| ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5 |
| |
| ; RUN: opt -mtriple=riscv64 -mattr=+m,+v -passes=slp-vectorizer -S -slp-enable-strided-stores < %s | FileCheck %s |
| |
| define void @constant_stride_2(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @constant_stride_2( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0:[0-9]+]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <8 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v8i8.p0.i64(<8 x i8> [[TMP1]], ptr align 1 [[GEP_S0]], i64 2, <8 x i1> splat (i1 true), i32 8) |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 2 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 4 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 6 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 8 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 10 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 12 |
| %gep_s7 = getelementptr i8, ptr %ps, i64 14 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| store i8 %load7, ptr %gep_s7 |
| |
| ret void |
| } |
| |
| define void @constant_stride_7(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @constant_stride_7( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <8 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v8i8.p0.i64(<8 x i8> [[TMP1]], ptr align 1 [[GEP_S0]], i64 7, <8 x i1> splat (i1 true), i32 8) |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 7 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 14 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 21 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 28 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 35 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 42 |
| %gep_s7 = getelementptr i8, ptr %ps, i64 49 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| store i8 %load7, ptr %gep_s7 |
| |
| ret void |
| } |
| |
| define void @constant_stride_0(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @constant_stride_0( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L1:%.*]] = getelementptr i8, ptr [[PL]], i64 1 |
| ; CHECK-NEXT: [[GEP_L2:%.*]] = getelementptr i8, ptr [[PL]], i64 2 |
| ; CHECK-NEXT: [[GEP_L3:%.*]] = getelementptr i8, ptr [[PL]], i64 3 |
| ; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr i8, ptr [[PL]], i64 4 |
| ; CHECK-NEXT: [[GEP_L5:%.*]] = getelementptr i8, ptr [[PL]], i64 5 |
| ; CHECK-NEXT: [[GEP_L6:%.*]] = getelementptr i8, ptr [[PL]], i64 6 |
| ; CHECK-NEXT: [[GEP_L7:%.*]] = getelementptr i8, ptr [[PL]], i64 7 |
| ; CHECK-NEXT: [[LOAD0:%.*]] = load i8, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: [[LOAD1:%.*]] = load i8, ptr [[GEP_L1]], align 1 |
| ; CHECK-NEXT: [[LOAD2:%.*]] = load i8, ptr [[GEP_L2]], align 1 |
| ; CHECK-NEXT: [[LOAD3:%.*]] = load i8, ptr [[GEP_L3]], align 1 |
| ; CHECK-NEXT: [[LOAD4:%.*]] = load i8, ptr [[GEP_L4]], align 1 |
| ; CHECK-NEXT: [[LOAD5:%.*]] = load i8, ptr [[GEP_L5]], align 1 |
| ; CHECK-NEXT: [[LOAD6:%.*]] = load i8, ptr [[GEP_L6]], align 1 |
| ; CHECK-NEXT: [[LOAD7:%.*]] = load i8, ptr [[GEP_L7]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD0]], ptr [[PS]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD1]], ptr [[PS]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD2]], ptr [[PS]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD3]], ptr [[PS]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD4]], ptr [[PS]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD5]], ptr [[PS]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD6]], ptr [[PS]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD7]], ptr [[PS]], align 1 |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| store i8 %load0, ptr %ps |
| store i8 %load1, ptr %ps |
| store i8 %load2, ptr %ps |
| store i8 %load3, ptr %ps |
| store i8 %load4, ptr %ps |
| store i8 %load5, ptr %ps |
| store i8 %load6, ptr %ps |
| store i8 %load7, ptr %ps |
| |
| ret void |
