CbC/CbC_llvm: clang/test/CodeGen/aarch64-bf16-ldst-intrinsics.c annotate

annotate clang/test/CodeGen/aarch64-bf16-ldst-intrinsics.c @ 266:00f31e85ec16 default tip

Added tag current for changeset 31d058e83c98

author	Shinji KONO <kono@ie.u-ryukyu.ac.jp>
date	Sat, 14 Oct 2023 10:13:55 +0900
parents	1f2b6ac9f198
children

rev	line source
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	2 // RUN: %clang_cc1 -triple aarch64 -target-feature +neon -target-feature +bf16 \
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	3 // RUN: -O2 -emit-llvm %s -o - \| FileCheck %s --check-prefixes=CHECK,CHECK64
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	4 // RUN: %clang_cc1 -triple armv8.6a-arm-none-eabi -target-feature +neon -target-feature +bf16 -mfloat-abi hard \
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	5 // RUN: -O2 -emit-llvm %s -o - \| FileCheck %s --check-prefixes=CHECK,CHECK32
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	6
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	7 // REQUIRES: arm-registered-target,aarch64-registered-target
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	8
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	9 #include "arm_neon.h"
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	10
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	11 // CHECK-LABEL: @test_vld1_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	12 // CHECK-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	13 // CHECK-NEXT: [[TMP1:%.]] = load <4 x bfloat>, ptr [[PTR:%.]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	14 // CHECK-NEXT: ret <4 x bfloat> [[TMP1]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	15 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	16 bfloat16x4_t test_vld1_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	17 return vld1_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	18 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	19
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	20 // CHECK-LABEL: @test_vld1q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	21 // CHECK-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	22 // CHECK-NEXT: [[TMP1:%.]] = load <8 x bfloat>, ptr [[PTR:%.]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	23 // CHECK-NEXT: ret <8 x bfloat> [[TMP1]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	24 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	25 bfloat16x8_t test_vld1q_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	26 return vld1q_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	27 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	28
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	29 // CHECK-LABEL: @test_vld1_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	30 // CHECK-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	31 // CHECK-NEXT: [[TMP0:%.]] = load bfloat, ptr [[PTR:%.]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	32 // CHECK-NEXT: [[VLD1_LANE:%.]] = insertelement <4 x bfloat> [[SRC:%.]], bfloat [[TMP0]], i64 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	33 // CHECK-NEXT: ret <4 x bfloat> [[VLD1_LANE]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	34 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	35 bfloat16x4_t test_vld1_lane_bf16(bfloat16_t const *ptr, bfloat16x4_t src) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	36 return vld1_lane_bf16(ptr, src, 0);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	37 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	38
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	39 // CHECK-LABEL: @test_vld1q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	40 // CHECK-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	41 // CHECK-NEXT: [[TMP0:%.]] = load bfloat, ptr [[PTR:%.]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	42 // CHECK-NEXT: [[VLD1_LANE:%.]] = insertelement <8 x bfloat> [[SRC:%.]], bfloat [[TMP0]], i64 7
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	43 // CHECK-NEXT: ret <8 x bfloat> [[VLD1_LANE]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	44 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	45 bfloat16x8_t test_vld1q_lane_bf16(bfloat16_t const *ptr, bfloat16x8_t src) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	46 return vld1q_lane_bf16(ptr, src, 7);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	47 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	48
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	49 // CHECK-LABEL: @test_vld1_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	50 // CHECK-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	51 // CHECK-NEXT: [[TMP0:%.]] = load bfloat, ptr [[PTR:%.]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	52 // CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x bfloat> poison, bfloat [[TMP0]], i64 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	53 // CHECK-NEXT: [[LANE:%.*]] = shufflevector <4 x bfloat> [[TMP1]], <4 x bfloat> poison, <4 x i32> zeroinitializer
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	54 // CHECK-NEXT: ret <4 x bfloat> [[LANE]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	55 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	56 bfloat16x4_t test_vld1_dup_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	57 return vld1_dup_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	58 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	59
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	60 // CHECK64-LABEL: @test_vld1_bf16_x2(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	61 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	62 // CHECK64-NEXT: [[VLD1XN:%.]] = tail call { <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld1x2.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	63 // CHECK64-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	64 // CHECK64-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	65 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X2_T:%.]] poison, <4 x bfloat> [[VLD1XN_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	66 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X2_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD1XN_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	67 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X2_T]] [[DOTFCA_0_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	68 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	69 // CHECK32-LABEL: @test_vld1_bf16_x2(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	70 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	71 // CHECK32-NEXT: [[VLD1XN:%.]] = tail call { <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld1x2.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	72 // CHECK32-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	73 // CHECK32-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	74 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	75 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	76 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [2 x <2 x i32>] poison, <2 x i32> [[TMP0]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	77 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [2 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP1]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	78 // CHECK32-NEXT: ret [2 x <2 x i32>] [[DOTFCA_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	79 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	80 bfloat16x4x2_t test_vld1_bf16_x2(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	81 return vld1_bf16_x2(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	82 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	83
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	84 // CHECK64-LABEL: @test_vld1q_bf16_x2(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	85 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	86 // CHECK64-NEXT: [[VLD1XN:%.]] = tail call { <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld1x2.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	87 // CHECK64-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	88 // CHECK64-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	89 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X2_T:%.]] poison, <8 x bfloat> [[VLD1XN_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	90 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X2_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD1XN_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	91 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X2_T]] [[DOTFCA_0_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	92 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	93 // CHECK32-LABEL: @test_vld1q_bf16_x2(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	94 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	95 // CHECK32-NEXT: [[VLD1XN:%.]] = tail call { <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld1x2.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	96 // CHECK32-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	97 // CHECK32-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	98 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	99 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	100 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [2 x <4 x i32>] poison, <4 x i32> [[TMP0]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	101 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [2 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP1]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	102 // CHECK32-NEXT: ret [2 x <4 x i32>] [[DOTFCA_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	103 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	104 bfloat16x8x2_t test_vld1q_bf16_x2(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	105 return vld1q_bf16_x2(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	106 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	107
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	108 // CHECK64-LABEL: @test_vld1_bf16_x3(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	109 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	110 // CHECK64-NEXT: [[VLD1XN:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld1x3.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	111 // CHECK64-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	112 // CHECK64-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	113 // CHECK64-NEXT: [[VLD1XN_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	114 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X3_T:%.]] poison, <4 x bfloat> [[VLD1XN_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	115 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD1XN_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	116 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_1_INSERT]], <4 x bfloat> [[VLD1XN_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	117 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	118 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	119 // CHECK32-LABEL: @test_vld1_bf16_x3(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	120 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	121 // CHECK32-NEXT: [[VLD1XN:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld1x3.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	122 // CHECK32-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	123 // CHECK32-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	124 // CHECK32-NEXT: [[VLD1XN_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	125 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	126 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	127 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_2_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	128 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [3 x <2 x i32>] poison, <2 x i32> [[TMP0]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	129 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [3 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP1]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	130 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [3 x <2 x i32>] [[DOTFCA_1_INSERT]], <2 x i32> [[TMP2]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	131 // CHECK32-NEXT: ret [3 x <2 x i32>] [[DOTFCA_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	132 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	133 bfloat16x4x3_t test_vld1_bf16_x3(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	134 return vld1_bf16_x3(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	135 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	136
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	137 // CHECK64-LABEL: @test_vld1q_bf16_x3(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	138 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	139 // CHECK64-NEXT: [[VLD1XN:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld1x3.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	140 // CHECK64-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	141 // CHECK64-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	142 // CHECK64-NEXT: [[VLD1XN_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	143 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X3_T:%.]] poison, <8 x bfloat> [[VLD1XN_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	144 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD1XN_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	145 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_1_INSERT]], <8 x bfloat> [[VLD1XN_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	146 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	147 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	148 // CHECK32-LABEL: @test_vld1q_bf16_x3(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	149 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	150 // CHECK32-NEXT: [[VLD1XN:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld1x3.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	151 // CHECK32-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	152 // CHECK32-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	153 // CHECK32-NEXT: [[VLD1XN_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	154 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	155 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	156 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_2_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	157 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [3 x <4 x i32>] poison, <4 x i32> [[TMP0]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	158 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [3 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP1]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	159 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [3 x <4 x i32>] [[DOTFCA_1_INSERT]], <4 x i32> [[TMP2]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	160 // CHECK32-NEXT: ret [3 x <4 x i32>] [[DOTFCA_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	161 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	162 bfloat16x8x3_t test_vld1q_bf16_x3(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	163 return vld1q_bf16_x3(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	164 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	165
