[llvm] [SLPVectorizer] Widen strided loads. (PR #153074)

[Alexey Bataev via llvm-commits]

================
@@ -0,0 +1,483 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -mtriple=riscv64 -mattr=+m,+v,+unaligned-vector-mem \
+; RUN: -riscv-v-slp-prefer-alt-opc-vectorization=true \
+; RUN: -passes=slp-vectorizer -S < %s | FileCheck %s
+; Function Attrs: nounwind uwtable vscale_range(8,1024)
+define i32 @x264_pixel_satd_8x4(ptr %pix1, i32  %i_pix1, ptr  %pix2, i32  %i_pix2) {
+; CHECK-LABEL: define i32 @x264_pixel_satd_8x4(
+; CHECK-SAME: ptr [[PIX1:%.*]], i32 [[I_PIX1:%.*]], ptr [[PIX2:%.*]], i32 [[I_PIX2:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[IDX_EXT:%.*]] = sext i32 [[I_PIX1]] to i64
+; CHECK-NEXT:    [[IDX_EXT63:%.*]] = sext i32 [[I_PIX2]] to i64
+; CHECK-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds nuw i8, ptr [[PIX1]], i64 4
+; CHECK-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds nuw i8, ptr [[PIX2]], i64 4
+; CHECK-NEXT:    [[TMP0:%.*]] = mul i64 [[IDX_EXT]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = call <4 x i32> @llvm.experimental.vp.strided.load.v4i32.p0.i64(ptr align 1 [[PIX1]], i64 [[TMP0]], <4 x i1> splat (i1 true), i32 4)
+; CHECK-NEXT:    [[TMP19:%.*]] = bitcast <4 x i32> [[TMP1]] to <16 x i8>
+; CHECK-NEXT:    [[TMP20:%.*]] = zext <16 x i8> [[TMP19]] to <16 x i32>
+; CHECK-NEXT:    [[TMP4:%.*]] = mul i64 [[IDX_EXT63]], 1
+; CHECK-NEXT:    [[TMP5:%.*]] = call <4 x i32> @llvm.experimental.vp.strided.load.v4i32.p0.i64(ptr align 1 [[PIX2]], i64 [[TMP4]], <4 x i1> splat (i1 true), i32 4)
+; CHECK-NEXT:    [[TMP28:%.*]] = bitcast <4 x i32> [[TMP5]] to <16 x i8>
+; CHECK-NEXT:    [[TMP29:%.*]] = zext <16 x i8> [[TMP28]] to <16 x i32>
+; CHECK-NEXT:    [[TMP30:%.*]] = sub nsw <16 x i32> [[TMP20]], [[TMP29]]
+; CHECK-NEXT:    [[TMP9:%.*]] = mul i64 [[IDX_EXT]], 1
+; CHECK-NEXT:    [[TMP10:%.*]] = call <4 x i32> @llvm.experimental.vp.strided.load.v4i32.p0.i64(ptr align 1 [[ARRAYIDX3]], i64 [[TMP9]], <4 x i1> splat (i1 true), i32 4)
+; CHECK-NEXT:    [[TMP38:%.*]] = bitcast <4 x i32> [[TMP10]] to <16 x i8>
+; CHECK-NEXT:    [[TMP39:%.*]] = zext <16 x i8> [[TMP38]] to <16 x i32>
+; CHECK-NEXT:    [[TMP13:%.*]] = mul i64 [[IDX_EXT63]], 1
+; CHECK-NEXT:    [[TMP14:%.*]] = call <4 x i32> @llvm.experimental.vp.strided.load.v4i32.p0.i64(ptr align 1 [[ARRAYIDX5]], i64 [[TMP13]], <4 x i1> splat (i1 true), i32 4)
+; CHECK-NEXT:    [[TMP47:%.*]] = bitcast <4 x i32> [[TMP14]] to <16 x i8>
+; CHECK-NEXT:    [[TMP48:%.*]] = zext <16 x i8> [[TMP47]] to <16 x i32>
+; CHECK-NEXT:    [[TMP49:%.*]] = sub nsw <16 x i32> [[TMP39]], [[TMP48]]
+; CHECK-NEXT:    [[TMP50:%.*]] = shl nsw <16 x i32> [[TMP49]], splat (i32 16)
+; CHECK-NEXT:    [[TMP51:%.*]] = add nsw <16 x i32> [[TMP50]], [[TMP30]]
+; CHECK-NEXT:    [[TMP52:%.*]] = shufflevector <16 x i32> [[TMP51]], <16 x i32> poison, <16 x i32> <i32 1, i32 0, i32 3, i32 2, i32 5, i32 4, i32 7, i32 6, i32 9, i32 8, i32 11, i32 10, i32 13, i32 12, i32 15, i32 14>
+; CHECK-NEXT:    [[TMP53:%.*]] = add nsw <16 x i32> [[TMP52]], [[TMP51]]
+; CHECK-NEXT:    [[TMP54:%.*]] = sub nsw <16 x i32> [[TMP52]], [[TMP51]]
+; CHECK-NEXT:    [[TMP55:%.*]] = shufflevector <16 x i32> [[TMP53]], <16 x i32> [[TMP54]], <16 x i32> <i32 0, i32 17, i32 2, i32 19, i32 4, i32 21, i32 6, i32 23, i32 8, i32 25, i32 10, i32 27, i32 12, i32 29, i32 14, i32 31>
----------------
alexey-bataev wrote:

I would like to avoid potential regressions, so would be good to provide perf numbers at least

https://github.com/llvm/llvm-project/pull/153074