[BasicAA] Scalable offset with scalable typesize. #80818
Merged
+32
−11
Conversation
This file contains 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
davemgreen
requested review from
preames,
harviniriawan,
efriedma-quic and
david-arm
|
@llvm/pr-subscribers-llvm-analysis Author: David Green (davemgreen) ChangesThis patch adds a simple alias analysis check for accesses that are scalable Full diff: https://github.com/llvm/llvm-project/pull/80818.diff 2 Files Affected:
diff --git a/llvm/lib/Analysis/BasicAliasAnalysis.cpp b/llvm/lib/Analysis/BasicAliasAnalysis.cpp
index 19c4393add6ab..4898dae3dbe08 100644
--- a/llvm/lib/Analysis/BasicAliasAnalysis.cpp
+++ b/llvm/lib/Analysis/BasicAliasAnalysis.cpp
@@ -1170,6 +1170,25 @@ AliasResult BasicAAResult::aliasGEP(
}
}
+ // VScale Alias Analysis - Given one scalable offset between accesses and a
+ // scalable typesize, we can divide each side by vscale, treating both values
+ // as a constant. We prove that Offset/vscale >= TypeSize/vscale.
+ if (DecompGEP1.VarIndices.size() == 1 && DecompGEP1.Offset.isZero() &&
+ PatternMatch::match(DecompGEP1.VarIndices[0].Val.V,
+ PatternMatch::m_VScale())) {
+ const VariableGEPIndex &ScalableVar = DecompGEP1.VarIndices[0];
+ APInt Scale =
+ ScalableVar.IsNegated ? -ScalableVar.Scale : ScalableVar.Scale;
+ LocationSize VLeftSize = Scale.isNegative() ? V1Size : V2Size;
+
+ // Note that we do not check that the typesize is scalable, as vscale >= 1
+ // so noalias still holds so long as the dependency distance is at least as
+ // big as the typesize.
+ if (VLeftSize.hasValue() &&
+ Scale.uge(VLeftSize.getValue().getKnownMinValue()))
+ return AliasResult::NoAlias;
+ }
+
// Bail on analysing scalable LocationSize
if (V1Size.isScalable() || V2Size.isScalable())
return AliasResult::MayAlias;
diff --git a/llvm/test/Analysis/BasicAA/vscale.ll b/llvm/test/Analysis/BasicAA/vscale.ll
index 3fff435463a6d..a8ed93be90ef4 100644
--- a/llvm/test/Analysis/BasicAA/vscale.ll
+++ b/llvm/test/Analysis/BasicAA/vscale.ll
@@ -309,6 +309,174 @@ define void @v1v2types(ptr %p) vscale_range(1,16) {
ret void
}
+; VScale intrinsic offset tests
+
+; CHECK-LABEL: vscale_neg_notscalable
+; CHECK-DAG: NoAlias: <4 x i32>* %p, <4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16, <4 x i32>* %p
+; CHECK-DAG: MayAlias: <4 x i32>* %m16, <4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <4 x i32>* %p, <4 x i32>* %vm16m16
+; CHECK-DAG: NoAlias: <4 x i32>* %vm16, <4 x i32>* %vm16m16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16, <4 x i32>* %vm16m16
+; CHECK-DAG: MayAlias: <4 x i32>* %m16pv16, <4 x i32>* %p
+; CHECK-DAG: NoAlias: <4 x i32>* %m16pv16, <4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16, <4 x i32>* %m16pv16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16pv16, <4 x i32>* %vm16m16
+define void @vscale_neg_notscalable(ptr %p) {
+ %v = call i64 @llvm.vscale.i64()
+ %vp = mul i64 %v, 16
+ %vm = mul i64 %v, -16
+ %vm16 = getelementptr i8, ptr %p, i64 %vm
+ %m16 = getelementptr <4 x i32>, ptr %p, i64 -1
+ %vm16m16 = getelementptr <4 x i32>, ptr %vm16, i64 -1
+ %m16pv16 = getelementptr i8, ptr %m16, i64 %vp
+ load <4 x i32>, ptr %p
+ load <4 x i32>, ptr %vm16
+ load <4 x i32>, ptr %m16
+ load <4 x i32>, ptr %vm16m16
+ load <4 x i32>, ptr %m16pv16
+ ret void
+}
+
+; CHECK-LABEL: vscale_neg_scalable
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %p, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %p
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %p, <vscale x 4 x i32>* %vm16m16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %vm16, <vscale x 4 x i32>* %vm16m16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %vm16m16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16pv16, <vscale x 4 x i32>* %p
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16pv16, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %m16pv16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %m16pv16, <vscale x 4 x i32>* %vm16m16
+define void @vscale_neg_scalable(ptr %p) {
+ %v = call i64 @llvm.vscale.i64()
+ %vp = mul i64 %v, 16
+ %vm = mul i64 %v, -16
+ %vm16 = getelementptr i8, ptr %p, i64 %vm
