[AArch64] Don't generate st2 for 64bit store that can use stp #69901
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
SamTebbs33,
labrinea,
jaykang10 and
sjoerdmeijer
|
@llvm/pr-subscribers-backend-aarch64 Author: David Green (davemgreen) ChangesD142966 made it so that st2 that do not start at element 0 use zip2 instead of st2. This extends that to any 64bit store that has a nearby load that can better become a LDP operation, which is expected to have a higher throughput. It searches up to 20 instructions away for a store to p+16 or p-16. Full diff: https://github.com/llvm/llvm-project/pull/69901.diff 3 Files Affected:
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index a16a102e472e709..8cf1f95c8b7ef05 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -15236,6 +15236,29 @@ bool AArch64TargetLowering::lowerInterleavedLoad(
return true;
}
+template <typename Iter>
+bool hasNearbyPairedStore(Iter It, Iter End, Value *Ptr, const DataLayout &DL) {
+ int MaxLookupDist = 20;
+ unsigned IdxWidth = DL.getIndexSizeInBits(0);
+ APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0);
+ const Value *PtrA1 =
+ Ptr->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
+
+ while (++It != End && !It->isDebugOrPseudoInst() && MaxLookupDist-- > 0) {
+ if (const auto *SI = dyn_cast<StoreInst>(&*It)) {
+ const Value *PtrB1 =
+ SI->getPointerOperand()->stripAndAccumulateInBoundsConstantOffsets(
+ DL, OffsetB);
+ if (PtrA1 == PtrB1 &&
+ (OffsetA.sextOrTrunc(IdxWidth) - OffsetB.sextOrTrunc(IdxWidth))
+ .abs() == 16)
+ return true;
+ }
+ }
+
+ return false;
+}
+
/// Lower an interleaved store into a stN intrinsic.
///
/// E.g. Lower an interleaved store (Factor = 3):
@@ -15327,8 +15350,15 @@ bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI,
return false;
}
// A 64bit st2 which does not start at element 0 will involved adding extra
- // ext elements, making the st2 unprofitable.
- if (Factor == 2 && SubVecTy->getPrimitiveSizeInBits() == 64 && Mask[0] != 0)
+ // ext elements making the st2 unprofitable, and if there is a nearby store
+ // that points to BaseAddr+16 or BaseAddr-16 then it can be better left as a
+ // zip;ldp pair which has higher throughput.
+ if (Factor == 2 && SubVecTy->getPrimitiveSizeInBits() == 64 &&
+ (Mask[0] != 0 ||
+ hasNearbyPairedStore(SI->getIterator(), SI->getParent()->end(),
+ BaseAddr, DL) ||
+ hasNearbyPairedStore(SI->getReverseIterator(), SI->getParent()->rend(),
+ BaseAddr, DL)))
return false;
Type *PtrTy = SI->getPointerOperandType();
diff --git a/llvm/test/CodeGen/AArch64/machine-cse-profitable-check.ll b/llvm/test/CodeGen/AArch64/machine-cse-profitable-check.ll
index d12240a9f4f3201..7f678d2c91dccc5 100644
--- a/llvm/test/CodeGen/AArch64/machine-cse-profitable-check.ll
+++ b/llvm/test/CodeGen/AArch64/machine-cse-profitable-check.ll
