1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -passes=slp-vectorizer -mtriple=riscv64 -mattr=+v \ 3; RUN: -riscv-v-slp-max-vf=0 -S | FileCheck %s 4 5; This should not be vectorized, as the cost of computing the offsets nullifies 6; the benefits of vectorizing: 7; 8; copy_with_offset_v2i8: 9; addi a0, a0, 8 10; vsetivli zero, 2, e8, mf8, ta, ma 11; vle8.v v8, (a0) 12; addi a1, a1, 16 13; vse8.v v8, (a1) 14; ret 15; 16; Compared to the scalar version where the offsets can be folded into the 17; addressing mode: 18; 19; copy_with_offset_v2i8: 20; lbu a2, 8(a0) 21; lbu a0, 9(a0) 22; sb a2, 16(a1) 23; sb a0, 17(a1) 24; ret 25 26define void @copy_with_offset_v2i8(ptr noalias %p, ptr noalias %q) { 27; CHECK-LABEL: @copy_with_offset_v2i8( 28; CHECK-NEXT: entry: 29; CHECK-NEXT: [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 8 30; CHECK-NEXT: [[X1:%.*]] = load i8, ptr [[P1]], align 1 31; CHECK-NEXT: [[Q1:%.*]] = getelementptr i8, ptr [[Q:%.*]], i32 16 32; CHECK-NEXT: store i8 [[X1]], ptr [[Q1]], align 1 33; CHECK-NEXT: [[P2:%.*]] = getelementptr i8, ptr [[P]], i32 9 34; CHECK-NEXT: [[X2:%.*]] = load i8, ptr [[P2]], align 1 35; CHECK-NEXT: [[Q2:%.*]] = getelementptr i8, ptr [[Q]], i32 17 36; CHECK-NEXT: store i8 [[X2]], ptr [[Q2]], align 1 37; CHECK-NEXT: ret void 38; 39entry: 40 %p1 = getelementptr i8, ptr %p, i32 8 41 %x1 = load i8, ptr %p1 42 %q1 = getelementptr i8, ptr %q, i32 16 43 store i8 %x1, ptr %q1 44 45 %p2 = getelementptr i8, ptr %p, i32 9 46 %x2 = load i8, ptr %p2 47 %q2 = getelementptr i8, ptr %q, i32 17 48 store i8 %x2, ptr %q2 49 50 ret void 51} 52 53; This on the other hand, should be vectorized as the vector savings outweigh 54; the GEP costs. 55define void @copy_with_offset_v4i8(ptr noalias %p, ptr noalias %q) { 56; CHECK-LABEL: @copy_with_offset_v4i8( 57; CHECK-NEXT: entry: 58; CHECK-NEXT: [[P1:%.*]] = getelementptr i8, ptr [[P:%.*]], i32 8 59; CHECK-NEXT: [[Q1:%.*]] = getelementptr i8, ptr [[Q:%.*]], i32 16 60; CHECK-NEXT: [[TMP0:%.*]] = load <4 x i8>, ptr [[P1]], align 1 61; CHECK-NEXT: store <4 x i8> [[TMP0]], ptr [[Q1]], align 1 62; CHECK-NEXT: ret void 63; 64entry: 65 %p1 = getelementptr i8, ptr %p, i32 8 66 %x1 = load i8, ptr %p1 67 %q1 = getelementptr i8, ptr %q, i32 16 68 store i8 %x1, ptr %q1 69 70 %p2 = getelementptr i8, ptr %p, i32 9 71 %x2 = load i8, ptr %p2 72 %q2 = getelementptr i8, ptr %q, i32 17 73 store i8 %x2, ptr %q2 74 75 %p3 = getelementptr i8, ptr %p, i32 10 76 %x3 = load i8, ptr %p3 77 %q3 = getelementptr i8, ptr %q, i32 18 78 store i8 %x3, ptr %q3 79 80 %p4 = getelementptr i8, ptr %p, i32 11 81 %x4 = load i8, ptr %p4 82 %q4 = getelementptr i8, ptr %q, i32 19 83 store i8 %x4, ptr %q4 84 85 ret void 86} 87