[CVP][LVI] Add support for InsertElementInst in LVI #99368
Merged
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
llvmbot
added
llvm:analysis
|
@llvm/pr-subscribers-llvm-analysis @llvm/pr-subscribers-llvm-transforms Author: Rajat Bajpai (rajatbajpai) ChangesCurrently, the LVI analysis pass doesn't support InsertElementInst vector instruction. Due to this, some optimization opportunities are missed. For example, in the below example, ICMP instruction can be folded but it doesn't. This change adds InsertElementInst support in the LVI analysis pass to fix the motivating example. Full diff: https://github.com/llvm/llvm-project/pull/99368.diff 2 Files Affected:
diff --git a/llvm/lib/Analysis/LazyValueInfo.cpp b/llvm/lib/Analysis/LazyValueInfo.cpp
index 92389f2896b8e..d28d4fa47fdae 100644
--- a/llvm/lib/Analysis/LazyValueInfo.cpp
+++ b/llvm/lib/Analysis/LazyValueInfo.cpp
@@ -428,6 +428,8 @@ class LazyValueInfoImpl {
std::optional<ValueLatticeElement> solveBlockValueIntrinsic(IntrinsicInst *II,
BasicBlock *BB);
std::optional<ValueLatticeElement>
+ solveBlockValueInsertElement(InsertElementInst *IEI, BasicBlock *BB);
+ std::optional<ValueLatticeElement>
solveBlockValueExtractValue(ExtractValueInst *EVI, BasicBlock *BB);
bool isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB);
void intersectAssumeOrGuardBlockValueConstantRange(Value *Val,
@@ -657,6 +659,9 @@ LazyValueInfoImpl::solveBlockValueImpl(Value *Val, BasicBlock *BB) {
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BBI))
return solveBlockValueBinaryOp(BO, BB);
+ if (auto *IEI = dyn_cast<InsertElementInst>(BBI))
+ return solveBlockValueInsertElement(IEI, BB);
+
if (auto *EVI = dyn_cast<ExtractValueInst>(BBI))
return solveBlockValueExtractValue(EVI, BB);
@@ -1038,6 +1043,37 @@ LazyValueInfoImpl::solveBlockValueIntrinsic(IntrinsicInst *II, BasicBlock *BB) {
MetadataVal);
}
+std::optional<ValueLatticeElement>
+LazyValueInfoImpl::solveBlockValueInsertElement(InsertElementInst *IEI,
+ BasicBlock *BB) {
+ std::optional<ValueLatticeElement> OptEltVal =
+ getBlockValue(IEI->getOperand(1), BB, IEI);
+ if (!OptEltVal)
+ return std::nullopt;
+ ValueLatticeElement &EltVal = *OptEltVal;
+
+ if (auto *CV = dyn_cast<ConstantVector>(IEI->getOperand(0))) {
+ // Must be vector of integers. Merge these elements to create
+ // the range.
+ for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
+ Constant *Elem = CV->getAggregateElement(i);
+ if (isa<PoisonValue>(Elem))
+ continue;
+ std::optional<ConstantRange> CR = getRangeFor(Elem, IEI, BB);
+ if (!CR)
+ return std::nullopt;
+ EltVal.mergeIn(ValueLatticeElement::getRange(*CR));
+ }
+ } else if (!isa<PoisonValue>(IEI->getOperand(0))) {
+ std::optional<ValueLatticeElement> OptVecResult =
+ solveBlockValueImpl(IEI->getOperand(0), BB);
+ if (!OptVecResult)
+ return std::nullopt;
+ EltVal.mergeIn(*OptVecResult);
+ }
+ return EltVal;
+}
+
std::optional<ValueLatticeElement>
LazyValueInfoImpl::solveBlockValueExtractValue(ExtractValueInst *EVI,
BasicBlock *BB) {
diff --git a/llvm/test/Transforms/CorrelatedValuePropagation/insertelement.ll b/llvm/test/Transforms/CorrelatedValuePropagation/insertelement.ll
new file mode 100644
index 0000000000000..769f431738342
--- /dev/null
+++ b/llvm/test/Transforms/CorrelatedValuePropagation/insertelement.ll
@@ -0,0 +1,76 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt < %s -passes=correlated-propagation -S | FileCheck %s
+
+;; Check if ICMP instruction is constant folded or not.
