[llvm] [MemCpyOpt] move SrcAlloca to the entry if transformation is performed (PR #67226)
Kohei Asano via llvm-commits
llvm-commits at lists.llvm.org
Sat Sep 23 01:33:34 PDT 2023
https://github.com/khei4 created https://github.com/llvm/llvm-project/pull/67226
This is fixup for https://github.com/llvm/llvm-project/pull/66618#discussion_r1328523770 .
This transformation checks whether allocas are static, if the transformation is performed. This patch moves the SrcAlloca to the entry of the BB when the optimization performed.
>From 681566e3760d3302d449b6d75ae3d500c71edbfb Mon Sep 17 00:00:00 2001
From: khei4 <kk.asano.luxy at gmail.com>
Date: Sat, 23 Sep 2023 17:28:35 +0900
Subject: [PATCH] [MemCpyOpt] move SrcAlloca to the entry if transformation is
performed
---
.../lib/Transforms/Scalar/MemCpyOptimizer.cpp | 18 +++----
llvm/test/Transforms/MemCpyOpt/stack-move.ll | 49 ++++++++-----------
2 files changed, 28 insertions(+), 39 deletions(-)
diff --git a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
index d87f2fb59814edf..2639a49fda71723 100644
--- a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
+++ b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -1458,8 +1458,8 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
// Check that src and dest are never captured, unescaped allocas. Also
// find the nearest common dominator and postdominator for all users in
- // order to shrink wrap the lifetimes, and instructions with noalias metadata
- // to remove them.
+ // order to shrink wrap the lifetimes, and instructions with noalias
+ // metadata to remove them.
SmallVector<Instruction *, 4> LifetimeMarkers;
SmallSet<Instruction *, 4> NoAliasInstrs;
@@ -1483,14 +1483,6 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
Worklist.pop_back();
for (const Use &U : I->uses()) {
auto *UI = cast<Instruction>(U.getUser());
- // TODO: We can perform the transformation if we move src alloca to
- // before the dominator of all uses. If any use that isn't dominated by
- // SrcAlloca exists, non-dominating uses will be produced.
- if (!DT->dominates(SrcAlloca, UI)) {
- LLVM_DEBUG(dbgs() << "Stack Move: SrcAlloca doesn't dominate all "
- "uses for the location, bailing\n");
- return false;
- }
if (Visited.size() >= MaxUsesToExplore) {
LLVM_DEBUG(
dbgs()
@@ -1600,7 +1592,11 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
if (!CaptureTrackingWithModRef(SrcAlloca, SrcModRefCallback))
return false;
- // We can do the transformation. First, align the allocas appropriately.
+ // We can do the transformation. First, move the SrcAlloca to the entry point.
+ // This is safe because any static alloca is in EntryBB.
+ SrcAlloca->moveBefore(*SrcAlloca->getParent(),
+ SrcAlloca->getParent()->getFirstInsertionPt());
+ // Align the allocas appropriately.
SrcAlloca->setAlignment(
std::max(SrcAlloca->getAlign(), DestAlloca->getAlign()));
diff --git a/llvm/test/Transforms/MemCpyOpt/stack-move.ll b/llvm/test/Transforms/MemCpyOpt/stack-move.ll
index f6be486dce8ac1e..a17a5530072ff82 100644
--- a/llvm/test/Transforms/MemCpyOpt/stack-move.ll
+++ b/llvm/test/Transforms/MemCpyOpt/stack-move.ll
@@ -94,6 +94,26 @@ define void @load_store() {
ret void
}
+define i32 @use_not_dominated_by_src_alloca() {
+; CHECK-LABEL: define i32 @use_not_dominated_by_src_alloca() {
+; CHECK-NEXT: [[SRC:%.*]] = alloca i8, align 4
+; CHECK-NEXT: [[DEST_GEP:%.*]] = getelementptr i64, ptr [[SRC]], i64 -1
+; CHECK-NEXT: [[DEST_USE:%.*]] = load i8, ptr [[DEST_GEP]], align 1
+; CHECK-NEXT: ret i32 0
+;
+ %dest = alloca i1, align 1
+ ; Replacing the use of dest with src causes no domination uses.
+ %dest.gep = getelementptr i64, ptr %dest, i64 -1
+ %dest.use = load i8, ptr %dest.gep, align 1
+ %src = alloca i8, align 4
+ %src.val = load i1, ptr %src, align 4
+
+ store i1 %src.val, ptr %dest, align 1
+
+ ret i32 0
+}
+
+
; Tests that merging two allocas shouldn't be more poisonous, smaller aligned src is valid.
define void @align_up() {
; CHECK-LABEL: define void @align_up() {
@@ -718,13 +738,12 @@ unr2:
}
-; TODO: merge allocas for multi basicblock loop case.
define void @multi_bb_loop(i32 %n) {
; CHECK-LABEL: define void @multi_bb_loop
; CHECK-SAME: (i32 [[N:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[NLT1:%.*]] = icmp slt i32 [[N]], 1
; CHECK-NEXT: [[SRC:%.*]] = alloca [[STRUCT_FOO:%.*]], align 8
+; CHECK-NEXT: [[NLT1:%.*]] = icmp slt i32 [[N]], 1
; CHECK-NEXT: store [[STRUCT_FOO]] { i32 0, i32 1, i32 42 }, ptr [[SRC]], align 4
; CHECK-NEXT: br i1 [[NLT1]], label [[LOOP_EXIT:%.*]], label [[LOOP_BODY:%.*]]
; CHECK: loop_body:
@@ -981,32 +1000,6 @@ bb2:
; Optimization failures follow:
-; TODO: we can merge those alloca if we move src alloca to the start of the BB.
-; Tests that a the optimization isn't performed,
-; when any use that isn't dominated by SrcAlloca exists.
-define i32 @use_not_dominated_by_src_alloca() {
-; CHECK-LABEL: define i32 @use_not_dominated_by_src_alloca() {
-; CHECK-NEXT: [[DEST:%.*]] = alloca i1, align 1
-; CHECK-NEXT: [[DEST_GEP:%.*]] = getelementptr i64, ptr [[DEST]], i64 -1
-; CHECK-NEXT: [[DEST_USE:%.*]] = load i8, ptr [[DEST_GEP]], align 1
-; CHECK-NEXT: [[SRC:%.*]] = alloca i8, align 4
-; CHECK-NEXT: [[SRC_VAL:%.*]] = load i1, ptr [[SRC]], align 4
-; CHECK-NEXT: store i1 [[SRC_VAL]], ptr [[DEST]], align 1
-; CHECK-NEXT: ret i32 0
-;
- %dest = alloca i1, align 1
- ; Replacing the use of dest with src causes no domination uses.
- %dest.gep = getelementptr i64, ptr %dest, i64 -1
- %dest.use = load i8, ptr %dest.gep, align 1
- %src = alloca i8, align 4
- %src.val = load i1, ptr %src, align 4
-
- store i1 %src.val, ptr %dest, align 1
-
- ret i32 0
-}
-
-
; Tests that a memcpy that doesn't completely overwrite a stack value is a use
; for the purposes of liveness analysis, not a definition.
define void @incomplete_memcpy() {
More information about the llvm-commits
mailing list