| } |
| |
| define void @two_constant_strides(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @two_constant_strides( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr i8, ptr [[PL]], i64 4 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[GEP_S4:%.*]] = getelementptr i8, ptr [[PS]], i64 9 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_L4]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP1]], ptr align 1 [[GEP_S0]], i64 2, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP2]], ptr align 1 [[GEP_S4]], i64 2, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 2 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 4 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 6 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 9 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 11 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 13 |
| %gep_s7 = getelementptr i8, ptr %ps, i64 15 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| store i8 %load7, ptr %gep_s7 |
| |
| ret void |
| } |
| |
| define void @constant_strides_variable_gap(ptr %pl, ptr %ps, i64 %gap) { |
| ; CHECK-LABEL: define void @constant_strides_variable_gap( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]], i64 [[GAP:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr i8, ptr [[PL]], i64 4 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[GAP_PS:%.*]] = getelementptr i8, ptr [[PS]], i64 [[GAP]] |
| ; CHECK-NEXT: [[GEP_S4:%.*]] = getelementptr i8, ptr [[GAP_PS]], i64 28 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L4]], align 1 |
| ; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP2]], ptr align 1 [[GEP_S0]], i64 7, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP1]], ptr align 1 [[GEP_S4]], i64 7, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 7 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 14 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 21 |
| |
| %gap_ps = getelementptr i8, ptr %ps, i64 %gap |
| %gep_s4 = getelementptr i8, ptr %gap_ps, i64 28 |
| %gep_s5 = getelementptr i8, ptr %gap_ps, i64 35 |
| %gep_s6 = getelementptr i8, ptr %gap_ps, i64 42 |
| %gep_s7 = getelementptr i8, ptr %gap_ps, i64 49 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| store i8 %load7, ptr %gep_s7 |
| |
| ret void |
| } |
| |
| define void @overlapping_strides(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @overlapping_strides( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr i8, ptr [[PL]], i64 4 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[GEP_S4:%.*]] = getelementptr i8, ptr [[PS]], i64 5 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_L4]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP1]], ptr align 1 [[GEP_S0]], i64 2, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP2]], ptr align 1 [[GEP_S4]], i64 3, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 2 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 4 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 6 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 5 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 8 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 11 |
| %gep_s7 = getelementptr i8, ptr %ps, i64 14 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| store i8 %load7, ptr %gep_s7 |
| |
| ret void |
| } |
| |
| define void @constant_stride_unit_stride(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @constant_stride_unit_stride( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr i8, ptr [[PL]], i64 4 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[GEP_S4:%.*]] = getelementptr i8, ptr [[PS]], i64 8 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L4]], align 1 |
| ; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP2]], ptr align 1 [[GEP_S0]], i64 2, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: store <4 x i8> [[TMP1]], ptr [[GEP_S4]], align 1 |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 2 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 4 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 6 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 8 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 9 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 10 |
| %gep_s7 = getelementptr i8, ptr %ps, i64 11 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| store i8 %load7, ptr %gep_s7 |
| |
| ret void |
| } |
| |
| define void @unit_stride_constant_stride(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @unit_stride_constant_stride( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr i8, ptr [[PL]], i64 4 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[GEP_S4:%.*]] = getelementptr i8, ptr [[PS]], i64 9 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_L4]], align 1 |
| ; CHECK-NEXT: store <4 x i8> [[TMP1]], ptr [[GEP_S0]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP2]], ptr align 1 [[GEP_S4]], i64 2, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 1 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 2 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 3 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 9 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 11 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 13 |
| %gep_s7 = getelementptr i8, ptr %ps, i64 15 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| store i8 %load7, ptr %gep_s7 |
| |
| ret void |
| } |
| |
| define void @overlap(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @overlap( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr i8, ptr [[PL]], i64 4 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[GEP_S4:%.*]] = getelementptr i8, ptr [[PS]], i64 5 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_L4]], align 1 |
| ; CHECK-NEXT: store <4 x i8> [[TMP1]], ptr [[GEP_S0]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v4i8.p0.i64(<4 x i8> [[TMP2]], ptr align 1 [[GEP_S4]], i64 2, <4 x i1> splat (i1 true), i32 4) |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 1 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 2 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 3 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 5 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 7 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 9 |
| %gep_s7 = getelementptr i8, ptr %ps, i64 11 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| store i8 %load7, ptr %gep_s7 |
| |
| ret void |
| } |
| |
| define void @overlap2(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @overlap2( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L2:%.*]] = getelementptr i8, ptr [[PL]], i64 2 |
| ; CHECK-NEXT: [[GEP_L3:%.*]] = getelementptr i8, ptr [[PL]], i64 3 |
| ; CHECK-NEXT: [[LOAD2:%.*]] = load i8, ptr [[GEP_L2]], align 1 |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[GEP_S2:%.*]] = getelementptr i8, ptr [[PS]], i64 4 |
| ; CHECK-NEXT: [[GEP_S3:%.*]] = getelementptr i8, ptr [[PS]], i64 6 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L3]], align 1 |
| ; CHECK-NEXT: [[TMP2:%.*]] = load <2 x i8>, ptr [[GEP_L0]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v2i8.p0.i64(<2 x i8> [[TMP2]], ptr align 1 [[GEP_S0]], i64 2, <2 x i1> splat (i1 true), i32 2) |
| ; CHECK-NEXT: store i8 [[LOAD2]], ptr [[GEP_S2]], align 1 |
| ; CHECK-NEXT: store <4 x i8> [[TMP1]], ptr [[GEP_S3]], align 1 |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 1 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 2 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 3 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 2 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 4 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 6 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 7 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 8 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 9 |