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	166 // CHECK64-LABEL: @test_vld1_bf16_x4(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	167 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	168 // CHECK64-NEXT: [[VLD1XN:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld1x4.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	169 // CHECK64-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	170 // CHECK64-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	171 // CHECK64-NEXT: [[VLD1XN_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	172 // CHECK64-NEXT: [[VLD1XN_FCA_3_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	173 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X4_T:%.]] poison, <4 x bfloat> [[VLD1XN_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	174 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD1XN_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	175 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_1_INSERT]], <4 x bfloat> [[VLD1XN_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	176 // CHECK64-NEXT: [[DOTFCA_0_3_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_2_INSERT]], <4 x bfloat> [[VLD1XN_FCA_3_EXTRACT]], 0, 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	177 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	178 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	179 // CHECK32-LABEL: @test_vld1_bf16_x4(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	180 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	181 // CHECK32-NEXT: [[VLD1XN:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld1x4.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	182 // CHECK32-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	183 // CHECK32-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	184 // CHECK32-NEXT: [[VLD1XN_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	185 // CHECK32-NEXT: [[VLD1XN_FCA_3_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD1XN]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	186 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	187 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	188 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_2_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	189 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x bfloat> [[VLD1XN_FCA_3_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	190 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [4 x <2 x i32>] poison, <2 x i32> [[TMP0]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	191 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP1]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	192 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_1_INSERT]], <2 x i32> [[TMP2]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	193 // CHECK32-NEXT: [[DOTFCA_3_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_2_INSERT]], <2 x i32> [[TMP3]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	194 // CHECK32-NEXT: ret [4 x <2 x i32>] [[DOTFCA_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	195 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	196 bfloat16x4x4_t test_vld1_bf16_x4(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	197 return vld1_bf16_x4(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	198 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	199
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	200 // CHECK64-LABEL: @test_vld1q_bf16_x4(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	201 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	202 // CHECK64-NEXT: [[VLD1XN:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld1x4.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	203 // CHECK64-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	204 // CHECK64-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	205 // CHECK64-NEXT: [[VLD1XN_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	206 // CHECK64-NEXT: [[VLD1XN_FCA_3_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	207 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X4_T:%.]] poison, <8 x bfloat> [[VLD1XN_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	208 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD1XN_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	209 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_1_INSERT]], <8 x bfloat> [[VLD1XN_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	210 // CHECK64-NEXT: [[DOTFCA_0_3_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_2_INSERT]], <8 x bfloat> [[VLD1XN_FCA_3_EXTRACT]], 0, 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	211 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	212 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	213 // CHECK32-LABEL: @test_vld1q_bf16_x4(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	214 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	215 // CHECK32-NEXT: [[VLD1XN:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld1x4.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	216 // CHECK32-NEXT: [[VLD1XN_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	217 // CHECK32-NEXT: [[VLD1XN_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	218 // CHECK32-NEXT: [[VLD1XN_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	219 // CHECK32-NEXT: [[VLD1XN_FCA_3_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD1XN]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	220 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	221 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	222 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_2_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	223 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <8 x bfloat> [[VLD1XN_FCA_3_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	224 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [4 x <4 x i32>] poison, <4 x i32> [[TMP0]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	225 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP1]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	226 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_1_INSERT]], <4 x i32> [[TMP2]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	227 // CHECK32-NEXT: [[DOTFCA_3_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_2_INSERT]], <4 x i32> [[TMP3]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	228 // CHECK32-NEXT: ret [4 x <4 x i32>] [[DOTFCA_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	229 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	230 bfloat16x8x4_t test_vld1q_bf16_x4(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	231 return vld1q_bf16_x4(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	232 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	233
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	234 // CHECK-LABEL: @test_vld1q_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	235 // CHECK-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	236 // CHECK-NEXT: [[TMP0:%.]] = load bfloat, ptr [[PTR:%.]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	237 // CHECK-NEXT: [[TMP1:%.*]] = insertelement <8 x bfloat> poison, bfloat [[TMP0]], i64 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	238 // CHECK-NEXT: [[LANE:%.*]] = shufflevector <8 x bfloat> [[TMP1]], <8 x bfloat> poison, <8 x i32> zeroinitializer
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	239 // CHECK-NEXT: ret <8 x bfloat> [[LANE]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	240 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	241 bfloat16x8_t test_vld1q_dup_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	242 return vld1q_dup_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	243 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	244
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	245 // CHECK64-LABEL: @test_vld2_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	246 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	247 // CHECK64-NEXT: [[VLD2:%.]] = tail call { <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld2.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	248 // CHECK64-NEXT: [[VLD2_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	249 // CHECK64-NEXT: [[VLD2_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	250 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X2_T:%.]] poison, <4 x bfloat> [[VLD2_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	251 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X2_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD2_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	252 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X2_T]] [[DOTFCA_0_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	253 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	254 // CHECK32-LABEL: @test_vld2_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	255 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	256 // CHECK32-NEXT: [[VLD2_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld2.v4bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	257 // CHECK32-NEXT: [[VLD2_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	258 // CHECK32-NEXT: [[VLD2_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	259 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD2_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	260 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x bfloat> [[VLD2_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	261 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [2 x <2 x i32>] poison, <2 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	262 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [2 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	263 // CHECK32-NEXT: ret [2 x <2 x i32>] [[DOTFCA_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	264 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	265 bfloat16x4x2_t test_vld2_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	266 return vld2_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	267 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	268
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	269 // CHECK64-LABEL: @test_vld2q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	270 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	271 // CHECK64-NEXT: [[VLD2:%.]] = tail call { <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld2.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	272 // CHECK64-NEXT: [[VLD2_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	273 // CHECK64-NEXT: [[VLD2_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	274 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X2_T:%.]] poison, <8 x bfloat> [[VLD2_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	275 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X2_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD2_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	276 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X2_T]] [[DOTFCA_0_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	277 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	278 // CHECK32-LABEL: @test_vld2q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	279 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	280 // CHECK32-NEXT: [[VLD2Q_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld2.v8bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	281 // CHECK32-NEXT: [[VLD2Q_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2Q_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	282 // CHECK32-NEXT: [[VLD2Q_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2Q_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	283 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD2Q_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	284 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <8 x bfloat> [[VLD2Q_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	285 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [2 x <4 x i32>] poison, <4 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	286 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [2 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	287 // CHECK32-NEXT: ret [2 x <4 x i32>] [[DOTFCA_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	288 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	289 bfloat16x8x2_t test_vld2q_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	290 return vld2q_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	291 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	292
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	293 // CHECK64-LABEL: @test_vld2_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	294 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	295 // CHECK64-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <4 x bfloat>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	296 // CHECK64-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <4 x bfloat>] [[SRC_COERCE]], 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	297 // CHECK64-NEXT: [[VLD2_LANE:%.]] = tail call { <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld2lane.v4bf16.p0(<4 x bfloat> [[SRC_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[SRC_COERCE_FCA_1_EXTRACT]], i64 1, ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	298 // CHECK64-NEXT: [[VLD2_LANE_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2_LANE]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	299 // CHECK64-NEXT: [[VLD2_LANE_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2_LANE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	300 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X2_T:%.]] poison, <4 x bfloat> [[VLD2_LANE_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	301 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X2_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD2_LANE_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	302 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X2_T]] [[DOTFCA_0_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	303 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	304 // CHECK32-LABEL: @test_vld2_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	305 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	306 // CHECK32-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <2 x i32>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	307 // CHECK32-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <2 x i32>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	308 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	309 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	310 // CHECK32-NEXT: [[VLD2_LANE_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld2lane.v4bf16.p0(ptr [[PTR:%.]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], i32 1, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	311 // CHECK32-NEXT: [[VLD2_LANE_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2_LANE_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	312 // CHECK32-NEXT: [[VLD2_LANE_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2_LANE_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	313 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x bfloat> [[VLD2_LANE_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	314 // CHECK32-NEXT: [[TMP4:%.*]] = bitcast <4 x bfloat> [[VLD2_LANE_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	315 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [2 x <2 x i32>] poison, <2 x i32> [[TMP3]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	316 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [2 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP4]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	317 // CHECK32-NEXT: ret [2 x <2 x i32>] [[DOTFCA_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	318 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	319 bfloat16x4x2_t test_vld2_lane_bf16(bfloat16_t const *ptr, bfloat16x4x2_t src) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	320 return vld2_lane_bf16(ptr, src, 1);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	321 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	322
