clang/test/CodeGen/AArch64/bf16-lane-intrinsics.c

   1 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py
   2 // RUN: %clang_cc1 -triple aarch64 -target-feature +neon -target-feature +bf16 \
   3 // RUN:  -disable-O0-optnone -emit-llvm %s -o - | opt -S -passes=mem2reg | FileCheck --check-prefix=CHECK-LE %s
   4 // RUN: %clang_cc1 -triple aarch64_be -target-feature +neon -target-feature +bf16 \
   5 // RUN:  -disable-O0-optnone -emit-llvm %s -o - | opt -S -passes=mem2reg | FileCheck --check-prefix=CHECK-BE %s
   6
   7 // REQUIRES: aarch64-registered-target || arm-registered-target
   8
   9 #include <arm_neon.h>
  10
  11 // CHECK-LE-LABEL: @test_vcopy_lane_bf16_v1(
  12 // CHECK-LE-NEXT:  entry:
  13 // CHECK-LE-NEXT:    [[VGET_LANE:%.*]] = extractelement <4 x bfloat> [[B:%.*]], i32 3
  14 // CHECK-LE-NEXT:    [[VSET_LANE:%.*]] = insertelement <4 x bfloat> [[A:%.*]], bfloat [[VGET_LANE]], i32 1
  15 // CHECK-LE-NEXT:    ret <4 x bfloat> [[VSET_LANE]]
  16 //
  17 // CHECK-BE-LABEL: @test_vcopy_lane_bf16_v1(
  18 // CHECK-BE-NEXT:  entry:
  19 // CHECK-BE-NEXT:    [[SHUFFLE:%.*]] = shufflevector <4 x bfloat> [[A:%.*]], <4 x bfloat> [[A]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  20 // CHECK-BE-NEXT:    [[SHUFFLE1:%.*]] = shufflevector <4 x bfloat> [[B:%.*]], <4 x bfloat> [[B]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  21 // CHECK-BE-NEXT:    [[VGET_LANE:%.*]] = extractelement <4 x bfloat> [[SHUFFLE1]], i32 3
  22 // CHECK-BE-NEXT:    [[VSET_LANE:%.*]] = insertelement <4 x bfloat> [[SHUFFLE]], bfloat [[VGET_LANE]], i32 1
  23 // CHECK-BE-NEXT:    [[SHUFFLE5:%.*]] = shufflevector <4 x bfloat> [[VSET_LANE]], <4 x bfloat> [[VSET_LANE]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  24 // CHECK-BE-NEXT:    ret <4 x bfloat> [[SHUFFLE5]]
  25 //
  26 bfloat16x4_t test_vcopy_lane_bf16_v1(bfloat16x4_t a, bfloat16x4_t b) {
  27   return vcopy_lane_bf16(a, 1, b, 3);
  28 }
  29
  30 // CHECK-LE-LABEL: @test_vcopy_lane_bf16_v2(
  31 // CHECK-LE-NEXT:  entry:
  32 // CHECK-LE-NEXT:    [[VGET_LANE:%.*]] = extractelement <4 x bfloat> [[B:%.*]], i32 0
  33 // CHECK-LE-NEXT:    [[VSET_LANE:%.*]] = insertelement <4 x bfloat> [[A:%.*]], bfloat [[VGET_LANE]], i32 2
  34 // CHECK-LE-NEXT:    ret <4 x bfloat> [[VSET_LANE]]
  35 //
  36 // CHECK-BE-LABEL: @test_vcopy_lane_bf16_v2(
  37 // CHECK-BE-NEXT:  entry:
  38 // CHECK-BE-NEXT:    [[SHUFFLE:%.*]] = shufflevector <4 x bfloat> [[A:%.*]], <4 x bfloat> [[A]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  39 // CHECK-BE-NEXT:    [[SHUFFLE1:%.*]] = shufflevector <4 x bfloat> [[B:%.*]], <4 x bfloat> [[B]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  40 // CHECK-BE-NEXT:    [[VGET_LANE:%.*]] = extractelement <4 x bfloat> [[SHUFFLE1]], i32 0