+
+define void @test1(ptr addrspace(1) %out) {
+; CHECK-LABEL: define void @test1(
+; CHECK-SAME: ptr addrspace(1) [[OUT:%.*]]) {
+; CHECK-NEXT: [[CALL:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x(), !range [[RNG0:![0-9]+]]
+; CHECK-NEXT: [[UDIV_LHS_TRUNC:%.*]] = trunc i32 [[CALL]] to i16
+; CHECK-NEXT: [[UDIV1:%.*]] = udiv i16 [[UDIV_LHS_TRUNC]], 5
+; CHECK-NEXT: [[UDIV_ZEXT:%.*]] = zext i16 [[UDIV1]] to i32
+; CHECK-NEXT: [[ADD1:%.*]] = add nuw nsw i32 [[UDIV_ZEXT]], 768
+; CHECK-NEXT: [[ADD2:%.*]] = add nuw nsw i32 [[UDIV_ZEXT]], 896
+; CHECK-NEXT: [[IE1:%.*]] = insertelement <2 x i32> poison, i32 [[ADD1]], i64 0
+; CHECK-NEXT: [[IE2:%.*]] = insertelement <2 x i32> [[IE1]], i32 [[ADD2]], i64 1
+; CHECK-NEXT: [[EI1:%.*]] = extractelement <2 x i1> <i1 true, i1 true>, i64 0
+; CHECK-NEXT: [[EI2:%.*]] = extractelement <2 x i1> <i1 true, i1 true>, i64 1
+; CHECK-NEXT: [[ADDUP:%.*]] = add i1 [[EI1]], [[EI2]]
+; CHECK-NEXT: [[ADDUP_UPCAST:%.*]] = zext i1 [[ADDUP]] to i32
+; CHECK-NEXT: store i32 [[ADDUP_UPCAST]], ptr addrspace(1) [[OUT]], align 4
+; CHECK-NEXT: ret void
+;
+ %call = call i32 @llvm.nvvm.read.ptx.sreg.tid.x(), !range !1
+ %udiv = udiv i32 %call, 5
+ %add1 = add i32 %udiv, 768
+ %add2 = add i32 %udiv, 896
+ %ie1 = insertelement <2 x i32> poison, i32 %add1, i64 0
+ %ie2 = insertelement <2 x i32> %ie1, i32 %add2, i64 1
+ %icmp1 = icmp slt <2 x i32> %ie2, <i32 1024, i32 1024>
+ %ei1 = extractelement <2 x i1> %icmp1, i64 0
+ %ei2 = extractelement <2 x i1> %icmp1, i64 1
+ %addUp = add i1 %ei1, %ei2
+ %addUp.upcast = zext i1 %addUp to i32
+ store i32 %addUp.upcast, ptr addrspace(1) %out, align 4
+ ret void
+}
+
+
+;; Check if LVI is able to handle constant vector operands
+;; in InsertElementInst and CVP is able to fold ICMP instruction.
+
+define void @test2(ptr addrspace(1) %out) {
+; CHECK-LABEL: define void @test2(
+; CHECK-SAME: ptr addrspace(1) [[OUT:%.*]]) {
+; CHECK-NEXT: [[CALL:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x(), !range [[RNG0]]
+; CHECK-NEXT: [[UDIV_LHS_TRUNC:%.*]] = trunc i32 [[CALL]] to i16
+; CHECK-NEXT: [[UDIV1:%.*]] = udiv i16 [[UDIV_LHS_TRUNC]], 5
+; CHECK-NEXT: [[UDIV_ZEXT:%.*]] = zext i16 [[UDIV1]] to i32
+; CHECK-NEXT: [[ADD2:%.*]] = add nuw nsw i32 [[UDIV_ZEXT]], 896
+; CHECK-NEXT: [[IE1:%.*]] = insertelement <2 x i32> <i32 poison, i32 1>, i32 [[ADD2]], i64 0
+; CHECK-NEXT: [[EI1:%.*]] = extractelement <2 x i1> <i1 true, i1 true>, i64 0
+; CHECK-NEXT: [[EI2:%.*]] = extractelement <2 x i1> <i1 true, i1 true>, i64 1
+; CHECK-NEXT: [[ADDUP:%.*]] = add i1 [[EI1]], [[EI2]]
+; CHECK-NEXT: [[ADDUP_UPCAST:%.*]] = zext i1 [[ADDUP]] to i32
+; CHECK-NEXT: store i32 [[ADDUP_UPCAST]], ptr addrspace(1) [[OUT]], align 4
+; CHECK-NEXT: ret void
+;
+ %call = call i32 @llvm.nvvm.read.ptx.sreg.tid.x(), !range !1
+ %udiv = udiv i32 %call, 5
+ %add2 = add i32 %udiv, 896
+ %ie1 = insertelement <2 x i32> <i32 poison, i32 1>, i32 %add2, i64 0
+ %icmp1 = icmp slt <2 x i32> %ie1, <i32 1024, i32 1024>
+ %ei1 = extractelement <2 x i1> %icmp1, i64 0
+ %ei2 = extractelement <2 x i1> %icmp1, i64 1
+ %addUp = add i1 %ei1, %ei2
+ %addUp.upcast = zext i1 %addUp to i32
+ store i32 %addUp.upcast, ptr addrspace(1) %out, align 4
+ ret void
+}
+
+
+!1 = !{i32 0, i32 640}
+;.