@@ -1,26 +1,14 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
-; RUN: llc -mtriple aarch64-none-linux-gnu < %s | FileCheck %s --check-prefixes=CHECK-BASE
-; RUN: llc -mtriple aarch64-none-linux-gnu -aggressive-machine-cse < %s | FileCheck %s --check-prefixes=CHECK-AGGRESSIVE-CSE
+; RUN: llc -mtriple aarch64-none-linux-gnu < %s | FileCheck %s
define void @foo(ptr %buf, <8 x i16> %a) {
-; CHECK-BASE-LABEL: foo:
-; CHECK-BASE: // %bb.0: // %entry
-; CHECK-BASE-NEXT: movi v2.2d, #0000000000000000
-; CHECK-BASE-NEXT: // kill: def $q0 killed $q0 def $q0_q1
-; CHECK-BASE-NEXT: zip2 v2.8h, v0.8h, v2.8h
-; CHECK-BASE-NEXT: movi v1.2d, #0000000000000000
-; CHECK-BASE-NEXT: st2 { v0.4h, v1.4h }, [x0], #16
-; CHECK-BASE-NEXT: str q2, [x0]
-; CHECK-BASE-NEXT: ret
-;
-; CHECK-AGGRESSIVE-CSE-LABEL: foo:
-; CHECK-AGGRESSIVE-CSE: // %bb.0: // %entry
-; CHECK-AGGRESSIVE-CSE-NEXT: // kill: def $q0 killed $q0 def $q0_q1
-; CHECK-AGGRESSIVE-CSE-NEXT: movi v1.2d, #0000000000000000
-; CHECK-AGGRESSIVE-CSE-NEXT: zip2 v2.8h, v0.8h, v1.8h
-; CHECK-AGGRESSIVE-CSE-NEXT: st2 { v0.4h, v1.4h }, [x0], #16
-; CHECK-AGGRESSIVE-CSE-NEXT: str q2, [x0]
-; CHECK-AGGRESSIVE-CSE-NEXT: ret
+; CHECK-LABEL: foo:
+; CHECK: // %bb.0: // %entry
+; CHECK-NEXT: movi v1.2d, #0000000000000000
+; CHECK-NEXT: zip1 v2.8h, v0.8h, v1.8h
+; CHECK-NEXT: zip2 v0.8h, v0.8h, v1.8h
+; CHECK-NEXT: stp q2, q0, [x0]
+; CHECK-NEXT: ret
entry:
%vzip.i = shufflevector <8 x i16> %a, <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 poison, i16 poison, i16 poison, i16 poison>, <8 x i32> <i32 0, i32 8, i32 1, i32 9, i32 2, i32 10, i32 3, i32 11>
%vzip1.i = shufflevector <8 x i16> %a, <8 x i16> <i16 poison, i16 poison, i16 poison, i16 poison, i16 0, i16 0, i16 0, i16 0>, <8 x i32> <i32 4, i32 12, i32 5, i32 13, i32 6, i32 14, i32 7, i32 15>
diff --git a/llvm/test/CodeGen/AArch64/vldn_shuffle.ll b/llvm/test/CodeGen/AArch64/vldn_shuffle.ll
index 5c73ba16972beb7..d5066aafb816d79 100644
--- a/llvm/test/CodeGen/AArch64/vldn_shuffle.ll
+++ b/llvm/test/CodeGen/AArch64/vldn_shuffle.ll
@@ -493,41 +493,38 @@ define void @transpose_s16_8x8(ptr nocapture noundef %0, ptr nocapture noundef %
define void @transpose_s16_8x8_(ptr nocapture noundef %0) {
; CHECK-LABEL: transpose_s16_8x8_:
; CHECK: // %bb.0:
-; CHECK-NEXT: mov x8, x0
+; CHECK-NEXT: ldp q0, q1, [x0]
+; CHECK-NEXT: ldp q2, q3, [x0, #32]
; CHECK-NEXT: ldp q4, q5, [x0, #64]
-; CHECK-NEXT: mov x9, x0
-; CHECK-NEXT: ldr q0, [x8, #16]!
-; CHECK-NEXT: mov x10, x0
-; CHECK-NEXT: ldr q3, [x0]
; CHECK-NEXT: ldp q6, q7, [x0, #96]
+; CHECK-NEXT: trn1 v16.8h, v0.8h, v1.8h
+; CHECK-NEXT: trn2 v0.8h, v0.8h, v1.8h
+; CHECK-NEXT: trn1 v1.8h, v2.8h, v3.8h
+; CHECK-NEXT: trn2 v2.8h, v2.8h, v3.8h
; CHECK-NEXT: trn1 v17.8h, v4.8h, v5.8h
-; CHECK-NEXT: ldr q1, [x9, #32]!
-; CHECK-NEXT: trn1 v16.8h, v3.8h, v0.8h
-; CHECK-NEXT: ldr q2, [x10, #48]!