  41 // CHECK-BE-NEXT:    [[VSET_LANE:%.*]] = insertelement <4 x bfloat> [[SHUFFLE]], bfloat [[VGET_LANE]], i32 2
  42 // CHECK-BE-NEXT:    [[SHUFFLE5:%.*]] = shufflevector <4 x bfloat> [[VSET_LANE]], <4 x bfloat> [[VSET_LANE]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  43 // CHECK-BE-NEXT:    ret <4 x bfloat> [[SHUFFLE5]]
  44 //
  45 bfloat16x4_t test_vcopy_lane_bf16_v2(bfloat16x4_t a, bfloat16x4_t b) {
  46   return vcopy_lane_bf16(a, 2, b, 0);
  47 }
  48
  49 // CHECK-LE-LABEL: @test_vcopyq_lane_bf16_v1(
  50 // CHECK-LE-NEXT:  entry:
  51 // CHECK-LE-NEXT:    [[VGET_LANE:%.*]] = extractelement <4 x bfloat> [[B:%.*]], i32 2
  52 // CHECK-LE-NEXT:    [[VSET_LANE:%.*]] = insertelement <8 x bfloat> [[A:%.*]], bfloat [[VGET_LANE]], i32 0
  53 // CHECK-LE-NEXT:    ret <8 x bfloat> [[VSET_LANE]]
  54 //
  55 // CHECK-BE-LABEL: @test_vcopyq_lane_bf16_v1(
  56 // CHECK-BE-NEXT:  entry:
  57 // CHECK-BE-NEXT:    [[SHUFFLE:%.*]] = shufflevector <8 x bfloat> [[A:%.*]], <8 x bfloat> [[A]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
  58 // CHECK-BE-NEXT:    [[SHUFFLE1:%.*]] = shufflevector <4 x bfloat> [[B:%.*]], <4 x bfloat> [[B]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  59 // CHECK-BE-NEXT:    [[VGET_LANE:%.*]] = extractelement <4 x bfloat> [[SHUFFLE1]], i32 2
  60 // CHECK-BE-NEXT:    [[VSET_LANE:%.*]] = insertelement <8 x bfloat> [[SHUFFLE]], bfloat [[VGET_LANE]], i32 0
  61 // CHECK-BE-NEXT:    [[SHUFFLE5:%.*]] = shufflevector <8 x bfloat> [[VSET_LANE]], <8 x bfloat> [[VSET_LANE]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
  62 // CHECK-BE-NEXT:    ret <8 x bfloat> [[SHUFFLE5]]
  63 //
  64 bfloat16x8_t test_vcopyq_lane_bf16_v1(bfloat16x8_t a, bfloat16x4_t b) {
  65   return vcopyq_lane_bf16(a, 0, b, 2);
  66 }
  67
  68 // CHECK-LE-LABEL: @test_vcopyq_lane_bf16_v2(
  69 // CHECK-LE-NEXT:  entry:
  70 // CHECK-LE-NEXT:    [[VGET_LANE:%.*]] = extractelement <4 x bfloat> [[B:%.*]], i32 0
  71 // CHECK-LE-NEXT:    [[VSET_LANE:%.*]] = insertelement <8 x bfloat> [[A:%.*]], bfloat [[VGET_LANE]], i32 6
  72 // CHECK-LE-NEXT:    ret <8 x bfloat> [[VSET_LANE]]
  73 //
  74 // CHECK-BE-LABEL: @test_vcopyq_lane_bf16_v2(
  75 // CHECK-BE-NEXT:  entry:
  76 // CHECK-BE-NEXT:    [[SHUFFLE:%.*]] = shufflevector <8 x bfloat> [[A:%.*]], <8 x bfloat> [[A]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
  77 // CHECK-BE-NEXT:    [[SHUFFLE1:%.*]] = shufflevector <4 x bfloat> [[B:%.*]], <4 x bfloat> [[B]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  78 // CHECK-BE-NEXT:    [[VGET_LANE:%.*]] = extractelement <4 x bfloat> [[SHUFFLE1]], i32 0
  79 // CHECK-BE-NEXT:    [[VSET_LANE:%.*]] = insertelement <8 x bfloat> [[SHUFFLE]], bfloat [[VGET_LANE]], i32 6
