32]], ptr @global.2, i32 0, i32 [[UREM_1]]
-; CHECK-NEXT: [[GEP_2:%.*]] = getelementptr inbounds [3 x i32], ptr [[GEP_1]], i32 0, i32 2
-; CHECK-NEXT: [[LOAD_2:%.*]] = load i32, ptr [[GEP_2]], align 4
-; CHECK-NEXT: [[UREM_2:%.*]] = urem i32 [[LOAD_2]], 1
-; CHECK-NEXT: [[GEP_3:%.*]] = getelementptr inbounds [1 x [3 x i32]], ptr @global.2, i32 0, i32 [[UREM_2]]
-; CHECK-NEXT: [[GEP_4:%.*]] = getelementptr inbounds [3 x i32], ptr [[GEP_3]], i32 0, i32 1
-; CHECK-NEXT: [[LOAD_3:%.*]] = load i32, ptr [[GEP_4]], align 4
-; CHECK-NEXT: ret i16 0
-;
-entry:
- %load.1 = load i32, ptr @global.1
- %urem.1 = urem i32 %load.1, 1
- %gep.1 = getelementptr inbounds [1 x [3 x i32]], ptr @global.2, i32 0, i32 %urem.1
- %gep.2 = getelementptr inbounds [3 x i32], ptr %gep.1, i32 0, i32 2
- %load.2 = load i32, ptr %gep.2
- %urem.2 = urem i32 %load.2, 1
- %gep.3 = getelementptr inbounds [1 x [3 x i32]], ptr @global.2, i32 0, i32 %urem.2
- %gep.4 = getelementptr inbounds [3 x i32], ptr %gep.3, i32 0, i32 1
- %load.3 = load i32, ptr %gep.4
- ret i16 0
-}
More information about the llvm-commits
mailing list