[SimplifyCFG] Add support for sinking instructions with multiple uses #95521
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sinking currently only supports instructions that have zero or one uses. Extend this to handle instructions with any number of uses, as long as all uses are consistent (i.e. the "same" for all sinking candidates). After llvm#94462 this is basically just a matter of looping over all uses instead of checking the first one only.
|
@llvm/pr-subscribers-llvm-transforms Author: Nikita Popov (nikic) ChangesSinking currently only supports instructions that have zero or one uses. Extend this to handle instructions with any number of uses, as long as all uses are consistent (i.e. the "same" for all sinking candidates). After #94462 this is basically just a matter of looping over all uses instead of checking the first one only. Full diff: https://github.com/llvm/llvm-project/pull/95521.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
index 9c7f90b0613a0..85d774d20daa0 100644
--- a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
@@ -1936,7 +1936,7 @@ static bool canSinkInstructions(
// Prune out obviously bad instructions to move. Each instruction must have
// exactly zero or one use, and we check later that use is by a single, common
// PHI instruction in the successor.
- bool HasUse = !Insts.front()->user_empty();
+ std::optional<unsigned> NumUses;
for (auto *I : Insts) {
// These instructions may change or break semantics if moved.
if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
@@ -1956,10 +1956,9 @@ static bool canSinkInstructions(
if (C->isInlineAsm() || C->cannotMerge() || C->isConvergent())
return false;
- // Each instruction must have zero or one use.
- if (HasUse && !I->hasOneUse())
- return false;
- if (!HasUse && !I->user_empty())
+ if (!NumUses)
+ NumUses = I->getNumUses();
+ else if (NumUses != I->getNumUses())
return false;
}
@@ -1987,8 +1986,7 @@ static bool canSinkInstructions(
// then the other phi operands must match the instructions from Insts. This
// also has to hold true for any phi nodes that would be created as a result
// of sinking. Both of these cases are represented by PhiOperands.
- if (HasUse) {
- const Use &U = *I0->use_begin();
+ for (const Use &U : I0->uses()) {
auto It = PHIOperands.find(&U);
if (It == PHIOperands.end())
// There may be uses in other blocks when sinking into a loop header.
@@ -2138,11 +2136,11 @@ static void sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
I0->andIRFlags(I);
}
- if (!I0->user_empty()) {
- // canSinkLastInstruction checked that all instructions were used by
- // one and only one PHI node. Find that now, RAUW it to our common
- // instruction and nuke it.
- auto *PN = cast<PHINode>(*I0->user_begin());
+ for (User *U : make_early_inc_range(I0->users())) {
+ // canSinkLastInstruction checked that all instructions are only used by
+ // phi nodes a way that allows replacing the phi node with the common
+ // instruction.
+ auto *PN = cast<PHINode>(U);
PN->replaceAllUsesWith(I0);
PN->eraseFromParent();
}
diff --git a/llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll b/llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll
index 7b2161351a794..be2f17dc1c780 100644
--- a/llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll
+++ b/llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll
@@ -1704,19 +1704,14 @@ return:
define ptr @multi_use_in_phi(i1 %cond, ptr %p, i64 %a, i64 %b) {
; CHECK-LABEL: @multi_use_in_phi(
-; CHECK-NEXT: br i1 [[COND:%.*]], label [[IF:%.*]], label [[ELSE:%.*]]
+; CHECK-NEXT: br i1 [[COND:%.*]], label [[IF:%.*]], label [[JOIN:%.*]]
; CHECK: if:
; CHECK-NEXT: call void @dummy()
-; CHECK-NEXT: [[GEP1_A:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[A:%.*]]
-; CHECK-NEXT: br label [[JOIN:%.*]]
-; CHECK: else:
-; CHECK-NEXT: [[GEP1_B:%.*]] = getelementptr i8, ptr [[P]], i64 [[A]]
; CHECK-NEXT: br label [[JOIN]]
