1 // NOTE
: Assertions have been autogenerated by utils
/update_cc_test_checks.py
2 // REQUIRES
: amdgpu-registered-target
3 // RUN
: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1200 -target-feature
+wavefrontsize32 -emit-llvm -o - %s | FileCheck %s --check-prefix
=CHECK-GFX1200
5 typedef int v2i __attribute__
((ext_vector_type(2)));
6 typedef short v8s __attribute__
((ext_vector_type(8)));
7 typedef half v8h __attribute__
((ext_vector_type(8)));
8 typedef __bf16 v8y __attribute__
((ext_vector_type(8)));
10 // CHECK-GFX1200-LABEL
: @test_amdgcn_global_load_tr_b64_v2i32
(
11 // CHECK-GFX1200-NEXT
: entry
:
12 // CHECK-GFX1200-NEXT
: [[TMP0
:%.
*]] = tail call
<2 x i32
> @llvm.amdgcn.global.load.tr.b64.v2i32
(ptr addrspace
(1) [[INPTR
:%.
*]])
13 // CHECK-GFX1200-NEXT
: ret
<2 x i32
> [[TMP0]]
15 v2i test_amdgcn_global_load_tr_b64_v2i32(global v2i* inptr)
17 return __builtin_amdgcn_global_load_tr_b64_v2i32(inptr);
20 // CHECK-GFX1200-LABEL: @test_amdgcn_global_load_tr_b128_v8i16(
21 // CHECK-GFX1200-NEXT: entry:
22 // CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x i16> @llvm.amdgcn.global.load.tr.b128.v8i16(ptr addrspace(1) [[INPTR:%.*]])
23 // CHECK-GFX1200-NEXT: ret <8 x i16> [[TMP0]]
25 v8s test_amdgcn_global_load_tr_b128_v8i16
(global v8s
* inptr
)
27 return __builtin_amdgcn_global_load_tr_b128_v8i16
(inptr);
30 // CHECK-GFX1200-LABEL
: @test_amdgcn_global_load_tr_b128_v8f16
(
31 // CHECK-GFX1200-NEXT
: entry
:
32 // CHECK-GFX1200-NEXT
: [[TMP0
:%.
*]] = tail call
<8 x half
> @llvm.amdgcn.global.load.tr.b128.v8f16
(ptr addrspace
(1) [[INPTR
:%.
*]])
33 // CHECK-GFX1200-NEXT
: ret
<8 x half
> [[TMP0]]
35 v8h test_amdgcn_global_load_tr_b128_v8f16(global v8h* inptr)
37 return __builtin_amdgcn_global_load_tr_b128_v8f16(inptr);
40 // CHECK-GFX1200-LABEL: @test_amdgcn_global_load_tr_b128_v8bf16(
41 // CHECK-GFX1200-NEXT: entry:
42 // CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x bfloat> @llvm.amdgcn.global.load.tr.b128.v8bf16(ptr addrspace(1) [[INPTR:%.*]])
43 // CHECK-GFX1200-NEXT: ret <8 x bfloat> [[TMP0]]
45 v8y test_amdgcn_global_load_tr_b128_v8bf16
(global v8y
* inptr
)
47 return __builtin_amdgcn_global_load_tr_b128_v8bf16
(inptr);