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	323 // CHECK64-LABEL: @test_vld2q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	324 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	325 // CHECK64-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <8 x bfloat>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	326 // CHECK64-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x bfloat>] [[SRC_COERCE]], 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	327 // CHECK64-NEXT: [[VLD2_LANE:%.]] = tail call { <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld2lane.v8bf16.p0(<8 x bfloat> [[SRC_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[SRC_COERCE_FCA_1_EXTRACT]], i64 7, ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	328 // CHECK64-NEXT: [[VLD2_LANE_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2_LANE]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	329 // CHECK64-NEXT: [[VLD2_LANE_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2_LANE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	330 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X2_T:%.]] poison, <8 x bfloat> [[VLD2_LANE_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	331 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X2_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD2_LANE_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	332 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X2_T]] [[DOTFCA_0_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	333 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	334 // CHECK32-LABEL: @test_vld2q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	335 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	336 // CHECK32-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <4 x i32>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	337 // CHECK32-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <4 x i32>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	338 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	339 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	340 // CHECK32-NEXT: [[VLD2Q_LANE_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld2lane.v8bf16.p0(ptr [[PTR:%.]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], i32 7, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	341 // CHECK32-NEXT: [[VLD2Q_LANE_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2Q_LANE_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	342 // CHECK32-NEXT: [[VLD2Q_LANE_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2Q_LANE_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	343 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <8 x bfloat> [[VLD2Q_LANE_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	344 // CHECK32-NEXT: [[TMP4:%.*]] = bitcast <8 x bfloat> [[VLD2Q_LANE_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	345 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [2 x <4 x i32>] poison, <4 x i32> [[TMP3]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	346 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [2 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP4]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	347 // CHECK32-NEXT: ret [2 x <4 x i32>] [[DOTFCA_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	348 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	349 bfloat16x8x2_t test_vld2q_lane_bf16(bfloat16_t const *ptr, bfloat16x8x2_t src) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	350 return vld2q_lane_bf16(ptr, src, 7);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	351 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	352
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	353 // CHECK64-LABEL: @test_vld3_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	354 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	355 // CHECK64-NEXT: [[VLD3:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld3.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	356 // CHECK64-NEXT: [[VLD3_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	357 // CHECK64-NEXT: [[VLD3_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	358 // CHECK64-NEXT: [[VLD3_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	359 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X3_T:%.]] poison, <4 x bfloat> [[VLD3_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	360 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD3_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	361 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_1_INSERT]], <4 x bfloat> [[VLD3_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	362 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	363 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	364 // CHECK32-LABEL: @test_vld3_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	365 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	366 // CHECK32-NEXT: [[VLD3_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld3.v4bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	367 // CHECK32-NEXT: [[VLD3_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	368 // CHECK32-NEXT: [[VLD3_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	369 // CHECK32-NEXT: [[VLD3_V_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	370 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD3_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	371 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x bfloat> [[VLD3_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	372 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x bfloat> [[VLD3_V_FCA_2_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	373 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [3 x <2 x i32>] poison, <2 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	374 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [3 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	375 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [3 x <2 x i32>] [[DOTFCA_1_INSERT]], <2 x i32> [[TMP3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	376 // CHECK32-NEXT: ret [3 x <2 x i32>] [[DOTFCA_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	377 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	378 bfloat16x4x3_t test_vld3_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	379 return vld3_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	380 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	381
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	382 // CHECK64-LABEL: @test_vld3q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	383 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	384 // CHECK64-NEXT: [[VLD3:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld3.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	385 // CHECK64-NEXT: [[VLD3_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	386 // CHECK64-NEXT: [[VLD3_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	387 // CHECK64-NEXT: [[VLD3_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	388 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X3_T:%.]] poison, <8 x bfloat> [[VLD3_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	389 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD3_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	390 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_1_INSERT]], <8 x bfloat> [[VLD3_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	391 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	392 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	393 // CHECK32-LABEL: @test_vld3q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	394 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	395 // CHECK32-NEXT: [[VLD3Q_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld3.v8bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	396 // CHECK32-NEXT: [[VLD3Q_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	397 // CHECK32-NEXT: [[VLD3Q_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	398 // CHECK32-NEXT: [[VLD3Q_V_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	399 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD3Q_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	400 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <8 x bfloat> [[VLD3Q_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	401 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <8 x bfloat> [[VLD3Q_V_FCA_2_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	402 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [3 x <4 x i32>] poison, <4 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	403 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [3 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	404 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [3 x <4 x i32>] [[DOTFCA_1_INSERT]], <4 x i32> [[TMP3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	405 // CHECK32-NEXT: ret [3 x <4 x i32>] [[DOTFCA_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	406 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	407 bfloat16x8x3_t test_vld3q_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	408 return vld3q_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	409 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	410
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	411 // CHECK64-LABEL: @test_vld3_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	412 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	413 // CHECK64-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <4 x bfloat>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	414 // CHECK64-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <4 x bfloat>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	415 // CHECK64-NEXT: [[SRC_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <4 x bfloat>] [[SRC_COERCE]], 2
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	416 // CHECK64-NEXT: [[VLD3_LANE:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld3lane.v4bf16.p0(<4 x bfloat> [[SRC_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[SRC_COERCE_FCA_1_EXTRACT]], <4 x bfloat> [[SRC_COERCE_FCA_2_EXTRACT]], i64 1, ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	417 // CHECK64-NEXT: [[VLD3_LANE_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_LANE]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	418 // CHECK64-NEXT: [[VLD3_LANE_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_LANE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	419 // CHECK64-NEXT: [[VLD3_LANE_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_LANE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	420 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X3_T:%.]] poison, <4 x bfloat> [[VLD3_LANE_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	421 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD3_LANE_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	422 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_1_INSERT]], <4 x bfloat> [[VLD3_LANE_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	423 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	424 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	425 // CHECK32-LABEL: @test_vld3_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	426 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	427 // CHECK32-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <2 x i32>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	428 // CHECK32-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <2 x i32>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	429 // CHECK32-NEXT: [[SRC_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <2 x i32>] [[SRC_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	430 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	431 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	432 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_2_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	433 // CHECK32-NEXT: [[VLD3_LANE_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld3lane.v4bf16.p0(ptr [[PTR:%.]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], <4 x bfloat> [[TMP2]], i32 1, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	434 // CHECK32-NEXT: [[VLD3_LANE_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_LANE_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	435 // CHECK32-NEXT: [[VLD3_LANE_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_LANE_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	436 // CHECK32-NEXT: [[VLD3_LANE_V_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_LANE_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	437 // CHECK32-NEXT: [[TMP4:%.*]] = bitcast <4 x bfloat> [[VLD3_LANE_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	438 // CHECK32-NEXT: [[TMP5:%.*]] = bitcast <4 x bfloat> [[VLD3_LANE_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	439 // CHECK32-NEXT: [[TMP6:%.*]] = bitcast <4 x bfloat> [[VLD3_LANE_V_FCA_2_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	440 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [3 x <2 x i32>] poison, <2 x i32> [[TMP4]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	441 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [3 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP5]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	442 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [3 x <2 x i32>] [[DOTFCA_1_INSERT]], <2 x i32> [[TMP6]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	443 // CHECK32-NEXT: ret [3 x <2 x i32>] [[DOTFCA_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	444 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	445 bfloat16x4x3_t test_vld3_lane_bf16(bfloat16_t const *ptr, bfloat16x4x3_t src) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	446 return vld3_lane_bf16(ptr, src, 1);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	447 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	448
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	449 // CHECK64-LABEL: @test_vld3q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	450 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	451 // CHECK64-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <8 x bfloat>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	452 // CHECK64-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <8 x bfloat>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	453 // CHECK64-NEXT: [[SRC_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <8 x bfloat>] [[SRC_COERCE]], 2
