146 lines · c
1// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 42// RUN: %clang_cc1 -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme-b16b16 -target-feature +sme-f16f16 -O2 -Werror -Wall -emit-llvm -o - %s | FileCheck %s3// RUN: %clang_cc1 -x c++ -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme-b16b16 -target-feature +sme-f8f16 -O2 -Werror -Wall -emit-llvm -o - %s | FileCheck %s -check-prefix CHECK-CXX4// RUN: %clang_cc1 -DSVE_OVERLOADED_FORMS -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme-b16b16 -target-feature +sme-f8f16 -O2 -Werror -Wall -emit-llvm -o - %s | FileCheck %s5// RUN: %clang_cc1 -DSVE_OVERLOADED_FORMS -x c++ -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme-b16b16 -target-feature +sme-f16f16 -O2 -Werror -Wall -emit-llvm -o - %s | FileCheck %s -check-prefix CHECK-CXX6 7// RUN: %clang_cc1 -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme-b16b16 -target-feature +sme-f16f16 -O2 -S -Werror -Wall -o /dev/null8 9// REQUIRES: aarch64-registered-target10 11#include <arm_sme.h>12 13#ifdef SVE_OVERLOADED_FORMS14#define SVE_ACLE_FUNC(A1,A2_UNUSED,A3) A1##A315#else16#define SVE_ACLE_FUNC(A1,A2,A3) A1##A2##A317#endif18 19// CHECK-LABEL: define dso_local void @test_svadd_za16_vg1x2_f16(20// CHECK-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {21// CHECK-NEXT: entry:22// CHECK-NEXT: tail call void @llvm.aarch64.sme.add.za16.vg1x2.nxv8f16(i32 [[SLICE]], <vscale x 8 x half> [[ZN_COERCE0]], <vscale x 8 x half> [[ZN_COERCE1]])23// CHECK-NEXT: ret void24//25// CHECK-CXX-LABEL: define dso_local void @_Z25test_svadd_za16_vg1x2_f16j13svfloat16x2_t(26// CHECK-CXX-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {27// CHECK-CXX-NEXT: entry:28// CHECK-CXX-NEXT: tail call void @llvm.aarch64.sme.add.za16.vg1x2.nxv8f16(i32 [[SLICE]], <vscale x 8 x half> [[ZN_COERCE0]], <vscale x 8 x half> [[ZN_COERCE1]])29// CHECK-CXX-NEXT: ret void30//31void test_svadd_za16_vg1x2_f16(uint32_t slice, svfloat16x2_t zn) __arm_streaming __arm_inout("za") {32 SVE_ACLE_FUNC(svadd_za16,_f16,_vg1x2)(slice, zn);33}34 35// CHECK-LABEL: define dso_local void @test_svadd_za16_vg1x4_f16(36// CHECK-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]], <vscale x 8 x half> [[ZN_COERCE2:%.*]], <vscale x 8 x half> [[ZN_COERCE3:%.*]]) local_unnamed_addr #[[ATTR0]] {37// CHECK-NEXT: entry:38// CHECK-NEXT: tail call void @llvm.aarch64.sme.add.za16.vg1x4.nxv8f16(i32 [[SLICE]], <vscale x 8 x half> [[ZN_COERCE0]], <vscale x 8 x half> [[ZN_COERCE1]], <vscale x 8 x half> [[ZN_COERCE2]], <vscale x 8 x half> [[ZN_COERCE3]])39// CHECK-NEXT: ret void40//41// CHECK-CXX-LABEL: define dso_local void @_Z25test_svadd_za16_vg1x4_f16j13svfloat16x4_t(42// CHECK-CXX-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]], <vscale x 8 x half> [[ZN_COERCE2:%.*]], <vscale x 8 x half> [[ZN_COERCE3:%.