Skip to content

InstCombine: sink loads with invariant.load metadata #112692

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
artagnon merged 2 commits into from
Oct 18, 2024
Merged

InstCombine: sink loads with invariant.load metadata #112692

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
90d75e5
InstCombine: sink loads with invariant.load metadata
artagnon Oct 17, 2024
af479a1
sink_instruction: improve test
artagnon Oct 18, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4822,7 +4822,8 @@ bool InstCombinerImpl::tryToSinkInstruction(Instruction *I,

// We can only sink load instructions if there is nothing between the load and
// the end of block that could change the value.
if (I->mayReadFromMemory()) {
if (I->mayReadFromMemory() &&
!I->hasMetadata(LLVMContext::MD_invariant_load)) {
// We don't want to do any sophisticated alias analysis, so we only check
// the instructions after I in I's parent block if we try to sink to its
// successor block.
Expand Down
123 changes: 119 additions & 4 deletions llvm/test/Transforms/InstCombine/sink_instruction.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,8 +86,8 @@ define i32 @test3(ptr nocapture readonly %P, i32 %i) {
; CHECK-LABEL: @test3(
; CHECK-NEXT: entry:
; CHECK-NEXT: switch i32 [[I:%.*]], label [[SW_EPILOG:%.*]] [
; CHECK-NEXT: i32 5, label [[SW_BB:%.*]]
; CHECK-NEXT: i32 2, label [[SW_BB]]
; CHECK-NEXT: i32 5, label [[SW_BB:%.*]]
; CHECK-NEXT: i32 2, label [[SW_BB]]
; CHECK-NEXT: ]
; CHECK: sw.bb:
; CHECK-NEXT: [[IDXPROM:%.*]] = sext i32 [[I]] to i64
Expand Down Expand Up @@ -190,8 +190,8 @@ define i32 @test6(ptr nocapture readonly %P, i32 %i, i1 %cond) {
; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[P:%.*]], i64 [[IDXPROM]]
; CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; CHECK-NEXT: switch i32 [[I]], label [[SW_BB:%.*]] [
; CHECK-NEXT: i32 5, label [[SW_EPILOG:%.*]]
; CHECK-NEXT: i32 2, label [[SW_EPILOG]]
; CHECK-NEXT: i32 5, label [[SW_EPILOG:%.*]]
; CHECK-NEXT: i32 2, label [[SW_EPILOG]]
; CHECK-NEXT: ]
; CHECK: sw.bb:
; CHECK-NEXT: br label [[SW_EPILOG]]
Expand Down Expand Up @@ -272,3 +272,118 @@ abort:
call void @abort()
unreachable
}

; Loads marked invariant can be sunk across critical edges.

define <4 x float> @invariant_load_metadata(ptr %p, i32 %cond) {
; CHECK-LABEL: @invariant_load_metadata(
; CHECK-NEXT: entry:
; CHECK-NEXT: [[C:%.*]] = icmp eq i32 [[COND:%.*]], 0
; CHECK-NEXT: br i1 [[C]], label [[BLOCK:%.*]], label [[END:%.*]]
; CHECK: block:
; CHECK-NEXT: call void @fn()
; CHECK-NEXT: br label [[END]]
; CHECK: end:
; CHECK-NEXT: [[V:%.*]] = load <4 x float>, ptr [[P:%.*]], align 16, !invariant.load [[META0:![0-9]+]]
; CHECK-NEXT: ret <4 x float> [[V]]
;
entry:
%v = load <4 x float>, ptr %p, !invariant.load !0
%c = icmp eq i32 %cond, 0
br i1 %c, label %block, label %end
block:
call void @fn()
br label %end
end:
ret <4 x float> %v
}

; Loads not marked invariant cannot be sunk across critical edges.

define <4 x float> @invariant_load_neg(ptr %p, i32 %cond) {
; CHECK-LABEL: @invariant_load_neg(
; CHECK-NEXT: entry:
; CHECK-NEXT: [[V:%.*]] = load <4 x float>, ptr [[P:%.*]], align 16
; CHECK-NEXT: [[C:%.*]] = icmp eq i32 [[COND:%.*]], 0
; CHECK-NEXT: br i1 [[C]], label [[BLOCK:%.*]], label [[END:%.*]]
; CHECK: block:
; CHECK-NEXT: call void @fn()
; CHECK-NEXT: br label [[END]]
; CHECK: end:
; CHECK-NEXT: ret <4 x float> [[V]]
;
entry:
%v = load <4 x float>, ptr %p
%c = icmp eq i32 %cond, 0
br i1 %c, label %block, label %end
block:
call void @fn()
br label %end
end:
ret <4 x float> %v
}

; Loads that aren't marked invariant but used in one branch
; can be sunk to that branch.

define void @invariant_load_use_in_br(ptr %p, i1 %cond) {
; CHECK-LABEL: @invariant_load_use_in_br(
; CHECK-NEXT: entry:
; CHECK-NEXT: br i1 [[COND:%.*]], label [[TRUE_BR:%.*]], label [[FALSE_BR:%.*]]
; CHECK: true.br:
; CHECK-NEXT: call void @fn()
; CHECK-NEXT: br label [[EXIT:%.*]]
; CHECK: false.br:
; CHECK-NEXT: [[VAL:%.*]] = load i32, ptr [[P:%.*]], align 4
; CHECK-NEXT: call void @fn(i32 [[VAL]])
; CHECK-NEXT: br label [[EXIT]]
; CHECK: exit:
; CHECK-NEXT: ret void
;
entry:
%val = load i32, ptr %p
br i1 %cond, label %true.br, label %false.br
true.br:
call void @fn()
br label %exit
false.br:
call void @fn(i32 %val)
br label %exit
exit:
ret void
}

; Invariant loads marked with metadata can be sunk past calls.

define void @invariant_load_metadata_call(ptr %p, i1 %cond) {
; CHECK-LABEL: @invariant_load_metadata_call(
; CHECK-NEXT: entry:
; CHECK-NEXT: call void @fn()
; CHECK-NEXT: br i1 [[COND:%.*]], label [[TRUE_BR:%.*]], label [[FALSE_BR:%.*]]
; CHECK: true.br:
; CHECK-NEXT: call void @fn()
; CHECK-NEXT: br label [[EXIT:%.*]]
; CHECK: false.br:
; CHECK-NEXT: [[VAL:%.*]] = load i32, ptr [[P:%.*]], align 4, !invariant.load [[META0]]
; CHECK-NEXT: call void @fn(i32 [[VAL]])
; CHECK-NEXT: br label [[EXIT]]
; CHECK: exit:
; CHECK-NEXT: ret void
;
entry:
%val = load i32, ptr %p, !invariant.load !0
call void @fn()
br i1 %cond, label %true.br, label %false.br
true.br:
call void @fn()
br label %exit
false.br:
call void @fn(i32 %val)
br label %exit
exit:
ret void
}

declare void @fn()

!0 = !{}