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	454 // CHECK64-NEXT: [[VLD3_LANE:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld3lane.v8bf16.p0(<8 x bfloat> [[SRC_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[SRC_COERCE_FCA_1_EXTRACT]], <8 x bfloat> [[SRC_COERCE_FCA_2_EXTRACT]], i64 7, ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	455 // CHECK64-NEXT: [[VLD3_LANE_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3_LANE]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	456 // CHECK64-NEXT: [[VLD3_LANE_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3_LANE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	457 // CHECK64-NEXT: [[VLD3_LANE_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3_LANE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	458 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X3_T:%.]] poison, <8 x bfloat> [[VLD3_LANE_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	459 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD3_LANE_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	460 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_1_INSERT]], <8 x bfloat> [[VLD3_LANE_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	461 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	462 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	463 // CHECK32-LABEL: @test_vld3q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	464 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	465 // CHECK32-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <4 x i32>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	466 // CHECK32-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <4 x i32>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	467 // CHECK32-NEXT: [[SRC_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <4 x i32>] [[SRC_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	468 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	469 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	470 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_2_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	471 // CHECK32-NEXT: [[VLD3Q_LANE_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld3lane.v8bf16.p0(ptr [[PTR:%.]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], <8 x bfloat> [[TMP2]], i32 7, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	472 // CHECK32-NEXT: [[VLD3Q_LANE_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_LANE_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	473 // CHECK32-NEXT: [[VLD3Q_LANE_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_LANE_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	474 // CHECK32-NEXT: [[VLD3Q_LANE_V_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_LANE_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	475 // CHECK32-NEXT: [[TMP4:%.*]] = bitcast <8 x bfloat> [[VLD3Q_LANE_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	476 // CHECK32-NEXT: [[TMP5:%.*]] = bitcast <8 x bfloat> [[VLD3Q_LANE_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	477 // CHECK32-NEXT: [[TMP6:%.*]] = bitcast <8 x bfloat> [[VLD3Q_LANE_V_FCA_2_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	478 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [3 x <4 x i32>] poison, <4 x i32> [[TMP4]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	479 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [3 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP5]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	480 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [3 x <4 x i32>] [[DOTFCA_1_INSERT]], <4 x i32> [[TMP6]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	481 // CHECK32-NEXT: ret [3 x <4 x i32>] [[DOTFCA_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	482 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	483 bfloat16x8x3_t test_vld3q_lane_bf16(bfloat16_t const *ptr, bfloat16x8x3_t src) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	484 return vld3q_lane_bf16(ptr, src, 7);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	485 // return vld3q_lane_bf16(ptr, src, 8);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	486 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	487
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	488 // CHECK64-LABEL: @test_vld4_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	489 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	490 // CHECK64-NEXT: [[VLD4:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld4.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	491 // CHECK64-NEXT: [[VLD4_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	492 // CHECK64-NEXT: [[VLD4_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	493 // CHECK64-NEXT: [[VLD4_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	494 // CHECK64-NEXT: [[VLD4_FCA_3_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	495 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X4_T:%.]] poison, <4 x bfloat> [[VLD4_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	496 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD4_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	497 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_1_INSERT]], <4 x bfloat> [[VLD4_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	498 // CHECK64-NEXT: [[DOTFCA_0_3_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_2_INSERT]], <4 x bfloat> [[VLD4_FCA_3_EXTRACT]], 0, 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	499 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	500 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	501 // CHECK32-LABEL: @test_vld4_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	502 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	503 // CHECK32-NEXT: [[VLD4_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld4.v4bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	504 // CHECK32-NEXT: [[VLD4_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	505 // CHECK32-NEXT: [[VLD4_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	506 // CHECK32-NEXT: [[VLD4_V_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	507 // CHECK32-NEXT: [[VLD4_V_FCA_3_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_V]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	508 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD4_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	509 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x bfloat> [[VLD4_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	510 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x bfloat> [[VLD4_V_FCA_2_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	511 // CHECK32-NEXT: [[TMP4:%.*]] = bitcast <4 x bfloat> [[VLD4_V_FCA_3_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	512 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [4 x <2 x i32>] poison, <2 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	513 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	514 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_1_INSERT]], <2 x i32> [[TMP3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	515 // CHECK32-NEXT: [[DOTFCA_3_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_2_INSERT]], <2 x i32> [[TMP4]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	516 // CHECK32-NEXT: ret [4 x <2 x i32>] [[DOTFCA_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	517 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	518 bfloat16x4x4_t test_vld4_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	519 return vld4_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	520 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	521
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	522 // CHECK64-LABEL: @test_vld4q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	523 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	524 // CHECK64-NEXT: [[VLD4:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld4.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	525 // CHECK64-NEXT: [[VLD4_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	526 // CHECK64-NEXT: [[VLD4_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	527 // CHECK64-NEXT: [[VLD4_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	528 // CHECK64-NEXT: [[VLD4_FCA_3_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	529 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X4_T:%.]] poison, <8 x bfloat> [[VLD4_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	530 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD4_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	531 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_1_INSERT]], <8 x bfloat> [[VLD4_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	532 // CHECK64-NEXT: [[DOTFCA_0_3_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_2_INSERT]], <8 x bfloat> [[VLD4_FCA_3_EXTRACT]], 0, 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	533 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	534 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	535 // CHECK32-LABEL: @test_vld4q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	536 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	537 // CHECK32-NEXT: [[VLD4Q_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld4.v8bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	538 // CHECK32-NEXT: [[VLD4Q_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	539 // CHECK32-NEXT: [[VLD4Q_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	540 // CHECK32-NEXT: [[VLD4Q_V_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	541 // CHECK32-NEXT: [[VLD4Q_V_FCA_3_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_V]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	542 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD4Q_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	543 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <8 x bfloat> [[VLD4Q_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	544 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <8 x bfloat> [[VLD4Q_V_FCA_2_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	545 // CHECK32-NEXT: [[TMP4:%.*]] = bitcast <8 x bfloat> [[VLD4Q_V_FCA_3_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	546 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [4 x <4 x i32>] poison, <4 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	547 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	548 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_1_INSERT]], <4 x i32> [[TMP3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	549 // CHECK32-NEXT: [[DOTFCA_3_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_2_INSERT]], <4 x i32> [[TMP4]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	550 // CHECK32-NEXT: ret [4 x <4 x i32>] [[DOTFCA_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	551 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	552 bfloat16x8x4_t test_vld4q_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	553 return vld4q_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	554 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	555
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	556 // CHECK64-LABEL: @test_vld4_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	557 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	558 // CHECK64-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <4 x bfloat>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	559 // CHECK64-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	560 // CHECK64-NEXT: [[SRC_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[SRC_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	561 // CHECK64-NEXT: [[SRC_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[SRC_COERCE]], 3
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	562 // CHECK64-NEXT: [[VLD4_LANE:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld4lane.v4bf16.p0(<4 x bfloat> [[SRC_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[SRC_COERCE_FCA_1_EXTRACT]], <4 x bfloat> [[SRC_COERCE_FCA_2_EXTRACT]], <4 x bfloat> [[SRC_COERCE_FCA_3_EXTRACT]], i64 1, ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	563 // CHECK64-NEXT: [[VLD4_LANE_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_LANE]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	564 // CHECK64-NEXT: [[VLD4_LANE_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_LANE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	565 // CHECK64-NEXT: [[VLD4_LANE_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_LANE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	566 // CHECK64-NEXT: [[VLD4_LANE_FCA_3_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_LANE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	567 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X4_T:%.]] poison, <4 x bfloat> [[VLD4_LANE_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	568 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD4_LANE_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	569 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_1_INSERT]], <4 x bfloat> [[VLD4_LANE_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	570 // CHECK64-NEXT: [[DOTFCA_0_3_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_2_INSERT]], <4 x bfloat> [[VLD4_LANE_FCA_3_EXTRACT]], 0, 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	571 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	572 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	573 // CHECK32-LABEL: @test_vld4_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	574 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	575 // CHECK32-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <2 x i32>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	576 // CHECK32-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	577 // CHECK32-NEXT: [[SRC_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[SRC_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	578 // CHECK32-NEXT: [[SRC_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[SRC_COERCE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	579 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	580 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	581 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_2_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	582 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <2 x i32> [[SRC_COERCE_FCA_3_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	583 // CHECK32-NEXT: [[VLD4_LANE_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld4lane.v4bf16.p0(ptr [[PTR:%.]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], <4 x bfloat> [[TMP2]], <4 x bfloat> [[TMP3]], i32 1, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	584 // CHECK32-NEXT: [[VLD4_LANE_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_LANE_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	585 // CHECK32-NEXT: [[VLD4_LANE_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_LANE_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	586 // CHECK32-NEXT: [[VLD4_LANE_V_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_LANE_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	587 // CHECK32-NEXT: [[VLD4_LANE_V_FCA_3_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_LANE_V]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	588 // CHECK32-NEXT: [[TMP5:%.*]] = bitcast <4 x bfloat> [[VLD4_LANE_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	589 // CHECK32-NEXT: [[TMP6:%.*]] = bitcast <4 x bfloat> [[VLD4_LANE_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	590 // CHECK32-NEXT: [[TMP7:%.*]] = bitcast <4 x bfloat> [[VLD4_LANE_V_FCA_2_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	591 // CHECK32-NEXT: [[TMP8:%.*]] = bitcast <4 x bfloat> [[VLD4_LANE_V_FCA_3_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	592 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [4 x <2 x i32>] poison, <2 x i32> [[TMP5]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	593 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP6]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	594 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_1_INSERT]], <2 x i32> [[TMP7]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	595 // CHECK32-NEXT: [[DOTFCA_3_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_2_INSERT]], <2 x i32> [[TMP8]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	596 // CHECK32-NEXT: ret [4 x <2 x i32>] [[DOTFCA_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	597 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	598 bfloat16x4x4_t test_vld4_lane_bf16(bfloat16_t const *ptr, bfloat16x4x4_t src) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	599 return vld4_lane_bf16(ptr, src, 1);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	600 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	601