  80 // CHECK-BE-NEXT:    [[SHUFFLE5:%.*]] = shufflevector <8 x bfloat> [[VSET_LANE]], <8 x bfloat> [[VSET_LANE]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
  81 // CHECK-BE-NEXT:    ret <8 x bfloat> [[SHUFFLE5]]
  82 //
  83 bfloat16x8_t test_vcopyq_lane_bf16_v2(bfloat16x8_t a, bfloat16x4_t b) {
  84   return vcopyq_lane_bf16(a, 6, b, 0);
  85 }
  86
  87 // CHECK-LE-LABEL: @test_vcopy_laneq_bf16_v1(
  88 // CHECK-LE-NEXT:  entry:
  89 // CHECK-LE-NEXT:    [[VGETQ_LANE:%.*]] = extractelement <8 x bfloat> [[B:%.*]], i32 7
  90 // CHECK-LE-NEXT:    [[VSET_LANE:%.*]] = insertelement <4 x bfloat> [[A:%.*]], bfloat [[VGETQ_LANE]], i32 0
  91 // CHECK-LE-NEXT:    ret <4 x bfloat> [[VSET_LANE]]
  92 //
  93 // CHECK-BE-LABEL: @test_vcopy_laneq_bf16_v1(
  94 // CHECK-BE-NEXT:  entry:
  95 // CHECK-BE-NEXT:    [[SHUFFLE:%.*]] = shufflevector <4 x bfloat> [[A:%.*]], <4 x bfloat> [[A]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
  96 // CHECK-BE-NEXT:    [[SHUFFLE1:%.*]] = shufflevector <8 x bfloat> [[B:%.*]], <8 x bfloat> [[B]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
  97 // CHECK-BE-NEXT:    [[VGETQ_LANE:%.*]] = extractelement <8 x bfloat> [[SHUFFLE1]], i32 7
  98 // CHECK-BE-NEXT:    [[VSET_LANE:%.*]] = insertelement <4 x bfloat> [[SHUFFLE]], bfloat [[VGETQ_LANE]], i32 0
  99 // CHECK-BE-NEXT:    [[SHUFFLE5:%.*]] = shufflevector <4 x bfloat> [[VSET_LANE]], <4 x bfloat> [[VSET_LANE]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
 100 // CHECK-BE-NEXT:    ret <4 x bfloat> [[SHUFFLE5]]
 101 //
 102 bfloat16x4_t test_vcopy_laneq_bf16_v1(bfloat16x4_t a, bfloat16x8_t b) {
 103   return vcopy_laneq_bf16(a, 0, b, 7);
 104 }
 105
 106 // CHECK-LE-LABEL: @test_vcopy_laneq_bf16_v2(
 107 // CHECK-LE-NEXT:  entry:
 108 // CHECK-LE-NEXT:    [[VGETQ_LANE:%.*]] = extractelement <8 x bfloat> [[B:%.*]], i32 4
 109 // CHECK-LE-NEXT:    [[VSET_LANE:%.*]] = insertelement <4 x bfloat> [[A:%.*]], bfloat [[VGETQ_LANE]], i32 3
 110 // CHECK-LE-NEXT:    ret <4 x bfloat> [[VSET_LANE]]
 111 //
 112 // CHECK-BE-LABEL: @test_vcopy_laneq_bf16_v2(
 113 // CHECK-BE-NEXT:  entry:
 114 // CHECK-BE-NEXT:    [[SHUFFLE:%.*]] = shufflevector <4 x bfloat> [[A:%.*]], <4 x bfloat> [[A]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
 115 // CHECK-BE-NEXT:    [[SHUFFLE1:%.*]] = shufflevector <8 x bfloat> [[B:%.*]], <8 x bfloat> [[B]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
 116 // CHECK-BE-NEXT:    [[VGETQ_LANE:%.*]] = extractelement <8 x bfloat> [[SHUFFLE1]], i32 4
 117 // CHECK-BE-NEXT:    [[VSET_LANE:%.*]] = insertelement <4 x bfloat> [[SHUFFLE]], bfloat [[VGETQ_LANE]], i32 3
 118 // CHECK-BE-NEXT:    [[SHUFFLE5:%.*]] = shufflevector <4 x bfloat> [[VSET_LANE]], <4 x bfloat> [[VSET_LANE]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>
 119 // CHECK-BE-NEXT:    ret <4 x bfloat> [[SHUFFLE5]]