+ %m16 = getelementptr <4 x i32>, ptr %p, i64 -1
+ %vm16m16 = getelementptr <4 x i32>, ptr %vm16, i64 -1
+ %m16pv16 = getelementptr i8, ptr %m16, i64 %vp
+ load <vscale x 4 x i32>, ptr %p
+ load <vscale x 4 x i32>, ptr %vm16
+ load <vscale x 4 x i32>, ptr %m16
+ load <vscale x 4 x i32>, ptr %vm16m16
+ load <vscale x 4 x i32>, ptr %m16pv16
+ ret void
+}
+
+; CHECK-LABEL: vscale_pos_notscalable
+; CHECK-DAG: NoAlias: <4 x i32>* %p, <4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16, <4 x i32>* %p
+; CHECK-DAG: MayAlias: <4 x i32>* %m16, <4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <4 x i32>* %p, <4 x i32>* %vm16m16
+; CHECK-DAG: NoAlias: <4 x i32>* %vm16, <4 x i32>* %vm16m16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16, <4 x i32>* %vm16m16
+; CHECK-DAG: MayAlias: <4 x i32>* %m16pv16, <4 x i32>* %p
+; CHECK-DAG: NoAlias: <4 x i32>* %m16pv16, <4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16, <4 x i32>* %m16pv16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16pv16, <4 x i32>* %vm16m16
+define void @vscale_pos_notscalable(ptr %p) {
+ %v = call i64 @llvm.vscale.i64()
+ %vp = mul i64 %v, 16
+ %vm = mul i64 %v, -16
+ %vm16 = getelementptr i8, ptr %p, i64 %vp
+ %m16 = getelementptr <4 x i32>, ptr %p, i64 1
+ %vm16m16 = getelementptr <4 x i32>, ptr %vm16, i64 1
+ %m16pv16 = getelementptr i8, ptr %m16, i64 %vm
+ load <4 x i32>, ptr %p
+ load <4 x i32>, ptr %vm16
+ load <4 x i32>, ptr %m16
+ load <4 x i32>, ptr %vm16m16
+ load <4 x i32>, ptr %m16pv16
+ ret void
+}
+
+; CHECK-LABEL: vscale_pos_scalable
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %p, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %p
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %p, <vscale x 4 x i32>* %vm16m16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %vm16, <vscale x 4 x i32>* %vm16m16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %vm16m16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16pv16, <vscale x 4 x i32>* %p
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16pv16, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %m16pv16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %m16pv16, <vscale x 4 x i32>* %vm16m16
+define void @vscale_pos_scalable(ptr %p) {
+ %v = call i64 @llvm.vscale.i64()
+ %vp = mul i64 %v, 16
+ %vm = mul i64 %v, -16
+ %vm16 = getelementptr i8, ptr %p, i64 %vp
+ %m16 = getelementptr <4 x i32>, ptr %p, i64 1
+ %vm16m16 = getelementptr <4 x i32>, ptr %vm16, i64 1
+ %m16pv16 = getelementptr i8, ptr %m16, i64 %vm
+ load <vscale x 4 x i32>, ptr %p
+ load <vscale x 4 x i32>, ptr %vm16
+ load <vscale x 4 x i32>, ptr %m16
+ load <vscale x 4 x i32>, ptr %vm16m16
+ load <vscale x 4 x i32>, ptr %m16pv16
+ ret void
+}
+
+; CHECK-LABEL: vscale_v1v2types
+; CHECK-DAG: MustAlias: <4 x i32>* %p, <vscale x 4 x i32>* %p
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %p, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <4 x i32>* %p, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %p, <4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <4 x i32>* %p, <4 x i32>* %vm16
+; CHECK-DAG: MustAlias: <4 x i32>* %vm16, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %p
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <4 x i32>* %p
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <4 x i32>* %vm16
+; CHECK-DAG: NoAlias: <4 x i32>* %m16, <vscale x 4 x i32>* %p
+; CHECK-DAG: NoAlias: <4 x i32>* %m16, <4 x i32>* %p
+; CHECK-DAG: MayAlias: <4 x i32>* %m16, <vscale x 4 x i32>* %vm16
+; CHECK-DAG: MayAlias: <4 x i32>* %m16, <4 x i32>* %vm16
+; CHECK-DAG: MustAlias: <4 x i32>* %m16, <vscale x 4 x i32>* %m16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %p, <vscale x 4 x i32>* %vp16
+; CHECK-DAG: NoAlias: <4 x i32>* %p, <vscale x 4 x i32>* %vp16
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %vm16, <vscale x 4 x i32>* %vp16
+; CHECK-DAG: NoAlias: <4 x i32>* %vm16, <vscale x 4 x i32>* %vp16
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %m16, <vscale x 4 x i32>* %vp16
+; CHECK-DAG: MayAlias: <4 x i32>* %m16, <vscale x 4 x i32>* %vp16
+define void @vscale_v1v2types(ptr %p) {
+ %v = call i64 @llvm.vscale.i64()
+ %vp = mul i64 %v, 16
+ %vm = mul i64 %v, -16