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	602 // CHECK64-LABEL: @test_vld4q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	603 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	604 // CHECK64-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <8 x bfloat>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	605 // CHECK64-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	606 // CHECK64-NEXT: [[SRC_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[SRC_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	607 // CHECK64-NEXT: [[SRC_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[SRC_COERCE]], 3
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	608 // CHECK64-NEXT: [[VLD4_LANE:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld4lane.v8bf16.p0(<8 x bfloat> [[SRC_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[SRC_COERCE_FCA_1_EXTRACT]], <8 x bfloat> [[SRC_COERCE_FCA_2_EXTRACT]], <8 x bfloat> [[SRC_COERCE_FCA_3_EXTRACT]], i64 7, ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	609 // CHECK64-NEXT: [[VLD4_LANE_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4_LANE]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	610 // CHECK64-NEXT: [[VLD4_LANE_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4_LANE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	611 // CHECK64-NEXT: [[VLD4_LANE_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4_LANE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	612 // CHECK64-NEXT: [[VLD4_LANE_FCA_3_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4_LANE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	613 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X4_T:%.]] poison, <8 x bfloat> [[VLD4_LANE_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	614 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD4_LANE_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	615 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_1_INSERT]], <8 x bfloat> [[VLD4_LANE_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	616 // CHECK64-NEXT: [[DOTFCA_0_3_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_2_INSERT]], <8 x bfloat> [[VLD4_LANE_FCA_3_EXTRACT]], 0, 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	617 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	618 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	619 // CHECK32-LABEL: @test_vld4q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	620 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	621 // CHECK32-NEXT: [[SRC_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <4 x i32>] [[SRC_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	622 // CHECK32-NEXT: [[SRC_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[SRC_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	623 // CHECK32-NEXT: [[SRC_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[SRC_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	624 // CHECK32-NEXT: [[SRC_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[SRC_COERCE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	625 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	626 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	627 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_2_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	628 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x i32> [[SRC_COERCE_FCA_3_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	629 // CHECK32-NEXT: [[VLD4Q_LANE_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld4lane.v8bf16.p0(ptr [[PTR:%.]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], <8 x bfloat> [[TMP2]], <8 x bfloat> [[TMP3]], i32 7, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	630 // CHECK32-NEXT: [[VLD4Q_LANE_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_LANE_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	631 // CHECK32-NEXT: [[VLD4Q_LANE_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_LANE_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	632 // CHECK32-NEXT: [[VLD4Q_LANE_V_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_LANE_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	633 // CHECK32-NEXT: [[VLD4Q_LANE_V_FCA_3_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_LANE_V]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	634 // CHECK32-NEXT: [[TMP5:%.*]] = bitcast <8 x bfloat> [[VLD4Q_LANE_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	635 // CHECK32-NEXT: [[TMP6:%.*]] = bitcast <8 x bfloat> [[VLD4Q_LANE_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	636 // CHECK32-NEXT: [[TMP7:%.*]] = bitcast <8 x bfloat> [[VLD4Q_LANE_V_FCA_2_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	637 // CHECK32-NEXT: [[TMP8:%.*]] = bitcast <8 x bfloat> [[VLD4Q_LANE_V_FCA_3_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	638 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [4 x <4 x i32>] poison, <4 x i32> [[TMP5]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	639 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP6]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	640 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_1_INSERT]], <4 x i32> [[TMP7]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	641 // CHECK32-NEXT: [[DOTFCA_3_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_2_INSERT]], <4 x i32> [[TMP8]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	642 // CHECK32-NEXT: ret [4 x <4 x i32>] [[DOTFCA_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	643 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	644 bfloat16x8x4_t test_vld4q_lane_bf16(bfloat16_t const *ptr, bfloat16x8x4_t src) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	645 return vld4q_lane_bf16(ptr, src, 7);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	646 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	647
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	648 // CHECK64-LABEL: @test_vld2_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	649 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	650 // CHECK64-NEXT: [[VLD2:%.]] = tail call { <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld2r.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	651 // CHECK64-NEXT: [[VLD2_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	652 // CHECK64-NEXT: [[VLD2_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	653 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X2_T:%.]] poison, <4 x bfloat> [[VLD2_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	654 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X2_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD2_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	655 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X2_T]] [[DOTFCA_0_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	656 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	657 // CHECK32-LABEL: @test_vld2_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	658 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	659 // CHECK32-NEXT: [[VLD2_DUP_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld2dup.v4bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	660 // CHECK32-NEXT: [[VLD2_DUP_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2_DUP_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	661 // CHECK32-NEXT: [[VLD2_DUP_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat> } [[VLD2_DUP_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	662 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD2_DUP_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	663 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x bfloat> [[VLD2_DUP_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	664 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [2 x <2 x i32>] poison, <2 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	665 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [2 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	666 // CHECK32-NEXT: ret [2 x <2 x i32>] [[DOTFCA_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	667 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	668 bfloat16x4x2_t test_vld2_dup_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	669 return vld2_dup_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	670 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	671
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	672 // CHECK64-LABEL: @test_vld2q_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	673 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	674 // CHECK64-NEXT: [[VLD2:%.]] = tail call { <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld2r.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	675 // CHECK64-NEXT: [[VLD2_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	676 // CHECK64-NEXT: [[VLD2_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	677 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X2_T:%.]] poison, <8 x bfloat> [[VLD2_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	678 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X2_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD2_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	679 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X2_T]] [[DOTFCA_0_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	680 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	681 // CHECK32-LABEL: @test_vld2q_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	682 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	683 // CHECK32-NEXT: [[VLD2Q_DUP_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld2dup.v8bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	684 // CHECK32-NEXT: [[VLD2Q_DUP_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2Q_DUP_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	685 // CHECK32-NEXT: [[VLD2Q_DUP_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat> } [[VLD2Q_DUP_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	686 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD2Q_DUP_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	687 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <8 x bfloat> [[VLD2Q_DUP_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	688 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [2 x <4 x i32>] poison, <4 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	689 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [2 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	690 // CHECK32-NEXT: ret [2 x <4 x i32>] [[DOTFCA_1_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	691 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	692 bfloat16x8x2_t test_vld2q_dup_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	693 return vld2q_dup_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	694 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	695
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	696 // CHECK64-LABEL: @test_vld3_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	697 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	698 // CHECK64-NEXT: [[VLD3:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld3r.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	699 // CHECK64-NEXT: [[VLD3_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	700 // CHECK64-NEXT: [[VLD3_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	701 // CHECK64-NEXT: [[VLD3_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	702 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X3_T:%.]] poison, <4 x bfloat> [[VLD3_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	703 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD3_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	704 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_1_INSERT]], <4 x bfloat> [[VLD3_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	705 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X3_T]] [[DOTFCA_0_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	706 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	707 // CHECK32-LABEL: @test_vld3_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	708 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	709 // CHECK32-NEXT: [[VLD3_DUP_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld3dup.v4bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	710 // CHECK32-NEXT: [[VLD3_DUP_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_DUP_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	711 // CHECK32-NEXT: [[VLD3_DUP_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_DUP_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	712 // CHECK32-NEXT: [[VLD3_DUP_V_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD3_DUP_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	713 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD3_DUP_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	714 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x bfloat> [[VLD3_DUP_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	715 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x bfloat> [[VLD3_DUP_V_FCA_2_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	716 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [3 x <2 x i32>] poison, <2 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	717 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [3 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	718 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [3 x <2 x i32>] [[DOTFCA_1_INSERT]], <2 x i32> [[TMP3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	719 // CHECK32-NEXT: ret [3 x <2 x i32>] [[DOTFCA_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	720 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	721 bfloat16x4x3_t test_vld3_dup_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	722 return vld3_dup_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	723 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	724
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	725 // CHECK64-LABEL: @test_vld3q_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	726 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	727 // CHECK64-NEXT: [[VLD3:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld3r.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	728 // CHECK64-NEXT: [[VLD3_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	729 // CHECK64-NEXT: [[VLD3_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	730 // CHECK64-NEXT: [[VLD3_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	731 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X3_T:%.]] poison, <8 x bfloat> [[VLD3_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	732 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD3_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	733 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_1_INSERT]], <8 x bfloat> [[VLD3_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	734 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X3_T]] [[DOTFCA_0_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	735 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	736 // CHECK32-LABEL: @test_vld3q_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	737 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	738 // CHECK32-NEXT: [[VLD3Q_DUP_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld3dup.v8bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	739 // CHECK32-NEXT: [[VLD3Q_DUP_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_DUP_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	740 // CHECK32-NEXT: [[VLD3Q_DUP_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_DUP_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	741 // CHECK32-NEXT: [[VLD3Q_DUP_V_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD3Q_DUP_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	742 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD3Q_DUP_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	743 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <8 x bfloat> [[VLD3Q_DUP_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	744 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <8 x bfloat> [[VLD3Q_DUP_V_FCA_2_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	745 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [3 x <4 x i32>] poison, <4 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	746 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [3 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	747 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [3 x <4 x i32>] [[DOTFCA_1_INSERT]], <4 x i32> [[TMP3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	748 // CHECK32-NEXT: ret [3 x <4 x i32>] [[DOTFCA_2_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	749 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	750 bfloat16x8x3_t test_vld3q_dup_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	751 return vld3q_dup_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	752 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	753
