1 ; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
2 ; RUN: llc -global-isel -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX908 %s
3 ; RUN: llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
4 ; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
6 define amdgpu_ps void @global_atomic_fadd_v2f16_no_rtn_intrinsic(ptr addrspace(1) %ptr, <2 x half> %data) {
7 ; GFX908-LABEL: name: global_atomic_fadd_v2f16_no_rtn_intrinsic
8 ; GFX908: bb.1 (%ir-block.0):
9 ; GFX908-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2
11 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
12 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
13 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
14 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
15 ; GFX908-NEXT: GLOBAL_ATOMIC_PK_ADD_F16 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
16 ; GFX908-NEXT: S_ENDPGM 0
17 ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_no_rtn_intrinsic
18 ; GFX90A_GFX940: bb.1 (%ir-block.0):
19 ; GFX90A_GFX940-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2
20 ; GFX90A_GFX940-NEXT: {{ $}}
21 ; GFX90A_GFX940-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
22 ; GFX90A_GFX940-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
23 ; GFX90A_GFX940-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
24 ; GFX90A_GFX940-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
25 ; GFX90A_GFX940-NEXT: GLOBAL_ATOMIC_PK_ADD_F16 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
26 ; GFX90A_GFX940-NEXT: S_ENDPGM 0
27 %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1.v2f16(ptr addrspace(1) %ptr, <2 x half> %data)
31 define amdgpu_ps void @global_atomic_fadd_v2f16_saddr_no_rtn_intrinsic(ptr addrspace(1) inreg %ptr, <2 x half> %data) {
32 ; GFX908-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_intrinsic
33 ; GFX908: bb.1 (%ir-block.0):
34 ; GFX908-NEXT: liveins: $sgpr0, $sgpr1, $vgpr0
36 ; GFX908-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
37 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
38 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
39 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
40 ; GFX908-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
41 ; GFX908-NEXT: GLOBAL_ATOMIC_PK_ADD_F16_SADDR [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
42 ; GFX908-NEXT: S_ENDPGM 0
43 ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_intrinsic
44 ; GFX90A_GFX940: bb.1 (%ir-block.0):
45 ; GFX90A_GFX940-NEXT: liveins: $sgpr0, $sgpr1, $vgpr0
46 ; GFX90A_GFX940-NEXT: {{ $}}
47 ; GFX90A_GFX940-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
48 ; GFX90A_GFX940-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
49 ; GFX90A_GFX940-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
50 ; GFX90A_GFX940-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
51 ; GFX90A_GFX940-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
52 ; GFX90A_GFX940-NEXT: GLOBAL_ATOMIC_PK_ADD_F16_SADDR [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
53 ; GFX90A_GFX940-NEXT: S_ENDPGM 0
54 %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1.v2f16(ptr addrspace(1) %ptr, <2 x half> %data)
58 define amdgpu_ps void @global_atomic_fadd_v2f16_no_rtn_flat_intrinsic(ptr addrspace(1) %ptr, <2 x half> %data) {
59 ; GFX908-LABEL: name: global_atomic_fadd_v2f16_no_rtn_flat_intrinsic
60 ; GFX908: bb.1 (%ir-block.0):
61 ; GFX908-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2
63 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
64 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
65 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
66 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
67 ; GFX908-NEXT: GLOBAL_ATOMIC_PK_ADD_F16 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
68 ; GFX908-NEXT: S_ENDPGM 0
69 ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_no_rtn_flat_intrinsic
70 ; GFX90A_GFX940: bb.1 (%ir-block.0):
71 ; GFX90A_GFX940-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2
72 ; GFX90A_GFX940-NEXT: {{ $}}
73 ; GFX90A_GFX940-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
74 ; GFX90A_GFX940-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
75 ; GFX90A_GFX940-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
76 ; GFX90A_GFX940-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
77 ; GFX90A_GFX940-NEXT: GLOBAL_ATOMIC_PK_ADD_F16 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
78 ; GFX90A_GFX940-NEXT: S_ENDPGM 0
79 %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1.v2f16(ptr addrspace(1) %ptr, <2 x half> %data)
83 define amdgpu_ps void @global_atomic_fadd_v2f16_saddr_no_rtn_flat_intrinsic(ptr addrspace(1) inreg %ptr, <2 x half> %data) {
84 ; GFX908-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_flat_intrinsic
85 ; GFX908: bb.1 (%ir-block.0):
86 ; GFX908-NEXT: liveins: $sgpr0, $sgpr1, $vgpr0
88 ; GFX908-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
89 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
90 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
91 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
92 ; GFX908-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
93 ; GFX908-NEXT: GLOBAL_ATOMIC_PK_ADD_F16_SADDR [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
94 ; GFX908-NEXT: S_ENDPGM 0
95 ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_flat_intrinsic
96 ; GFX90A_GFX940: bb.1 (%ir-block.0):
97 ; GFX90A_GFX940-NEXT: liveins: $sgpr0, $sgpr1, $vgpr0
98 ; GFX90A_GFX940-NEXT: {{ $}}
99 ; GFX90A_GFX940-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
100 ; GFX90A_GFX940-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
101 ; GFX90A_GFX940-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
102 ; GFX90A_GFX940-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
103 ; GFX90A_GFX940-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
104 ; GFX90A_GFX940-NEXT: GLOBAL_ATOMIC_PK_ADD_F16_SADDR [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
105 ; GFX90A_GFX940-NEXT: S_ENDPGM 0
106 %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1.v2f16(ptr addrspace(1) %ptr, <2 x half> %data)
110 declare <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1.v2f16(ptr addrspace(1), <2 x half>)
111 declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1.v2f16(ptr addrspace(1), <2 x half>)