[llvm] [LV] Support interleaving in order to reduce stalls (PR #197959)

Florian Hahn via llvm-commits llvm-commits at lists.llvm.org
Tue Jul 21 07:53:17 PDT 2026


================
@@ -0,0 +1,1198 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
+; RUN: opt -passes=loop-vectorize -enable-interleave-to-reduce-stalls -force-target-max-vector-interleave=4 -force-target-load-latency=1 -S %s | FileCheck %s --check-prefixes=CHECK-LATENCY1
+; RUN: opt -passes=loop-vectorize -enable-interleave-to-reduce-stalls -force-target-max-vector-interleave=4 -force-target-load-latency=2 -S %s | FileCheck %s --check-prefixes=CHECK-LATENCY2
+; RUN: opt -passes=loop-vectorize -enable-interleave-to-reduce-stalls -force-target-max-vector-interleave=4 -force-target-load-latency=8 -S %s | FileCheck %s --check-prefixes=CHECK-LATENCY8
+; RUN: opt -passes=loop-vectorize -enable-interleave-to-reduce-stalls -force-target-max-vector-interleave=4 -mcpu=cortex-a510 -S %s | FileCheck %s --check-prefixes=CHECK-A510
+; RUN: opt -passes=loop-vectorize -enable-interleave-to-reduce-stalls -force-target-max-vector-interleave=4 -mcpu=cortex-a320 -S %s | FileCheck %s --check-prefixes=CHECK-A320
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"
+target triple = "aarch64-unknown-none-elf"
+
+; In these tests we expect that the loop is interleaved enough that we have as
+; many loads as the load latency. The exception is for latency 8, due to the
+; max interleave factor being 4.
+
+define void @i16_add1(ptr noalias readonly %src, ptr noalias writeonly %dst, i16 %argval) {
+; CHECK-LATENCY1-LABEL: define void @i16_add1(
+; CHECK-LATENCY1-SAME: ptr noalias readonly [[SRC:%.*]], ptr noalias writeonly [[DST:%.*]], i16 [[ARGVAL:%.*]]) {
+; CHECK-LATENCY1-NEXT:  [[ENTRY:.*:]]
+; CHECK-LATENCY1-NEXT:    br label %[[VECTOR_PH:.*]]
+; CHECK-LATENCY1:       [[VECTOR_PH]]:
+; CHECK-LATENCY1-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i16> poison, i16 [[ARGVAL]], i64 0
+; CHECK-LATENCY1-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i16> [[BROADCAST_SPLATINSERT]], <8 x i16> poison, <8 x i32> zeroinitializer
+; CHECK-LATENCY1-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK-LATENCY1:       [[VECTOR_BODY]]:
+; CHECK-LATENCY1-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-LATENCY1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds nuw i16, ptr [[SRC]], i64 [[INDEX]]
+; CHECK-LATENCY1-NEXT:    [[WIDE_LOAD:%.*]] = load <8 x i16>, ptr [[TMP0]], align 2
+; CHECK-LATENCY1-NEXT:    [[TMP1:%.*]] = add <8 x i16> [[WIDE_LOAD]], [[BROADCAST_SPLAT]]
+; CHECK-LATENCY1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw i16, ptr [[DST]], i64 [[INDEX]]
+; CHECK-LATENCY1-NEXT:    store <8 x i16> [[TMP1]], ptr [[TMP2]], align 2
+; CHECK-LATENCY1-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
+; CHECK-LATENCY1-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
+; CHECK-LATENCY1-NEXT:    br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK-LATENCY1:       [[MIDDLE_BLOCK]]:
+; CHECK-LATENCY1-NEXT:    br label %[[EXIT:.*]]
+; CHECK-LATENCY1:       [[EXIT]]:
+; CHECK-LATENCY1-NEXT:    ret void
+;
+; CHECK-LATENCY2-LABEL: define void @i16_add1(
+; CHECK-LATENCY2-SAME: ptr noalias readonly [[SRC:%.*]], ptr noalias writeonly [[DST:%.*]], i16 [[ARGVAL:%.*]]) {
+; CHECK-LATENCY2-NEXT:  [[ENTRY:.*:]]
+; CHECK-LATENCY2-NEXT:    br label %[[VECTOR_PH:.*]]
+; CHECK-LATENCY2:       [[VECTOR_PH]]:
+; CHECK-LATENCY2-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i16> poison, i16 [[ARGVAL]], i64 0
+; CHECK-LATENCY2-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i16> [[BROADCAST_SPLATINSERT]], <8 x i16> poison, <8 x i32> zeroinitializer
+; CHECK-LATENCY2-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK-LATENCY2:       [[VECTOR_BODY]]:
+; CHECK-LATENCY2-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-LATENCY2-NEXT:    [[TMP0:%.*]] = getelementptr inbounds nuw i16, ptr [[SRC]], i64 [[INDEX]]
----------------
fhahn wrote:

Can we share check lines where the same interleave factor is chosen? Here it looks like those are the same as ` CHECK-A510-NEXT: ` at a quick glance.

https://github.com/llvm/llvm-project/pull/197959


More information about the llvm-commits mailing list