+ %vp16 = getelementptr i8, ptr %p, i64 %vp
+ %vm16 = getelementptr i8, ptr %p, i64 %vm
+ %m16 = getelementptr <4 x i32>, ptr %p, i64 -1
+ load <vscale x 4 x i32>, ptr %p
+ load <4 x i32>, ptr %p
+ load <vscale x 4 x i32>, ptr %vm16
+ load <4 x i32>, ptr %vm16
+ load <vscale x 4 x i32>, ptr %m16
+ load <4 x i32>, ptr %m16
+ load <vscale x 4 x i32>, ptr %vp16
+ ret void
+}
+
+; CHECK-LABEL: twovscales
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %vp161, <vscale x 4 x i32>* %vp162
+; CHECK-DAG: NoAlias: <vscale x 4 x i32>* %vp161, <vscale x 4 x i32>* %vp161b
+; CHECK-DAG: MayAlias: <vscale x 4 x i32>* %vp161b, <vscale x 4 x i32>* %vp162
+define void @twovscales(ptr %p) {
+ %v1 = call i64 @llvm.vscale.i64()
+ %v2 = call i64 @llvm.vscale.i64()
+ %vp1 = mul i64 %v1, 16
+ %vp2 = mul i64 %v2, 16
+ %vp3 = mul i64 %v1, 17
+ %vp161 = getelementptr i8, ptr %p, i64 %vp1
+ %vp162 = getelementptr i8, ptr %p, i64 %vp2
+ %vp161b = getelementptr i8, ptr %vp161, i64 %vp3
+ load <vscale x 4 x i32>, ptr %vp161
+ load <vscale x 4 x i32>, ptr %vp162
+ load <vscale x 4 x i32>, ptr %vp161b
+ ret void
+}
+
; getelementptr recursion
; CHECK-LABEL: gep_recursion_level_1
|
nikic
reviewed
Feb 6, 2024
Hmm. Yeah maybe. I was thinking of offsets similar to typesizes, and it would be a strange access than might wrap around depending on the vscale. I'll add a IsNSW check if you think that's better - it looks like a few cases in the tests get a little worse, but we might be able to improve it in the future with vscale_range to show the extreme would not overflow. |
davemgreen
force-pushed
the
gh-aa-vscalexvscale
branch
from
4004e4d
to
b1a2dbe
Compare
|
Added a check for |
nikic
approved these changes
Feb 7, 2024
| if (DecompGEP1.VarIndices.size() == 1 && DecompGEP1.VarIndices[0].IsNSW && | ||
| DecompGEP1.Offset.isZero() && | ||
| PatternMatch::match(DecompGEP1.VarIndices[0].Val.V, | ||
| PatternMatch::m_VScale())) { |
There was a problem hiding this comment.
I think we also need DecompGEP1.VarIndices[0].Val.TruncBits == 0.
This patch adds a simple alias analysis check for accesses that are scalable with a offset between them that is also trivially scalable (there are no other constant/variable offsets). We essentially divide each side by vscale and are left needing to check that the offset >= typesize.
davemgreen
force-pushed
the
gh-aa-vscalexvscale
branch
from
b1a2dbe
to
399eee3
Compare
davemgreen
added a commit
to davemgreen/llvm-project
that referenced
this pull request
Feb 8, 2024
This extends llvm#80818 when IsNSW is lost (possibly due to looking through multiple GEPs), to check the vscale_range for an access that will not overflow even with the maximum range.
davemgreen
added a commit
to davemgreen/llvm-project
that referenced
this pull request
Feb 8, 2024
This is a fix for llvm#80818, as pointed out in llvm#81144 it should be checking the abs of Scale. The added test changes from NoAlias to MayAlias.
davemgreen
added a commit
that referenced
this pull request
Feb 9, 2024
davemgreen
added a commit
to davemgreen/llvm-project
that referenced
this pull request
Feb 9, 2024
This extends llvm#80818 when IsNSW is lost (possibly due to looking through multiple GEPs), to check the vscale_range for an access that will not overflow even with the maximum range.
davemgreen
added a commit
to davemgreen/llvm-project
that referenced
this pull request
Feb 12, 2024
This extends llvm#80818 when IsNSW is lost (possibly due to looking through multiple GEPs), to check the vscale_range for an access that will not overflow even with the maximum range.
davemgreen
added a commit
that referenced
this pull request
Feb 12, 2024
This extends #80818 when IsNSW is lost (possibly due to looking through multiple GEPs), to check the vscale_range for an access that will not overflow even with the maximum range.
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.
This patch adds a simple alias analysis check for accesses that are scalable
with a offset between them that is also trivially scalable (there are no other
constant/variable offsets). We essentially divide each side by vscale and are
left needing to check that the offset >= typesize.