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	754 // CHECK64-LABEL: @test_vld4_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	755 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	756 // CHECK64-NEXT: [[VLD4:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.aarch64.neon.ld4r.v4bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	757 // CHECK64-NEXT: [[VLD4_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	758 // CHECK64-NEXT: [[VLD4_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	759 // CHECK64-NEXT: [[VLD4_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	760 // CHECK64-NEXT: [[VLD4_FCA_3_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	761 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X4X4_T:%.]] poison, <4 x bfloat> [[VLD4_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	762 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_0_INSERT]], <4 x bfloat> [[VLD4_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	763 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_1_INSERT]], <4 x bfloat> [[VLD4_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	764 // CHECK64-NEXT: [[DOTFCA_0_3_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_2_INSERT]], <4 x bfloat> [[VLD4_FCA_3_EXTRACT]], 0, 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	765 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X4X4_T]] [[DOTFCA_0_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	766 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	767 // CHECK32-LABEL: @test_vld4_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	768 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	769 // CHECK32-NEXT: [[VLD4_DUP_V:%.]] = tail call { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } @llvm.arm.neon.vld4dup.v4bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	770 // CHECK32-NEXT: [[VLD4_DUP_V_FCA_0_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_DUP_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	771 // CHECK32-NEXT: [[VLD4_DUP_V_FCA_1_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_DUP_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	772 // CHECK32-NEXT: [[VLD4_DUP_V_FCA_2_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_DUP_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	773 // CHECK32-NEXT: [[VLD4_DUP_V_FCA_3_EXTRACT:%.*]] = extractvalue { <4 x bfloat>, <4 x bfloat>, <4 x bfloat>, <4 x bfloat> } [[VLD4_DUP_V]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	774 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x bfloat> [[VLD4_DUP_V_FCA_0_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	775 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x bfloat> [[VLD4_DUP_V_FCA_1_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	776 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x bfloat> [[VLD4_DUP_V_FCA_2_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	777 // CHECK32-NEXT: [[TMP4:%.*]] = bitcast <4 x bfloat> [[VLD4_DUP_V_FCA_3_EXTRACT]] to <2 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	778 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [4 x <2 x i32>] poison, <2 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	779 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_0_INSERT]], <2 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	780 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_1_INSERT]], <2 x i32> [[TMP3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	781 // CHECK32-NEXT: [[DOTFCA_3_INSERT:%.*]] = insertvalue [4 x <2 x i32>] [[DOTFCA_2_INSERT]], <2 x i32> [[TMP4]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	782 // CHECK32-NEXT: ret [4 x <2 x i32>] [[DOTFCA_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	783 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	784 bfloat16x4x4_t test_vld4_dup_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	785 return vld4_dup_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	786 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	787
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	788 // CHECK64-LABEL: @test_vld4q_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	789 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	790 // CHECK64-NEXT: [[VLD4:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.aarch64.neon.ld4r.v8bf16.p0(ptr [[PTR:%.]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	791 // CHECK64-NEXT: [[VLD4_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	792 // CHECK64-NEXT: [[VLD4_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	793 // CHECK64-NEXT: [[VLD4_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	794 // CHECK64-NEXT: [[VLD4_FCA_3_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	795 // CHECK64-NEXT: [[DOTFCA_0_0_INSERT:%.]] = insertvalue [[STRUCT_BFLOAT16X8X4_T:%.]] poison, <8 x bfloat> [[VLD4_FCA_0_EXTRACT]], 0, 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	796 // CHECK64-NEXT: [[DOTFCA_0_1_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_0_INSERT]], <8 x bfloat> [[VLD4_FCA_1_EXTRACT]], 0, 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	797 // CHECK64-NEXT: [[DOTFCA_0_2_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_1_INSERT]], <8 x bfloat> [[VLD4_FCA_2_EXTRACT]], 0, 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	798 // CHECK64-NEXT: [[DOTFCA_0_3_INSERT:%.*]] = insertvalue [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_2_INSERT]], <8 x bfloat> [[VLD4_FCA_3_EXTRACT]], 0, 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	799 // CHECK64-NEXT: ret [[STRUCT_BFLOAT16X8X4_T]] [[DOTFCA_0_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	800 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	801 // CHECK32-LABEL: @test_vld4q_dup_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	802 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	803 // CHECK32-NEXT: [[VLD4Q_DUP_V:%.]] = tail call { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } @llvm.arm.neon.vld4dup.v8bf16.p0(ptr [[PTR:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	804 // CHECK32-NEXT: [[VLD4Q_DUP_V_FCA_0_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_DUP_V]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	805 // CHECK32-NEXT: [[VLD4Q_DUP_V_FCA_1_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_DUP_V]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	806 // CHECK32-NEXT: [[VLD4Q_DUP_V_FCA_2_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_DUP_V]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	807 // CHECK32-NEXT: [[VLD4Q_DUP_V_FCA_3_EXTRACT:%.*]] = extractvalue { <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat> } [[VLD4Q_DUP_V]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	808 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <8 x bfloat> [[VLD4Q_DUP_V_FCA_0_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	809 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <8 x bfloat> [[VLD4Q_DUP_V_FCA_1_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	810 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <8 x bfloat> [[VLD4Q_DUP_V_FCA_2_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	811 // CHECK32-NEXT: [[TMP4:%.*]] = bitcast <8 x bfloat> [[VLD4Q_DUP_V_FCA_3_EXTRACT]] to <4 x i32>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	812 // CHECK32-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [4 x <4 x i32>] poison, <4 x i32> [[TMP1]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	813 // CHECK32-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_0_INSERT]], <4 x i32> [[TMP2]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	814 // CHECK32-NEXT: [[DOTFCA_2_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_1_INSERT]], <4 x i32> [[TMP3]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	815 // CHECK32-NEXT: [[DOTFCA_3_INSERT:%.*]] = insertvalue [4 x <4 x i32>] [[DOTFCA_2_INSERT]], <4 x i32> [[TMP4]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	816 // CHECK32-NEXT: ret [4 x <4 x i32>] [[DOTFCA_3_INSERT]]
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	817 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	818 bfloat16x8x4_t test_vld4q_dup_bf16(bfloat16_t const *ptr) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	819 return vld4q_dup_bf16(ptr);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	820 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	821
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	822 // CHECK64-LABEL: @test_vst1_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	823 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	824 // CHECK64-NEXT: store <4 x bfloat> [[VAL:%.]], ptr [[PTR:%.]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	825 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	826 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	827 // CHECK32-LABEL: @test_vst1_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	828 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	829 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst1.p0.v4bf16(ptr [[PTR:%.]], <4 x bfloat> [[VAL:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	830 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	831 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	832 void test_vst1_bf16(bfloat16_t *ptr, bfloat16x4_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	833 vst1_bf16(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	834 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	835
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	836 // CHECK64-LABEL: @test_vst1q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	837 // CHECK64-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	838 // CHECK64-NEXT: store <8 x bfloat> [[VAL:%.]], ptr [[PTR:%.]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	839 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	840 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	841 // CHECK32-LABEL: @test_vst1q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	842 // CHECK32-NEXT: entry:
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	843 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst1.p0.v8bf16(ptr [[PTR:%.]], <8 x bfloat> [[VAL:%.]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	844 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	845 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	846 void test_vst1q_bf16(bfloat16_t *ptr, bfloat16x8_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	847 vst1q_bf16(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	848 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	849
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	850 // CHECK-LABEL: @test_vst1_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	851 // CHECK-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	852 // CHECK-NEXT: [[TMP0:%.]] = extractelement <4 x bfloat> [[VAL:%.]], i64 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	853 // CHECK-NEXT: store bfloat [[TMP0]], ptr [[PTR:%.*]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	854 // CHECK-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	855 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	856 void test_vst1_lane_bf16(bfloat16_t *ptr, bfloat16x4_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	857 vst1_lane_bf16(ptr, val, 1);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	858 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	859
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	860 // CHECK-LABEL: @test_vst1q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	861 // CHECK-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	862 // CHECK-NEXT: [[TMP0:%.]] = extractelement <8 x bfloat> [[VAL:%.]], i64 7
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	863 // CHECK-NEXT: store bfloat [[TMP0]], ptr [[PTR:%.*]], align 2
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	864 // CHECK-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	865 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	866 void test_vst1q_lane_bf16(bfloat16_t *ptr, bfloat16x8_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	867 vst1q_lane_bf16(ptr, val, 7);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	868 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	869
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	870 // CHECK64-LABEL: @test_vst1_bf16_x2(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	871 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	872 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	873 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <4 x bfloat>] [[VAL_COERCE]], 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	874 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st1x2.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	875 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	876 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	877 // CHECK32-LABEL: @test_vst1_bf16_x2(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	878 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	879 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	880 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	881 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	882 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	883 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst1x2.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	884 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	885 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	886 void test_vst1_bf16_x2(bfloat16_t *ptr, bfloat16x4x2_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	887 vst1_bf16_x2(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	888 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	889
