| 1 |
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6 |
| 2 |
; RUN: opt -passes=vector-combine -mtriple=arm64-apple-darwinos -S %s | FileCheck %s |
| 3 |
|
| 4 |
define i32 @load_v4i8_bitcast_to_i32(ptr %x) { |
| 5 |
; CHECK-LABEL: define i32 @load_v4i8_bitcast_to_i32( |
| 6 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 7 |
; CHECK-NEXT: [[R_SCALAR:%.*]] = load i32, ptr [[X]], align 4 |
| 8 |
; CHECK-NEXT: ret i32 [[R_SCALAR]] |
| 9 |
; |
| 10 |
%lv = load <4 x i8>, ptr %x |
| 11 |
%r = bitcast <4 x i8> %lv to i32 |
| 12 |
ret i32 %r |
| 13 |
} |
| 14 |
|
| 15 |
define i64 @load_v2i32_bitcast_to_i64(ptr %x) { |
| 16 |
; CHECK-LABEL: define i64 @load_v2i32_bitcast_to_i64( |
| 17 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 18 |
; CHECK-NEXT: [[R_SCALAR:%.*]] = load i64, ptr [[X]], align 8 |
| 19 |
; CHECK-NEXT: ret i64 [[R_SCALAR]] |
| 20 |
; |
| 21 |
%lv = load <2 x i32>, ptr %x |
| 22 |
%r = bitcast <2 x i32> %lv to i64 |
| 23 |
ret i64 %r |
| 24 |
} |
| 25 |
|
| 26 |
define float @load_v4i8_bitcast_to_float(ptr %x) { |
| 27 |
; CHECK-LABEL: define float @load_v4i8_bitcast_to_float( |
| 28 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 29 |
; CHECK-NEXT: [[R_SCALAR:%.*]] = load float, ptr [[X]], align 4 |
| 30 |
; CHECK-NEXT: ret float [[R_SCALAR]] |
| 31 |
; |
| 32 |
%lv = load <4 x i8>, ptr %x |
| 33 |
%r = bitcast <4 x i8> %lv to float |
| 34 |
ret float %r |
| 35 |
} |
| 36 |
|
| 37 |
define float @load_v2i16_bitcast_to_float(ptr %x) { |
| 38 |
; CHECK-LABEL: define float @load_v2i16_bitcast_to_float( |
| 39 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 40 |
; CHECK-NEXT: [[R_SCALAR:%.*]] = load float, ptr [[X]], align 4 |
| 41 |
; CHECK-NEXT: ret float [[R_SCALAR]] |
| 42 |
; |
| 43 |
%lv = load <2 x i16>, ptr %x |
| 44 |
%r = bitcast <2 x i16> %lv to float |
| 45 |
ret float %r |
| 46 |
} |
| 47 |
|
| 48 |
define double @load_v4i16_bitcast_to_double(ptr %x) { |
| 49 |
; CHECK-LABEL: define double @load_v4i16_bitcast_to_double( |
| 50 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 51 |
; CHECK-NEXT: [[LV:%.*]] = load <4 x i16>, ptr [[X]], align 8 |
| 52 |
; CHECK-NEXT: [[R_SCALAR:%.*]] = bitcast <4 x i16> [[LV]] to double |
| 53 |
; CHECK-NEXT: ret double [[R_SCALAR]] |
| 54 |
; |
| 55 |
%lv = load <4 x i16>, ptr %x |
| 56 |
%r = bitcast <4 x i16> %lv to double |
| 57 |
ret double %r |
| 58 |
} |
| 59 |
|
| 60 |
define double @load_v2i32_bitcast_to_double(ptr %x) { |
| 61 |
; CHECK-LABEL: define double @load_v2i32_bitcast_to_double( |
| 62 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 63 |
; CHECK-NEXT: [[LV:%.*]] = load <2 x i32>, ptr [[X]], align 8 |
| 64 |
; CHECK-NEXT: [[R_SCALAR:%.*]] = bitcast <2 x i32> [[LV]] to double |
| 65 |
; CHECK-NEXT: ret double [[R_SCALAR]] |
| 66 |
; |
| 67 |
%lv = load <2 x i32>, ptr %x |
| 68 |
%r = bitcast <2 x i32> %lv to double |
| 69 |
ret double %r |
| 70 |
} |
| 71 |
|
| 72 |
; Multiple users with the same bitcast type should be scalarized. |
| 73 |