+; CHECK: [[RNG0]] = !{i32 0, i32 640}
+;.
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
nikic
reviewed
Jul 17, 2024
llvm/test/Transforms/CorrelatedValuePropagation/insertelement.ll
Outdated
Show resolved
Hide resolved
rajatbajpai
force-pushed
the
dev/rbajpai/lvi_insertelement
branch
2 times, most recently
from
4e223c3 to
5e2a073
Compare
rajatbajpai
force-pushed
the
dev/rbajpai/lvi_insertelement
branch
from
5e2a073 to
0dfc0f9
Compare
|
Will merge the changes after getting the clean pipeline. Thank you for reviewing this change. |
rajatbajpai
force-pushed
the
dev/rbajpai/lvi_insertelement
branch
from
0dfc0f9 to
e6145b4
Compare
Currently, the LVI analysis pass doesn't support InsertElementInst vector instruction. Due to this, some optimization opportunities are missed. For example, in the below example, ICMP instruction can be folded but it doesn't. ``` ... %ie1 = insertelement <2 x i32> poison, i32 10, i64 0 %ie2 = insertelement <2 x i32> %ie1, i32 20, i64 1 %icmp1 = icmp <2 x i1> %ie2, <i32 40, i32 40> ... ``` This change adds InsertElementInst support in the LVI analysis pass to fix the motivating example.
rajatbajpai
force-pushed
the
dev/rbajpai/lvi_insertelement
branch
from
e6145b4 to
530f0c6
Compare
|
The Linux build is failing and the error doesn't seem to be related to my change. I'll wait some time for the setup to recover. |
|
@rajatbajpai You can ignore the linux build. |
|
@nikic could you please merge this PR for me? I don't have permission to merge a PR. Thanks! |
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/56/builds/2815 Here is the relevant piece of the build log for the reference: |
yuxuanchen1997
pushed a commit
that referenced
this pull request
Jul 25, 2024
Summary: Currently, the LVI analysis pass doesn't support InsertElementInst vector instruction. Due to this, some optimization opportunities are missed. For example, in the below example, ICMP instruction can be folded but it doesn't. ``` ... %ie1 = insertelement <2 x i32> poison, i32 10, i64 0 %ie2 = insertelement <2 x i32> %ie1, i32 20, i64 1 %icmp = icmp <2 x i1> %ie2, <i32 40, i32 40> ... ``` This change adds InsertElementInst support in the LVI analysis pass to fix the motivating example. Test Plan: Reviewers: Subscribers: Tasks: Tags: Differential Revision: https://phabricator.intern.facebook.com/D60251386
|
Hello, I bisected a crash back to this patch: |
|
Reduced test case: @g = external global i32
define <2 x i16> @test() {
%ins = insertelement <2 x i16> poison, i16 ptrtoint (ptr @g to i16), i32 0
ret <2 x i16> %ins
} |
|
@mikaelholmen Should be fixed by 7a7a426. |
|
That was quick. Thanks! |
|
Thanks @mikaelholmen, for reporting the issue and @nikic for a quick fix. |
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.
Currently, the LVI analysis pass doesn't support InsertElementInst vector instruction. Due to this, some optimization opportunities are missed. For example, in the below example, ICMP instruction can be folded but it doesn't.
This change adds InsertElementInst support in the LVI analysis pass to fix the motivating example.