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	890 // CHECK64-LABEL: @test_vst1q_bf16_x2(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	891 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	892 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	893 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x bfloat>] [[VAL_COERCE]], 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	894 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st1x2.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	895 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	896 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	897 // CHECK32-LABEL: @test_vst1q_bf16_x2(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	898 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	899 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	900 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	901 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	902 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	903 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst1x2.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	904 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	905 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	906 void test_vst1q_bf16_x2(bfloat16_t *ptr, bfloat16x8x2_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	907 vst1q_bf16_x2(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	908 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	909
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	910 // CHECK64-LABEL: @test_vst1_bf16_x3(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	911 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	912 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	913 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	914 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE]], 2
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	915 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st1x3.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	916 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	917 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	918 // CHECK32-LABEL: @test_vst1_bf16_x3(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	919 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	920 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	921 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	922 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	923 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	924 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	925 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	926 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst1x3.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], <4 x bfloat> [[TMP2]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	927 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	928 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	929 void test_vst1_bf16_x3(bfloat16_t *ptr, bfloat16x4x3_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	930 vst1_bf16_x3(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	931 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	932
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	933 // CHECK64-LABEL: @test_vst1q_bf16_x3(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	934 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	935 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	936 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	937 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE]], 2
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	938 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st1x3.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	939 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	940 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	941 // CHECK32-LABEL: @test_vst1q_bf16_x3(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	942 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	943 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	944 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	945 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	946 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	947 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	948 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	949 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst1x3.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], <8 x bfloat> [[TMP2]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	950 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	951 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	952 void test_vst1q_bf16_x3(bfloat16_t *ptr, bfloat16x8x3_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	953 vst1q_bf16_x3(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	954 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	955
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	956 // CHECK64-LABEL: @test_vst1_bf16_x4(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	957 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	958 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	959 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	960 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	961 // CHECK64-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 3
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	962 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st1x4.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_3_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	963 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	964 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	965 // CHECK32-LABEL: @test_vst1_bf16_x4(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	966 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	967 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	968 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	969 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	970 // CHECK32-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	971 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	972 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	973 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	974 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_3_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	975 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst1x4.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], <4 x bfloat> [[TMP2]], <4 x bfloat> [[TMP3]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	976 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	977 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	978 void test_vst1_bf16_x4(bfloat16_t *ptr, bfloat16x4x4_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	979 vst1_bf16_x4(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	980 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	981
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	982 // CHECK64-LABEL: @test_vst1q_bf16_x4(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	983 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	984 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	985 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	986 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	987 // CHECK64-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 3
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	988 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st1x4.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_3_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	989 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	990 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	991 // CHECK32-LABEL: @test_vst1q_bf16_x4(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	992 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	993 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	994 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	995 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	996 // CHECK32-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	997 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	998 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	999 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1000 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_3_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1001 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst1x4.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], <8 x bfloat> [[TMP2]], <8 x bfloat> [[TMP3]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1002 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1003 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1004 void test_vst1q_bf16_x4(bfloat16_t *ptr, bfloat16x8x4_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1005 vst1q_bf16_x4(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1006 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1007
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1008 // CHECK64-LABEL: @test_vst2_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1009 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1010 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1011 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <4 x bfloat>] [[VAL_COERCE]], 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1012 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st2.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1013 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1014 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1015 // CHECK32-LABEL: @test_vst2_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1016 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1017 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1018 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1019 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1020 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1021 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst2.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1022 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1023 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1024 void test_vst2_bf16(bfloat16_t *ptr, bfloat16x4x2_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1025 vst2_bf16(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1026 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1027
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1028 // CHECK64-LABEL: @test_vst2q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1029 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1030 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1031 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x bfloat>] [[VAL_COERCE]], 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1032 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st2.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1033 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1034 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1035 // CHECK32-LABEL: @test_vst2q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1036 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1037 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1038 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1039 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1040 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1041 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst2.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1042 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1043 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1044 void test_vst2q_bf16(bfloat16_t *ptr, bfloat16x8x2_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1045 vst2q_bf16(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1046 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1047
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1048 // CHECK64-LABEL: @test_vst2_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1049 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1050 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1051 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <4 x bfloat>] [[VAL_COERCE]], 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1052 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st2lane.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], i64 1, ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1053 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1054 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1055 // CHECK32-LABEL: @test_vst2_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1056 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1057 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1058 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1059 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1060 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1061 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst2lane.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], i32 1, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1062 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1063 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1064 void test_vst2_lane_bf16(bfloat16_t *ptr, bfloat16x4x2_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1065 vst2_lane_bf16(ptr, val, 1);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1066 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1067
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1068 // CHECK64-LABEL: @test_vst2q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1069 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1070 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1071 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x bfloat>] [[VAL_COERCE]], 1
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1072 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st2lane.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], i64 7, ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1073 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1074 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1075 // CHECK32-LABEL: @test_vst2q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1076 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1077 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [2 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1078 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1079 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1080 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1081 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst2lane.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], i32 7, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1082 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1083 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1084 void test_vst2q_lane_bf16(bfloat16_t *ptr, bfloat16x8x2_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1085 vst2q_lane_bf16(ptr, val, 7);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1086 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1087
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1088 // CHECK64-LABEL: @test_vst3_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1089 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1090 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1091 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1092 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE]], 2
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1093 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st3.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1094 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1095 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1096 // CHECK32-LABEL: @test_vst3_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1097 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1098 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1099 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1100 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1101 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1102 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1103 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1104 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst3.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], <4 x bfloat> [[TMP2]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1105 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1106 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1107 void test_vst3_bf16(bfloat16_t *ptr, bfloat16x4x3_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1108 vst3_bf16(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1109 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1110