| |
| store i8 %load0, ptr %gep_s0 |
| store i8 %load1, ptr %gep_s1 |
| store i8 %load2, ptr %gep_s2 |
| store i8 %load3, ptr %gep_s3 |
| store i8 %load4, ptr %gep_s4 |
| store i8 %load5, ptr %gep_s5 |
| store i8 %load6, ptr %gep_s6 |
| |
| ret void |
| } |
| |
| define void @vf_ordering_issue(ptr %pl, ptr %ps) { |
| ; CHECK-LABEL: define void @vf_ordering_issue( |
| ; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] { |
| ; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr i8, ptr [[PL]], i64 0 |
| ; CHECK-NEXT: [[GEP_L1:%.*]] = getelementptr i8, ptr [[PL]], i64 38 |
| ; CHECK-NEXT: [[GEP_L14:%.*]] = getelementptr i8, ptr [[PL]], i64 92 |
| ; CHECK-NEXT: [[GEP_S11:%.*]] = getelementptr i8, ptr [[PL]], i64 33 |
| ; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr i8, ptr [[PL]], i64 4 |
| ; CHECK-NEXT: [[GEP_L20:%.*]] = getelementptr i8, ptr [[PL]], i64 13 |
| ; CHECK-NEXT: [[GEP_L21:%.*]] = getelementptr i8, ptr [[PL]], i64 83 |
| ; CHECK-NEXT: [[GEP_L22:%.*]] = getelementptr i8, ptr [[PL]], i64 32 |
| ; CHECK-NEXT: [[GEP_L3:%.*]] = getelementptr i8, ptr [[PL]], i64 15 |
| ; CHECK-NEXT: [[GEP_L23:%.*]] = getelementptr i8, ptr [[PL]], i64 23 |
| ; CHECK-NEXT: [[LOAD5:%.*]] = load i8, ptr [[GEP_S0]], align 1 |
| ; CHECK-NEXT: [[LOAD7:%.*]] = load i8, ptr [[GEP_L1]], align 1 |
| ; CHECK-NEXT: [[LOAD14:%.*]] = load i8, ptr [[GEP_L14]], align 1 |
| ; CHECK-NEXT: [[LOAD15:%.*]] = load i8, ptr [[GEP_S11]], align 1 |
| ; CHECK-NEXT: [[LOAD20:%.*]] = load i8, ptr [[GEP_L20]], align 1 |
| ; CHECK-NEXT: [[LOAD21:%.*]] = load i8, ptr [[GEP_L21]], align 1 |
| ; CHECK-NEXT: [[LOAD22:%.*]] = load i8, ptr [[GEP_L22]], align 1 |
| ; CHECK-NEXT: [[LOAD23:%.*]] = load i8, ptr [[GEP_L23]], align 1 |
| ; CHECK-NEXT: [[GEP_S5:%.*]] = getelementptr i8, ptr [[PS]], i64 0 |
| ; CHECK-NEXT: [[GEP_S7:%.*]] = getelementptr i8, ptr [[PS]], i64 2 |
| ; CHECK-NEXT: [[GEP_S9:%.*]] = getelementptr i8, ptr [[PS]], i64 4 |
| ; CHECK-NEXT: [[GEP_S3:%.*]] = getelementptr i8, ptr [[PS]], i64 6 |
| ; CHECK-NEXT: [[GEP_S4:%.*]] = getelementptr i8, ptr [[PS]], i64 8 |
| ; CHECK-NEXT: [[GEP_S12:%.*]] = getelementptr i8, ptr [[PS]], i64 16 |
| ; CHECK-NEXT: [[GEP_S13:%.*]] = getelementptr i8, ptr [[PS]], i64 18 |
| ; CHECK-NEXT: [[GEP_S14:%.*]] = getelementptr i8, ptr [[PS]], i64 20 |
| ; CHECK-NEXT: [[GEP_S15:%.*]] = getelementptr i8, ptr [[PS]], i64 22 |
| ; CHECK-NEXT: [[GEP_S23:%.*]] = getelementptr i8, ptr [[PS]], i64 38 |
| ; CHECK-NEXT: [[TMP1:%.*]] = load <8 x i8>, ptr [[GEP_L4]], align 1 |
| ; CHECK-NEXT: [[TMP2:%.*]] = add <8 x i8> [[TMP1]], splat (i8 1) |
| ; CHECK-NEXT: [[TMP3:%.*]] = load <8 x i8>, ptr [[GEP_L3]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD5]], ptr [[GEP_S5]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD7]], ptr [[GEP_S7]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD14]], ptr [[GEP_S9]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD15]], ptr [[GEP_S3]], align 1 |
| ; CHECK-NEXT: store <8 x i8> [[TMP2]], ptr [[GEP_S4]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD20]], ptr [[GEP_S12]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD21]], ptr [[GEP_S13]], align 1 |
| ; CHECK-NEXT: store i8 [[LOAD22]], ptr [[GEP_S14]], align 1 |
| ; CHECK-NEXT: call void @llvm.experimental.vp.strided.store.v8i8.p0.i64(<8 x i8> [[TMP3]], ptr align 1 [[GEP_S15]], i64 2, <8 x i1> splat (i1 true), i32 8) |
| ; CHECK-NEXT: store i8 [[LOAD23]], ptr [[GEP_S23]], align 1 |
| ; CHECK-NEXT: ret void |
| ; |
| %gep_l0 = getelementptr i8, ptr %pl, i64 0 |
| %gep_l1 = getelementptr i8, ptr %pl, i64 38 |
| %gep_l2 = getelementptr i8, ptr %pl, i64 92 |
| %gep_l3 = getelementptr i8, ptr %pl, i64 33 |
| %gep_l4 = getelementptr i8, ptr %pl, i64 4 |
| %gep_l5 = getelementptr i8, ptr %pl, i64 5 |
| %gep_l6 = getelementptr i8, ptr %pl, i64 6 |
| %gep_l7 = getelementptr i8, ptr %pl, i64 7 |