 120 //
 121 bfloat16x4_t test_vcopy_laneq_bf16_v2(bfloat16x4_t a, bfloat16x8_t b) {
 122   return vcopy_laneq_bf16(a, 3, b, 4);
 123 }
 124
 125 // CHECK-LE-LABEL: @test_vcopyq_laneq_bf16_v1(
 126 // CHECK-LE-NEXT:  entry:
 127 // CHECK-LE-NEXT:    [[VGETQ_LANE:%.*]] = extractelement <8 x bfloat> [[B:%.*]], i32 7
 128 // CHECK-LE-NEXT:    [[VSET_LANE:%.*]] = insertelement <8 x bfloat> [[A:%.*]], bfloat [[VGETQ_LANE]], i32 3
 129 // CHECK-LE-NEXT:    ret <8 x bfloat> [[VSET_LANE]]
 130 //
 131 // CHECK-BE-LABEL: @test_vcopyq_laneq_bf16_v1(
 132 // CHECK-BE-NEXT:  entry:
 133 // CHECK-BE-NEXT:    [[SHUFFLE:%.*]] = shufflevector <8 x bfloat> [[A:%.*]], <8 x bfloat> [[A]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
 134 // CHECK-BE-NEXT:    [[SHUFFLE1:%.*]] = shufflevector <8 x bfloat> [[B:%.*]], <8 x bfloat> [[B]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
 135 // CHECK-BE-NEXT:    [[VGETQ_LANE:%.*]] = extractelement <8 x bfloat> [[SHUFFLE1]], i32 7
 136 // CHECK-BE-NEXT:    [[VSET_LANE:%.*]] = insertelement <8 x bfloat> [[SHUFFLE]], bfloat [[VGETQ_LANE]], i32 3
 137 // CHECK-BE-NEXT:    [[SHUFFLE5:%.*]] = shufflevector <8 x bfloat> [[VSET_LANE]], <8 x bfloat> [[VSET_LANE]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
 138 // CHECK-BE-NEXT:    ret <8 x bfloat> [[SHUFFLE5]]
 139 //
 140 bfloat16x8_t test_vcopyq_laneq_bf16_v1(bfloat16x8_t a, bfloat16x8_t b) {
 141   return vcopyq_laneq_bf16(a, 3, b, 7);
 142
 143 }
 144
 145 // CHECK-LE-LABEL: @test_vcopyq_laneq_bf16_v2(
 146 // CHECK-LE-NEXT:  entry:
 147 // CHECK-LE-NEXT:    [[VGETQ_LANE:%.*]] = extractelement <8 x bfloat> [[B:%.*]], i32 2
 148 // CHECK-LE-NEXT:    [[VSET_LANE:%.*]] = insertelement <8 x bfloat> [[A:%.*]], bfloat [[VGETQ_LANE]], i32 6
 149 // CHECK-LE-NEXT:    ret <8 x bfloat> [[VSET_LANE]]
 150 //
 151 // CHECK-BE-LABEL: @test_vcopyq_laneq_bf16_v2(
 152 // CHECK-BE-NEXT:  entry:
 153 // CHECK-BE-NEXT:    [[SHUFFLE:%.*]] = shufflevector <8 x bfloat> [[A:%.*]], <8 x bfloat> [[A]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
 154 // CHECK-BE-NEXT:    [[SHUFFLE1:%.*]] = shufflevector <8 x bfloat> [[B:%.*]], <8 x bfloat> [[B]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
 155 // CHECK-BE-NEXT:    [[VGETQ_LANE:%.*]] = extractelement <8 x bfloat> [[SHUFFLE1]], i32 2
 156 // CHECK-BE-NEXT:    [[VSET_LANE:%.*]] = insertelement <8 x bfloat> [[SHUFFLE]], bfloat [[VGETQ_LANE]], i32 6
 157 // CHECK-BE-NEXT:    [[SHUFFLE5:%.*]] = shufflevector <8 x bfloat> [[VSET_LANE]], <8 x bfloat> [[VSET_LANE]], <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
 158 // CHECK-BE-NEXT:    ret <8 x bfloat> [[SHUFFLE5]]
 159 //
 160 bfloat16x8_t test_vcopyq_laneq_bf16_v2(bfloat16x8_t a, bfloat16x8_t b) {
 161   return vcopyq_laneq_bf16(a, 6, b, 2);
 162 }
 163