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1111 // CHECK64-LABEL: @test_vst3q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1112 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1113 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1114 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1115 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE]], 2
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1116 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st3.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1117 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1118 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1119 // CHECK32-LABEL: @test_vst3q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1120 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1121 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1122 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1123 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1124 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1125 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1126 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1127 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst3.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], <8 x bfloat> [[TMP2]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1128 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1129 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1130 void test_vst3q_bf16(bfloat16_t *ptr, bfloat16x8x3_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1131 vst3q_bf16(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1132 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1133
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1134 // CHECK64-LABEL: @test_vst3_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1135 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1136 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1137 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1138 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <4 x bfloat>] [[VAL_COERCE]], 2
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1139 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st3lane.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], i64 1, ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1140 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1141 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1142 // CHECK32-LABEL: @test_vst3_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1143 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1144 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1145 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1146 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <2 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1147 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1148 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1149 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1150 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst3lane.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], <4 x bfloat> [[TMP2]], i32 1, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1151 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1152 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1153 void test_vst3_lane_bf16(bfloat16_t *ptr, bfloat16x4x3_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1154 vst3_lane_bf16(ptr, val, 1);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1155 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1156
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1157 // CHECK64-LABEL: @test_vst3q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1158 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1159 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1160 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1161 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <8 x bfloat>] [[VAL_COERCE]], 2
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1162 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st3lane.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], i64 7, ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1163 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1164 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1165 // CHECK32-LABEL: @test_vst3q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1166 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1167 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1168 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1169 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [3 x <4 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1170 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1171 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1172 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1173 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst3lane.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], <8 x bfloat> [[TMP2]], i32 7, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1174 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1175 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1176 void test_vst3q_lane_bf16(bfloat16_t *ptr, bfloat16x8x3_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1177 vst3q_lane_bf16(ptr, val, 7);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1178 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1179
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1180 // CHECK64-LABEL: @test_vst4_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1181 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1182 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1183 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1184 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1185 // CHECK64-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 3
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1186 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st4.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_3_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1187 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1188 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1189 // CHECK32-LABEL: @test_vst4_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1190 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1191 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1192 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1193 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1194 // CHECK32-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1195 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1196 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1197 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1198 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_3_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1199 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst4.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], <4 x bfloat> [[TMP2]], <4 x bfloat> [[TMP3]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1200 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1201 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1202 void test_vst4_bf16(bfloat16_t *ptr, bfloat16x4x4_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1203 vst4_bf16(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1204 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1205
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1206 // CHECK64-LABEL: @test_vst4q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1207 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1208 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1209 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1210 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1211 // CHECK64-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 3
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1212 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st4.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_3_EXTRACT]], ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1213 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1214 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1215 // CHECK32-LABEL: @test_vst4q_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1216 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1217 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1218 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1219 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1220 // CHECK32-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1221 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1222 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1223 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1224 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_3_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1225 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst4.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], <8 x bfloat> [[TMP2]], <8 x bfloat> [[TMP3]], i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1226 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1227 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1228 void test_vst4q_bf16(bfloat16_t *ptr, bfloat16x8x4_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1229 vst4q_bf16(ptr, val);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1230 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1231
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1232 // CHECK64-LABEL: @test_vst4_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1233 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1234 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1235 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1236 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1237 // CHECK64-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <4 x bfloat>] [[VAL_COERCE]], 3
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1238 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st4lane.v4bf16.p0(<4 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], <4 x bfloat> [[VAL_COERCE_FCA_3_EXTRACT]], i64 1, ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1239 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1240 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1241 // CHECK32-LABEL: @test_vst4_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1242 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1243 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1244 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1245 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1246 // CHECK32-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <2 x i32>] [[VAL_COERCE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1247 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1248 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1249 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <4 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1250 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <2 x i32> [[VAL_COERCE_FCA_3_EXTRACT]] to <4 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1251 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst4lane.p0.v4bf16(ptr [[PTR:%.*]], <4 x bfloat> [[TMP0]], <4 x bfloat> [[TMP1]], <4 x bfloat> [[TMP2]], <4 x bfloat> [[TMP3]], i32 1, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1252 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1253 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1254 void test_vst4_lane_bf16(bfloat16_t *ptr, bfloat16x4x4_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1255 vst4_lane_bf16(ptr, val, 1);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1256 }
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1257
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1258 // CHECK64-LABEL: @test_vst4q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1259 // CHECK64-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1260 // CHECK64-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1261 // CHECK64-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1262 // CHECK64-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1263 // CHECK64-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <8 x bfloat>] [[VAL_COERCE]], 3
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1264 // CHECK64-NEXT: tail call void @llvm.aarch64.neon.st4lane.v8bf16.p0(<8 x bfloat> [[VAL_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_1_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_2_EXTRACT]], <8 x bfloat> [[VAL_COERCE_FCA_3_EXTRACT]], i64 7, ptr [[PTR:%.*]])
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1265 // CHECK64-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1266 //
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1267 // CHECK32-LABEL: @test_vst4q_lane_bf16(
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1268 // CHECK32-NEXT: entry:
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1269 // CHECK32-NEXT: [[VAL_COERCE_FCA_0_EXTRACT:%.]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE:%.]], 0
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1270 // CHECK32-NEXT: [[VAL_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 1
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1271 // CHECK32-NEXT: [[VAL_COERCE_FCA_2_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 2
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1272 // CHECK32-NEXT: [[VAL_COERCE_FCA_3_EXTRACT:%.*]] = extractvalue [4 x <4 x i32>] [[VAL_COERCE]], 3
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1273 // CHECK32-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_0_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1274 // CHECK32-NEXT: [[TMP1:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_1_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1275 // CHECK32-NEXT: [[TMP2:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_2_EXTRACT]] to <8 x bfloat>
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1276 // CHECK32-NEXT: [[TMP3:%.*]] = bitcast <4 x i32> [[VAL_COERCE_FCA_3_EXTRACT]] to <8 x bfloat>
252 1f2b6ac9f198 LLVM16-1 Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: 236 diff changeset	1277 // CHECK32-NEXT: tail call void @llvm.arm.neon.vst4lane.p0.v8bf16(ptr [[PTR:%.*]], <8 x bfloat> [[TMP0]], <8 x bfloat> [[TMP1]], <8 x bfloat> [[TMP2]], <8 x bfloat> [[TMP3]], i32 7, i32 2)
236 c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1278 // CHECK32-NEXT: ret void
c4bab56944e8 LLVM 16 kono parents: 221 diff changeset	1279 //
221 79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1280 void test_vst4q_lane_bf16(bfloat16_t *ptr, bfloat16x8x4_t val) {
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1281 vst4q_lane_bf16(ptr, val, 7);
79ff65ed7e25 LLVM12 Original Shinji KONO <kono@ie.u-ryukyu.ac.jp> parents: diff changeset	1282 }

Mercurial > hg > CbC > CbC_llvm

annotate clang/test/CodeGen/aarch64-bf16-ldst-intrinsics.c @ 266:00f31e85ec16 default tip