| %gep_l8 = getelementptr i8, ptr %pl, i64 8 |
| %gep_l9 = getelementptr i8, ptr %pl, i64 9 |
| %gep_l10 = getelementptr i8, ptr %pl, i64 10 |
| %gep_l11 = getelementptr i8, ptr %pl, i64 11 |
| %gep_l12 = getelementptr i8, ptr %pl, i64 13 |
| %gep_l13 = getelementptr i8, ptr %pl, i64 83 |
| %gep_l14 = getelementptr i8, ptr %pl, i64 32 |
| %gep_l15 = getelementptr i8, ptr %pl, i64 15 |
| %gep_l16 = getelementptr i8, ptr %pl, i64 16 |
| %gep_l17 = getelementptr i8, ptr %pl, i64 17 |
| %gep_l18 = getelementptr i8, ptr %pl, i64 18 |
| %gep_l19 = getelementptr i8, ptr %pl, i64 19 |
| %gep_l20 = getelementptr i8, ptr %pl, i64 20 |
| %gep_l21 = getelementptr i8, ptr %pl, i64 21 |
| %gep_l22 = getelementptr i8, ptr %pl, i64 22 |
| %gep_l23 = getelementptr i8, ptr %pl, i64 23 |
| |
| %load0 = load i8, ptr %gep_l0 |
| %load1 = load i8, ptr %gep_l1 |
| %load2 = load i8, ptr %gep_l2 |
| %load3 = load i8, ptr %gep_l3 |
| %load4 = load i8, ptr %gep_l4 |
| %load5 = load i8, ptr %gep_l5 |
| %load6 = load i8, ptr %gep_l6 |
| %load7 = load i8, ptr %gep_l7 |
| %load8 = load i8, ptr %gep_l8 |
| %load9 = load i8, ptr %gep_l9 |
| %load10 = load i8, ptr %gep_l10 |
| %load11 = load i8, ptr %gep_l11 |
| %load12 = load i8, ptr %gep_l12 |
| %load13 = load i8, ptr %gep_l13 |
| %load14 = load i8, ptr %gep_l14 |
| %load15 = load i8, ptr %gep_l15 |
| %load16 = load i8, ptr %gep_l16 |
| %load17 = load i8, ptr %gep_l17 |
| %load18 = load i8, ptr %gep_l18 |
| %load19 = load i8, ptr %gep_l19 |
| %load20 = load i8, ptr %gep_l20 |
| %load21 = load i8, ptr %gep_l21 |
| %load22 = load i8, ptr %gep_l22 |
| %load23 = load i8, ptr %gep_l23 |
| |
| %add4 = add i8 %load4, 1 |
| %add5 = add i8 %load5, 1 |
| %add6 = add i8 %load6, 1 |
| %add7 = add i8 %load7, 1 |
| %add8 = add i8 %load8, 1 |
| %add9 = add i8 %load9, 1 |
| %add10 = add i8 %load10, 1 |
| %add11 = add i8 %load11, 1 |
| |
| %gep_s0 = getelementptr i8, ptr %ps, i64 0 |
| %gep_s1 = getelementptr i8, ptr %ps, i64 2 |
| %gep_s2 = getelementptr i8, ptr %ps, i64 4 |
| %gep_s3 = getelementptr i8, ptr %ps, i64 6 |
| %gep_s4 = getelementptr i8, ptr %ps, i64 8 |
| %gep_s5 = getelementptr i8, ptr %ps, i64 9 |
| %gep_s6 = getelementptr i8, ptr %ps, i64 10 |
| %gep_s7 = getelementptr i8, ptr %ps, i64 11 |
| %gep_s8 = getelementptr i8, ptr %ps, i64 12 |
| %gep_s9 = getelementptr i8, ptr %ps, i64 13 |
| %gep_s10 = getelementptr i8, ptr %ps, i64 14 |
| %gep_s11 = getelementptr i8, ptr %ps, i64 15 |
| %gep_s12 = getelementptr i8, ptr %ps, i64 16 |
| %gep_s13 = getelementptr i8, ptr %ps, i64 18 |
| %gep_s14 = getelementptr i8, ptr %ps, i64 20 |
| %gep_s15 = getelementptr i8, ptr %ps, i64 22 |
| %gep_s16 = getelementptr i8, ptr %ps, i64 24 |
| %gep_s17 = getelementptr i8, ptr %ps, i64 26 |
| %gep_s18 = getelementptr i8, ptr %ps, i64 28 |
| %gep_s19 = getelementptr i8, ptr %ps, i64 30 |
| %gep_s20 = getelementptr i8, ptr %ps, i64 32 |
| %gep_s21 = getelementptr i8, ptr %ps, i64 34 |
| %gep_s22 = getelementptr i8, ptr %ps, i64 36 |
| %gep_s23 = getelementptr i8, ptr %ps, i64 38 |
| |
| store i8 %load0 , ptr %gep_s0 |
| store i8 %load1 , ptr %gep_s1 |
| store i8 %load2 , ptr %gep_s2 |
| store i8 %load3 , ptr %gep_s3 |
| store i8 %add4 , ptr %gep_s4 |
| store i8 %add5 , ptr %gep_s5 |
| store i8 %add6 , ptr %gep_s6 |
| store i8 %add7 , ptr %gep_s7 |
| store i8 %add8 , ptr %gep_s8 |
| store i8 %add9 , ptr %gep_s9 |
| store i8 %add10 , ptr %gep_s10 |
| store i8 %add11 , ptr %gep_s11 |
| store i8 %load12 , ptr %gep_s12 |
| store i8 %load13 , ptr %gep_s13 |
| store i8 %load14 , ptr %gep_s14 |
| store i8 %load15 , ptr %gep_s15 |
| store i8 %load16 , ptr %gep_s16 |
| store i8 %load17 , ptr %gep_s17 |
| store i8 %load18 , ptr %gep_s18 |
| store i8 %load19 , ptr %gep_s19 |
| store i8 %load20 , ptr %gep_s20 |
| store i8 %load21 , ptr %gep_s21 |
| store i8 %load22 , ptr %gep_s22 |
| store i8 %load23 , ptr %gep_s23 |
| |
| ret void |
| } |