define i32 @load_v4i8_bitcast_multiple_users_same_type(ptr %x) { |
| 74 |
; CHECK-LABEL: define i32 @load_v4i8_bitcast_multiple_users_same_type( |
| 75 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 76 |
; CHECK-NEXT: [[LV_SCALAR:%.*]] = load i32, ptr [[X]], align 4 |
| 77 |
; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV_SCALAR]], [[LV_SCALAR]] |
| 78 |
; CHECK-NEXT: ret i32 [[ADD]] |
| 79 |
; |
| 80 |
%lv = load <4 x i8>, ptr %x |
| 81 |
%r1 = bitcast <4 x i8> %lv to i32 |
| 82 |
%r2 = bitcast <4 x i8> %lv to i32 |
| 83 |
%add = add i32 %r1, %r2 |
| 84 |
ret i32 %add |
| 85 |
} |
| 86 |
|
| 87 |
; Different bitcast types should not be scalarized. |
| 88 |
define i32 @load_v4i8_bitcast_multiple_users_different_types(ptr %x) { |
| 89 |
; CHECK-LABEL: define i32 @load_v4i8_bitcast_multiple_users_different_types( |
| 90 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 91 |
; CHECK-NEXT: [[LV:%.*]] = load <4 x i8>, ptr [[X]], align 4 |
| 92 |
; CHECK-NEXT: [[R1:%.*]] = bitcast <4 x i8> [[LV]] to i32 |
| 93 |
; CHECK-NEXT: [[R2:%.*]] = bitcast <4 x i8> [[LV]] to float |
| 94 |
; CHECK-NEXT: [[R2_INT:%.*]] = bitcast float [[R2]] to i32 |
| 95 |
; CHECK-NEXT: [[ADD:%.*]] = add i32 [[R1]], [[R2_INT]] |
| 96 |
; CHECK-NEXT: ret i32 [[ADD]] |
| 97 |
; |
| 98 |
%lv = load <4 x i8>, ptr %x |
| 99 |
%r1 = bitcast <4 x i8> %lv to i32 |
| 100 |
%r2 = bitcast <4 x i8> %lv to float |
| 101 |
%r2.int = bitcast float %r2 to i32 |
| 102 |
%add = add i32 %r1, %r2.int |
| 103 |
ret i32 %add |
| 104 |
} |
| 105 |
|
| 106 |
; Bitcast to vector should not be scalarized. |
| 107 |
define <2 x i16> @load_v4i8_bitcast_to_vector(ptr %x) { |
| 108 |
; CHECK-LABEL: define <2 x i16> @load_v4i8_bitcast_to_vector( |
| 109 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 110 |
; CHECK-NEXT: [[LV:%.*]] = load <4 x i8>, ptr [[X]], align 4 |
| 111 |
; CHECK-NEXT: [[R:%.*]] = bitcast <4 x i8> [[LV]] to <2 x i16> |
| 112 |
; CHECK-NEXT: ret <2 x i16> [[R]] |
| 113 |
; |
| 114 |
%lv = load <4 x i8>, ptr %x |
| 115 |
%r = bitcast <4 x i8> %lv to <2 x i16> |
| 116 |
ret <2 x i16> %r |
| 117 |
} |
| 118 |
|
| 119 |
; Load with both bitcast users and other users should not be scalarized. |
| 120 |
define i32 @load_v4i8_mixed_users(ptr %x) { |
| 121 |
; CHECK-LABEL: define i32 @load_v4i8_mixed_users( |
| 122 |
; CHECK-SAME: ptr [[X:%.*]]) { |
| 123 |
; CHECK-NEXT: [[LV:%.*]] = load <4 x i8>, ptr [[X]], align 4 |
| 124 |
; CHECK-NEXT: [[R1:%.*]] = bitcast <4 x i8> [[LV]] to i32 |
| 125 |
; CHECK-NEXT: [[R2:%.*]] = extractelement <4 x i8> [[LV]], i32 0 |
| 126 |
; CHECK-NEXT: [[R2_EXT:%.*]] = zext i8 [[R2]] to i32 |
| 127 |
; CHECK-NEXT: [[ADD:%.*]] = add i32 [[R1]], [[R2_EXT]] |
| 128 |
; CHECK-NEXT: ret i32 [[ADD]] |
| 129 |
; |
| 130 |
%lv = load <4 x i8>, ptr %x |
| 131 |
%r1 = bitcast <4 x i8> %lv to i32 |
| 132 |
%r2 = extractelement <4 x i8> %lv, i32 0 |
| 133 |
%r2.ext = zext i8 %r2 to i32 |
| 134 |
%add = add i32 %r1, %r2.ext |
| 135 |
ret i32 %add |
| 136 |
} |