[llvm] [AArch64] Set MaxInterleaving to 4 for Neoverse V2 (PR #100385)

[Sjoerd Meijer via llvm-commits]

https://github.com/sjoerdmeijer updated https://github.com/llvm/llvm-project/pull/100385

>From 0160e7ac7f28c7256f123fe360086a6fd802085a Mon Sep 17 00:00:00 2001
From: Sjoerd Meijer <smeijer at nvidia.com>
Date: Wed, 24 Jul 2024 17:49:43 +0530
Subject: [PATCH] [AArch64] Set MaxInterleaving to 4 for Neoverse V2 and V3

This helps loop based benchmarks quite a lot, SPEC INT is unaffected.
---
 llvm/lib/Target/AArch64/AArch64Subtarget.cpp              | 6 ++++--
 .../LoopVectorize/AArch64/interleaving-load-store.ll      | 8 ++++++++
 .../LoopVectorize/AArch64/interleaving-reduction.ll       | 8 ++++++++
 3 files changed, 20 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64Subtarget.cpp b/llvm/lib/Target/AArch64/AArch64Subtarget.cpp
index 32a355fe38f1c..bb88197edd301 100644
--- a/llvm/lib/Target/AArch64/AArch64Subtarget.cpp
+++ b/llvm/lib/Target/AArch64/AArch64Subtarget.cpp
@@ -233,10 +233,12 @@ void AArch64Subtarget::initializeProperties(bool HasMinSize) {
     PrefLoopAlignment = Align(32);
     MaxBytesForLoopAlignment = 16;
     break;
-  case NeoverseN2:
-  case NeoverseN3:
   case NeoverseV2:
   case NeoverseV3:
+    MaxInterleaveFactor = 4;
+    LLVM_FALLTHROUGH;
+  case NeoverseN2:
+  case NeoverseN3:
     PrefFunctionAlignment = Align(16);
     PrefLoopAlignment = Align(32);
     MaxBytesForLoopAlignment = 16;
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleaving-load-store.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleaving-load-store.ll
index 0e54bd15e5ea5..0441e91c9d1d3 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/interleaving-load-store.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleaving-load-store.ll
@@ -5,6 +5,8 @@
 ; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-macos -mcpu=apple-a14 -S %s | FileCheck --check-prefix=INTERLEAVE-4 %s
 ; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-macos -mcpu=apple-a15 -S %s | FileCheck --check-prefix=INTERLEAVE-4 %s
 ; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-macos -mcpu=apple-a16 -S %s | FileCheck --check-prefix=INTERLEAVE-4 %s
+; RUN: opt -passes=loop-vectorize -mtriple=arm64 -mcpu=neoverse-v2 -S %s | FileCheck --check-prefix=INTERLEAVE-4 %s
+; RUN: opt -passes=loop-vectorize -mtriple=arm64 -mcpu=neoverse-v3 -S %s | FileCheck --check-prefix=INTERLEAVE-4-VLA %s
 
 ; Tests for selecting interleave counts for loops with loads and stores.
 
@@ -225,6 +227,12 @@ define void @interleave_single_load_store(ptr %src, ptr %dst, i64 %N, i8 %a, i8
 ; INTERLEAVE-2:       exit:
 ; INTERLEAVE-2-NEXT:    ret void
 ;
+; INTERLEAVE-4-VLA-LABEL: @interleave_single_load_store(
+; INTERLEAVE-4-VLA:       call <vscale x 16 x i8> @llvm.smax.nxv16i8(
+; INTERLEAVE-4-VLA-NEXT:  call <vscale x 16 x i8> @llvm.smax.nxv16i8(
+; INTERLEAVE-4-VLA-NEXT:  call <vscale x 16 x i8> @llvm.smax.nxv16i8(
+; INTERLEAVE-4-VLA-NEXT:  call <vscale x 16 x i8> @llvm.smax.nxv16i8(
+;
 entry:
   br label %loop
 
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleaving-reduction.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleaving-reduction.ll
index 72d528d8748ba..9c207edf0f160 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/interleaving-reduction.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleaving-reduction.ll
@@ -5,6 +5,8 @@
 ; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-macos -mcpu=apple-a14 -S %s | FileCheck --check-prefix=INTERLEAVE-4 %s
 ; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-macos -mcpu=apple-a15 -S %s | FileCheck --check-prefix=INTERLEAVE-4 %s
 ; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-macos -mcpu=apple-a16 -S %s | FileCheck --check-prefix=INTERLEAVE-4 %s
+; RUN: opt -passes=loop-vectorize -mtriple=arm64 -mcpu=neoverse-v2 -S %s | FileCheck --check-prefix=INTERLEAVE-4 %s
+; RUN: opt -passes=loop-vectorize -mtriple=arm64 -mcpu=neoverse-v3 -S %s | FileCheck --check-prefix=INTERLEAVE-4-VLA %s
 
 ; Tests for selecting the interleave count for loops with reductions.
 
@@ -117,6 +119,12 @@ define i32 @interleave_integer_reduction(ptr %src, i64 %N) {
 ; INTERLEAVE-2-NEXT:    [[RED_NEXT_LCSSA:%.*]] = phi i32 [ [[RED_NEXT]], [[LOOP]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
 ; INTERLEAVE-2-NEXT:    ret i32 [[RED_NEXT_LCSSA]]
 ;
+; INTERLEAVE-4-VLA-LABEL: @interleave_integer_reduction(
+; INTERLEAVE-4-VLA:       add <vscale x 4 x i32>
+; INTERLEAVE-4-VLA-NEXT:  add <vscale x 4 x i32>
+; INTERLEAVE-4-VLA-NEXT:  add <vscale x 4 x i32>
+; INTERLEAVE-4-VLA-NEXT:  add <vscale x 4 x i32>
+;
 entry:
   br label %loop