*]]) local_unnamed_addr #[[ATTR0]] {43// CHECK-CXX-NEXT: entry:44// CHECK-CXX-NEXT: tail call void @llvm.aarch64.sme.add.za16.vg1x4.nxv8f16(i32 [[SLICE]], <vscale x 8 x half> [[ZN_COERCE0]], <vscale x 8 x half> [[ZN_COERCE1]], <vscale x 8 x half> [[ZN_COERCE2]], <vscale x 8 x half> [[ZN_COERCE3]])45// CHECK-CXX-NEXT: ret void46//47void test_svadd_za16_vg1x4_f16(uint32_t slice, svfloat16x4_t zn) __arm_streaming __arm_inout("za") {48 SVE_ACLE_FUNC(svadd_za16,_f16,_vg1x4)(slice, zn);49}50 51// CHECK-LABEL: define dso_local void @test_svsub_za16_vg1x2_f16(52// CHECK-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0]] {53// CHECK-NEXT: entry:54// CHECK-NEXT: tail call void @llvm.aarch64.sme.sub.za16.vg1x2.nxv8f16(i32 [[SLICE]], <vscale x 8 x half> [[ZN_COERCE0]], <vscale x 8 x half> [[ZN_COERCE1]])55// CHECK-NEXT: ret void56//57// CHECK-CXX-LABEL: define dso_local void @_Z25test_svsub_za16_vg1x2_f16j13svfloat16x2_t(58// CHECK-CXX-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0]] {59// CHECK-CXX-NEXT: entry:60// CHECK-CXX-NEXT: tail call void @llvm.aarch64.sme.sub.za16.vg1x2.nxv8f16(i32 [[SLICE]], <vscale x 8 x half> [[ZN_COERCE0]], <vscale x 8 x half> [[ZN_COERCE1]])61// CHECK-CXX-NEXT: ret void62//63void test_svsub_za16_vg1x2_f16(uint32_t slice, svfloat16x2_t zn) __arm_streaming __arm_inout("za") {64 SVE_ACLE_FUNC(svsub_za16,_f16,_vg1x2)(slice, zn);65}66 67// CHECK-LABEL: define dso_local void @test_svsub_za16_vg1x4_f16(68// CHECK-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]], <vscale x 8 x half> [[ZN_COERCE2:%.*]], <vscale x 8 x half> [[ZN_COERCE3:%.*]]) local_unnamed_addr #[[ATTR0]] {69// CHECK-NEXT: entry:70// CHECK-NEXT: tail call void @llvm.aarch64.sme.sub.za16.vg1x4.nxv8f16(i32 [[SLICE]], <vscale x 8 x half> [[ZN_COERCE0]], <vscale x 8 x half> [[ZN_COERCE1]], <vscale x 8 x half> [[ZN_COERCE2]], <vscale x 8 x half> [[ZN_COERCE3]])71// CHECK-NEXT: ret void72//73// CHECK-CXX-LABEL: define dso_local void @_Z25test_svsub_za16_vg1x4_f16j13svfloat16x4_t(74// CHECK-CXX-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]], <vscale x 8 x half> [[ZN_COERCE2:%.*]], <vscale x 8 x half> [[ZN_COERCE3:%.*]]) local_unnamed_addr #[[ATTR0]] {75// CHECK-CXX-NEXT: entry:76// CHECK-CXX-NEXT: tail call void @llvm.aarch64.sme.sub.za16.vg1x4.nxv8f16(i32 [[SLICE]], <vscale x 8 x half> [[ZN_COERCE0]], <vscale x 8 x half> [[ZN_COERCE1]], <vscale x 8 x half> [[ZN_COERCE2]], <vscale x 8 x half> [[ZN_COERCE3]])77// CHECK-CXX-NEXT: ret void78//79void test_svsub_za16_vg1x4_f16(uint32_t slice, svfloat16x4_t zn) __arm_streaming __arm_inout("za") {80 SVE_ACLE_FUNC(svsub_za16,_f16,_vg1x4)(slice, zn);81}82 83// CHECK-LABEL: define dso_local void @test_svadd_za16_vg1x2_bf16(84// CHECK-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0]] {85// CHECK-NEXT: entry:86// CHECK-NEXT: tail call void @llvm.aarch64.sme.add.za16.vg1x2.nxv8bf16(i32 [[SLICE]], <vscale x 8 x bfloat> [[ZN_COERCE0]], <vscale x 8 x bfloat> [[ZN_COERCE1]])87// CHECK-NEXT: ret void88//89// CHECK-CXX-LABEL: define dso_local void @_Z26test_svadd_za16_vg1x2_bf16j14svbfloat16x2_t(90// CHECK-CXX-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0]] {91// CHECK-CXX-NEXT: entry:92// CHECK-CXX-NEXT: tail call void @llvm.aarch64.sme.add.za16.vg1x2.nxv8bf16(i32 [[SLICE]], <vscale x 8 x bfloat> [[ZN_COERCE0]], <vscale x 8 x bfloat> [[ZN_COERCE1]])93// CHECK-CXX-NEXT: ret void94//95void test_svadd_za16_vg1x2_bf16(uint32_t slice, svbfloat16x2_t zn) __arm_streaming __arm_inout("za") {96 SVE_ACLE_FUNC(svadd_za16,_bf16,_vg1x2)(slice, zn);97}98 99// CHECK-LABEL: define dso_local void @test_svadd_za16_vg1x4_bf16(100// CHECK-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE2:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE3:%.