-; CHECK-NEXT: trn2 v4.8h, v4.8h, v5.8h
-; CHECK-NEXT: trn1 v19.8h, v6.8h, v7.8h
-; CHECK-NEXT: trn2 v0.8h, v3.8h, v0.8h
-; CHECK-NEXT: trn2 v3.8h, v6.8h, v7.8h
-; CHECK-NEXT: trn1 v18.8h, v1.8h, v2.8h
-; CHECK-NEXT: trn2 v1.8h, v1.8h, v2.8h
+; CHECK-NEXT: trn2 v3.8h, v4.8h, v5.8h
+; CHECK-NEXT: trn1 v18.8h, v6.8h, v7.8h
+; CHECK-NEXT: trn2 v4.8h, v6.8h, v7.8h
; CHECK-NEXT: trn1 v5.4s, v16.4s, v17.4s
+; CHECK-NEXT: trn1 v7.4s, v0.4s, v3.4s
; CHECK-NEXT: trn2 v16.4s, v16.4s, v17.4s
-; CHECK-NEXT: trn1 v20.4s, v0.4s, v4.4s
-; CHECK-NEXT: trn1 v6.4s, v18.4s, v19.4s
-; CHECK-NEXT: trn2 v17.4s, v18.4s, v19.4s
-; CHECK-NEXT: trn2 v18.4s, v0.4s, v4.4s
-; CHECK-NEXT: trn1 v21.4s, v1.4s, v3.4s
-; CHECK-NEXT: trn2 v19.4s, v1.4s, v3.4s
-; CHECK-NEXT: zip2 v0.4s, v5.4s, v6.4s
-; CHECK-NEXT: zip2 v2.4s, v16.4s, v17.4s
-; CHECK-NEXT: st2 { v5.2s, v6.2s }, [x0]
-; CHECK-NEXT: zip2 v1.4s, v20.4s, v21.4s
-; CHECK-NEXT: zip2 v3.4s, v18.4s, v19.4s
-; CHECK-NEXT: st2 { v20.2s, v21.2s }, [x8]
-; CHECK-NEXT: st2 { v16.2s, v17.2s }, [x9]
-; CHECK-NEXT: st2 { v18.2s, v19.2s }, [x10]
-; CHECK-NEXT: stp q0, q1, [x0, #64]
-; CHECK-NEXT: stp q2, q3, [x0, #96]
+; CHECK-NEXT: trn1 v6.4s, v1.4s, v18.4s
+; CHECK-NEXT: trn1 v19.4s, v2.4s, v4.4s
+; CHECK-NEXT: trn2 v1.4s, v1.4s, v18.4s
+; CHECK-NEXT: trn2 v0.4s, v0.4s, v3.4s
+; CHECK-NEXT: trn2 v2.4s, v2.4s, v4.4s
+; CHECK-NEXT: zip1 v3.4s, v5.4s, v6.4s
+; CHECK-NEXT: zip1 v4.4s, v7.4s, v19.4s
+; CHECK-NEXT: zip1 v17.4s, v16.4s, v1.4s
+; CHECK-NEXT: zip1 v18.4s, v0.4s, v2.4s
+; CHECK-NEXT: zip2 v5.4s, v5.4s, v6.4s
+; CHECK-NEXT: zip2 v1.4s, v16.4s, v1.4s
+; CHECK-NEXT: zip2 v0.4s, v0.4s, v2.4s
+; CHECK-NEXT: stp q3, q4, [x0]
+; CHECK-NEXT: zip2 v3.4s, v7.4s, v19.4s
+; CHECK-NEXT: stp q17, q18, [x0, #32]
+; CHECK-NEXT: stp q1, q0, [x0, #96]
+; CHECK-NEXT: stp q5, q3, [x0, #64]
; CHECK-NEXT: ret
%2 = load <8 x i16>, ptr %0, align 16
%3 = getelementptr inbounds <8 x i16>, ptr %0, i64 1
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
D142966 made it so that st2 that do not start at element 0 use zip2 instead of st2. This extends that to any 64bit store that has a nearby load that can better become a LDP operation, which is expected to have a higher throughput. It searches up to 20 instructions away for a store to p+16 or p-16.
davemgreen
force-pushed
the
gh-a64-st2stp
branch
from
d4cbab1
to
8bca87a
Compare
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.
D142966 made it so that st2 that do not start at element 0 use zip2 instead of st2. This extends that to any 64bit store that has a nearby load that can better become a LDP operation, which is expected to have a higher throughput. It searches up to 20 instructions away for a store to p+16 or p-16.