; CHECK: join:
-; CHECK-NEXT: [[GEP1_B_SINK:%.*]] = phi ptr [ [[GEP1_B]], [[ELSE]] ], [ [[GEP1_A]], [[IF]] ]
-; CHECK-NEXT: [[PHI1:%.*]] = phi ptr [ [[GEP1_A]], [[IF]] ], [ [[GEP1_B]], [[ELSE]] ]
-; CHECK-NEXT: [[GEP2_B:%.*]] = getelementptr i8, ptr [[GEP1_B_SINK]], i64 [[B:%.*]]
-; CHECK-NEXT: call void @use.ptr(ptr [[PHI1]])
+; CHECK-NEXT: [[GEP1_B:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[A:%.*]]
+; CHECK-NEXT: [[GEP2_B:%.*]] = getelementptr i8, ptr [[GEP1_B]], i64 [[B:%.*]]
+; CHECK-NEXT: call void @use.ptr(ptr [[GEP1_B]])
; CHECK-NEXT: ret ptr [[GEP2_B]]
;
br i1 %cond, label %if, label %else
@@ -1778,23 +1773,16 @@ join:
define i64 @multi_use_in_block(i1 %cond, ptr %p, i64 %a, i64 %b) {
; CHECK-LABEL: @multi_use_in_block(
-; CHECK-NEXT: br i1 [[COND:%.*]], label [[IF:%.*]], label [[ELSE:%.*]]
+; CHECK-NEXT: br i1 [[COND:%.*]], label [[IF:%.*]], label [[JOIN:%.*]]
; CHECK: if:
; CHECK-NEXT: call void @dummy()
-; CHECK-NEXT: [[GEP1_A:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[A:%.*]]
-; CHECK-NEXT: [[V_A:%.*]] = load i64, ptr [[GEP1_A]], align 8
-; CHECK-NEXT: [[GEP2_A:%.*]] = getelementptr i8, ptr [[GEP1_A]], i64 [[V_A]]
-; CHECK-NEXT: br label [[JOIN:%.*]]
-; CHECK: else:
-; CHECK-NEXT: [[GEP1_B:%.*]] = getelementptr i8, ptr [[P]], i64 [[A]]
-; CHECK-NEXT: [[V_B:%.*]] = load i64, ptr [[GEP1_B]], align 8
-; CHECK-NEXT: [[GEP2_B:%.*]] = getelementptr i8, ptr [[GEP1_B]], i64 [[V_B]]
; CHECK-NEXT: br label [[JOIN]]
; CHECK: join:
-; CHECK-NEXT: [[PHI1:%.*]] = phi i64 [ [[V_A]], [[IF]] ], [ [[V_B]], [[ELSE]] ]
-; CHECK-NEXT: [[PHI2:%.*]] = phi ptr [ [[GEP2_A]], [[IF]] ], [ [[GEP2_B]], [[ELSE]] ]
-; CHECK-NEXT: call void @use.ptr(ptr [[PHI2]])
-; CHECK-NEXT: ret i64 [[PHI1]]
+; CHECK-NEXT: [[GEP1_B:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[A:%.*]]
+; CHECK-NEXT: [[V_B:%.*]] = load i64, ptr [[GEP1_B]], align 8
+; CHECK-NEXT: [[GEP2_B:%.*]] = getelementptr i8, ptr [[GEP1_B]], i64 [[V_B]]
+; CHECK-NEXT: call void @use.ptr(ptr [[GEP2_B]])
+; CHECK-NEXT: ret i64 [[V_B]]
;
br i1 %cond, label %if, label %else
@@ -1824,19 +1812,17 @@ define i64 @multi_use_in_block_inconsistent(i1 %cond, ptr %p, i64 %a, i64 %b) {
; CHECK: if:
; CHECK-NEXT: call void @dummy()
; CHECK-NEXT: [[GEP1_A:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[A:%.*]]
-; CHECK-NEXT: [[V_A:%.*]] = load i64, ptr [[GEP1_A]], align 8
-; CHECK-NEXT: [[GEP2_A:%.*]] = getelementptr i8, ptr [[GEP1_A]], i64 [[V_A]]
; CHECK-NEXT: br label [[JOIN:%.*]]
; CHECK: else:
; CHECK-NEXT: [[GEP1_B:%.*]] = getelementptr i8, ptr [[P]], i64 [[A]]
-; CHECK-NEXT: [[V_B:%.*]] = load i64, ptr [[P]], align 8
-; CHECK-NEXT: [[GEP2_B:%.*]] = getelementptr i8, ptr [[GEP1_B]], i64 [[V_B]]
; CHECK-NEXT: br label [[JOIN]]
; CHECK: join:
-; CHECK-NEXT: [[PHI1:%.*]] = phi i64 [ [[V_A]], [[IF]] ], [ [[V_B]], [[ELSE]] ]
-; CHECK-NEXT: [[PHI2:%.*]] = phi ptr [ [[GEP2_A]], [[IF]] ], [ [[GEP2_B]], [[ELSE]] ]
-; CHECK-NEXT: call void @use.ptr(ptr [[PHI2]])
-; CHECK-NEXT: ret i64 [[PHI1]]
+; CHECK-NEXT: [[P_SINK:%.*]] = phi ptr [ [[P]], [[ELSE]] ], [ [[GEP1_A]], [[IF]] ]
+; CHECK-NEXT: [[GEP1_B_SINK:%.*]] = phi ptr [ [[GEP1_B]], [[ELSE]] ], [ [[GEP1_A]], [[IF]] ]
+; CHECK-NEXT: [[V_B:%.*]] = load i64, ptr [[P_SINK]], align 8
+; CHECK-NEXT: [[GEP2_B:%.*]] = getelementptr i8, ptr [[GEP1_B_SINK]], i64 [[V_B]]
+; CHECK-NEXT: call void @use.ptr(ptr [[GEP2_B]])
+; CHECK-NEXT: ret i64 [[V_B]]
;
br i1 %cond, label %if, label %else
|
fhahn
reviewed
Jun 14, 2024
|
This patch is causing a 20% performance regression in Details in #96838 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Sinking currently only supports instructions that have zero or one uses. Extend this to handle instructions with any number of uses, as long as all uses are consistent (i.e. the "same" for all sinking candidates).
After #94462 this is basically just a matter of looping over all uses instead of checking the first one only.