*]]) local_unnamed_addr #[[ATTR0]] {101// CHECK-NEXT: entry:102// CHECK-NEXT: tail call void @llvm.aarch64.sme.add.za16.vg1x4.nxv8bf16(i32 [[SLICE]], <vscale x 8 x bfloat> [[ZN_COERCE0]], <vscale x 8 x bfloat> [[ZN_COERCE1]], <vscale x 8 x bfloat> [[ZN_COERCE2]], <vscale x 8 x bfloat> [[ZN_COERCE3]])103// CHECK-NEXT: ret void104//105// CHECK-CXX-LABEL: define dso_local void @_Z26test_svadd_za16_vg1x4_bf16j14svbfloat16x4_t(106// CHECK-CXX-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE2:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE3:%.*]]) local_unnamed_addr #[[ATTR0]] {107// CHECK-CXX-NEXT: entry:108// CHECK-CXX-NEXT: tail call void @llvm.aarch64.sme.add.za16.vg1x4.nxv8bf16(i32 [[SLICE]], <vscale x 8 x bfloat> [[ZN_COERCE0]], <vscale x 8 x bfloat> [[ZN_COERCE1]], <vscale x 8 x bfloat> [[ZN_COERCE2]], <vscale x 8 x bfloat> [[ZN_COERCE3]])109// CHECK-CXX-NEXT: ret void110//111void test_svadd_za16_vg1x4_bf16(uint32_t slice, svbfloat16x4_t zn) __arm_streaming __arm_inout("za") {112 SVE_ACLE_FUNC(svadd_za16,_bf16,_vg1x4)(slice, zn);113}114 115// CHECK-LABEL: define dso_local void @test_svsub_za16_vg1x2_bf16(116// CHECK-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0]] {117// CHECK-NEXT: entry:118// CHECK-NEXT: tail call void @llvm.aarch64.sme.sub.za16.vg1x2.nxv8bf16(i32 [[SLICE]], <vscale x 8 x bfloat> [[ZN_COERCE0]], <vscale x 8 x bfloat> [[ZN_COERCE1]])119// CHECK-NEXT: ret void120//121// CHECK-CXX-LABEL: define dso_local void @_Z26test_svsub_za16_vg1x2_bf16j14svbfloat16x2_t(122// CHECK-CXX-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0]] {123// CHECK-CXX-NEXT: entry:124// CHECK-CXX-NEXT: tail call void @llvm.aarch64.sme.sub.za16.vg1x2.nxv8bf16(i32 [[SLICE]], <vscale x 8 x bfloat> [[ZN_COERCE0]], <vscale x 8 x bfloat> [[ZN_COERCE1]])125// CHECK-CXX-NEXT: ret void126//127void test_svsub_za16_vg1x2_bf16(uint32_t slice, svbfloat16x2_t zn) __arm_streaming __arm_inout("za") {128 SVE_ACLE_FUNC(svsub_za16,_bf16,_vg1x2)(slice, zn);129}130 131// CHECK-LABEL: define dso_local void @test_svsub_za16_vg1x4_bf16(132// CHECK-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE2:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE3:%.*]]) local_unnamed_addr #[[ATTR0]] {133// CHECK-NEXT: entry:134// CHECK-NEXT: tail call void @llvm.aarch64.sme.sub.za16.vg1x4.nxv8bf16(i32 [[SLICE]], <vscale x 8 x bfloat> [[ZN_COERCE0]], <vscale x 8 x bfloat> [[ZN_COERCE1]], <vscale x 8 x bfloat> [[ZN_COERCE2]], <vscale x 8 x bfloat> [[ZN_COERCE3]])135// CHECK-NEXT: ret void136//137// CHECK-CXX-LABEL: define dso_local void @_Z26test_svsub_za16_vg1x4_bf16j14svbfloat16x4_t(138// CHECK-CXX-SAME: i32 noundef [[SLICE:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE2:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE3:%.*]]) local_unnamed_addr #[[ATTR0]] {139// CHECK-CXX-NEXT: entry:140// CHECK-CXX-NEXT: tail call void @llvm.aarch64.sme.sub.za16.vg1x4.nxv8bf16(i32 [[SLICE]], <vscale x 8 x bfloat> [[ZN_COERCE0]], <vscale x 8 x bfloat> [[ZN_COERCE1]], <vscale x 8 x bfloat> [[ZN_COERCE2]], <vscale x 8 x bfloat> [[ZN_COERCE3]])141// CHECK-CXX-NEXT: ret void142//143void test_svsub_za16_vg1x4_bf16(uint32_t slice, svbfloat16x4_t zn) __arm_streaming __arm_inout("za") {144 SVE_ACLE_FUNC(svsub_za16,_bf16,_vg1x4)(slice, zn);145}146