1; RUN: opt < %s -cost-model -analyze -mtriple=thumbv7-apple-ios6.0.0 -march=arm -mcpu=cortex-a8 | FileCheck %s --check-prefix=COST 2; RUN: llc -mtriple=arm-eabi -mattr=+neon %s -o - | FileCheck %s 3; Make sure that ARM backend with NEON handles vselect. 4 5define void @vmax_v4i32(<4 x i32>* %m, <4 x i32> %a, <4 x i32> %b) { 6; CHECK: vmax.s32 {{q[0-9]+}}, {{q[0-9]+}}, {{q[0-9]+}} 7 %cmpres = icmp sgt <4 x i32> %a, %b 8 %maxres = select <4 x i1> %cmpres, <4 x i32> %a, <4 x i32> %b 9 store <4 x i32> %maxres, <4 x i32>* %m 10 ret void 11} 12 13%T0_10 = type <16 x i16> 14%T1_10 = type <16 x i1> 15; CHECK-LABEL: func_blend10: 16define void @func_blend10(%T0_10* %loadaddr, %T0_10* %loadaddr2, 17 %T1_10* %blend, %T0_10* %storeaddr) { 18 %v0 = load %T0_10, %T0_10* %loadaddr 19 %v1 = load %T0_10, %T0_10* %loadaddr2 20 %c = icmp slt %T0_10 %v0, %v1 21; CHECK: vmin.s16 22; CHECK: vmin.s16 23; COST: func_blend10 24; COST: cost of 2 {{.*}} select 25 %r = select %T1_10 %c, %T0_10 %v0, %T0_10 %v1 26 store %T0_10 %r, %T0_10* %storeaddr 27 ret void 28} 29%T0_14 = type <8 x i32> 30%T1_14 = type <8 x i1> 31; CHECK-LABEL: func_blend14: 32define void @func_blend14(%T0_14* %loadaddr, %T0_14* %loadaddr2, 33 %T1_14* %blend, %T0_14* %storeaddr) { 34 %v0 = load %T0_14, %T0_14* %loadaddr 35 %v1 = load %T0_14, %T0_14* %loadaddr2 36 %c = icmp slt %T0_14 %v0, %v1 37; CHECK: vmin.s32 38; CHECK: vmin.s32 39; COST: func_blend14 40; COST: cost of 2 {{.*}} select 41 %r = select %T1_14 %c, %T0_14 %v0, %T0_14 %v1 42 store %T0_14 %r, %T0_14* %storeaddr 43 ret void 44} 45%T0_15 = type <16 x i32> 46%T1_15 = type <16 x i1> 47; CHECK-LABEL: func_blend15: 48define void @func_blend15(%T0_15* %loadaddr, %T0_15* %loadaddr2, 49 %T1_15* %blend, %T0_15* %storeaddr) { 50; CHECK: vmin.s32 51; CHECK: vmin.s32 52 %v0 = load %T0_15, %T0_15* %loadaddr 53 %v1 = load %T0_15, %T0_15* %loadaddr2 54 %c = icmp slt %T0_15 %v0, %v1 55; COST: func_blend15 56; COST: cost of 4 {{.*}} select 57 %r = select %T1_15 %c, %T0_15 %v0, %T0_15 %v1 58 store %T0_15 %r, %T0_15* %storeaddr 59 ret void 60} 61 62; We adjusted the cost model of the following selects. When we improve code 63; lowering we also need to adjust the cost. 64%T0_18 = type <4 x i64> 65%T1_18 = type <4 x i1> 66; CHECK-LABEL: func_blend18: 67define void @func_blend18(%T0_18* %loadaddr, %T0_18* %loadaddr2, 68 %T1_18* %blend, %T0_18* %storeaddr) { 69; CHECK: vbsl 70; CHECK: vbsl 71 %v0 = load %T0_18, %T0_18* %loadaddr 72 %v1 = load %T0_18, %T0_18* %loadaddr2 73 %c = icmp slt %T0_18 %v0, %v1 74; COST: func_blend18 75; COST: cost of 19 {{.*}} select 76 %r = select %T1_18 %c, %T0_18 %v0, %T0_18 %v1 77 store %T0_18 %r, %T0_18* %storeaddr 78 ret void 79} 80%T0_19 = type <8 x i64> 81%T1_19 = type <8 x i1> 82; CHECK-LABEL: func_blend19: 83define void @func_blend19(%T0_19* %loadaddr, %T0_19* %loadaddr2, 84 %T1_19* %blend, %T0_19* %storeaddr) { 85; CHECK: vbsl 86; CHECK: vbsl 87; CHECK: vbsl 88; CHECK: vbsl 89 %v0 = load %T0_19, %T0_19* %loadaddr 90 %v1 = load %T0_19, %T0_19* %loadaddr2 91 %c = icmp slt %T0_19 %v0, %v1 92; COST: func_blend19 93; COST: cost of 50 {{.*}} select 94 %r = select %T1_19 %c, %T0_19 %v0, %T0_19 %v1 95 store %T0_19 %r, %T0_19* %storeaddr 96 ret void 97} 98%T0_20 = type <16 x i64> 99%T1_20 = type <16 x i1> 100; CHECK-LABEL: func_blend20: 101define void @func_blend20(%T0_20* %loadaddr, %T0_20* %loadaddr2, 102 %T1_20* %blend, %T0_20* %storeaddr) { 103; CHECK: vbsl 104; CHECK: vbsl 105; CHECK: vbsl 106; CHECK: vbsl 107; CHECK: vbsl 108; CHECK: vbsl 109; CHECK: vbsl 110; CHECK: vbsl 111 %v0 = load %T0_20, %T0_20* %loadaddr 112 %v1 = load %T0_20, %T0_20* %loadaddr2 113 %c = icmp slt %T0_20 %v0, %v1 114; COST: func_blend20 115; COST: cost of 100 {{.*}} select 116 %r = select %T1_20 %c, %T0_20 %v0, %T0_20 %v1 117 store %T0_20 %r, %T0_20* %storeaddr 118 ret void 119} 120