Skip to content

Commit f433c3b

Browse files
arsenmarsenm
authored
AMDGPU: Add tests for atomicrmw handling of new metadata (#89248)
Add baseline tests which should comprehensively test the new atomic metadata. Test codegen / expansion, and preservation in a few transforms. New metadata defined in #85052
1 parent d9169ff commit f433c3b

20 files changed

+22034
-1649
lines changed

‎llvm/test/CodeGen/AMDGPU/flat-atomic-fadd.v2f16.ll‎

Lines changed: 76 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -35,3 +35,79 @@ define amdgpu_ps <2 x half> @flat_atomic_fadd_v2f16_rtn_intrinsic(ptr %ptr, <2 x
3535
}
3636

3737
declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1.v2f16(ptr, <2 x half>)
38+
39+
define <2 x half> @flat_agent_atomic_fadd_ret_v2f16(ptr %ptr, <2 x half> %val) {
40+
; GFX940-LABEL: name: flat_agent_atomic_fadd_ret_v2f16
41+
; GFX940: bb.0 (%ir-block.0):
42+
; GFX940-NEXT: successors: %bb.1(0x80000000)
43+
; GFX940-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2
44+
; GFX940-NEXT: {{ $}}
45+
; GFX940-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
46+
; GFX940-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
47+
; GFX940-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
48+
; GFX940-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
49+
; GFX940-NEXT: [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
50+
; GFX940-NEXT: [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
51+
; GFX940-NEXT: [[FLAT_LOAD_DWORD:%[0-9]+]]:vgpr_32 = FLAT_LOAD_DWORD [[COPY4]], 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.ptr)
52+
; GFX940-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
53+
; GFX940-NEXT: {{ $}}
54+
; GFX940-NEXT: bb.1.atomicrmw.start:
55+
; GFX940-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
56+
; GFX940-NEXT: {{ $}}
57+
; GFX940-NEXT: [[PHI:%[0-9]+]]:sreg_64 = PHI [[S_MOV_B64_]], %bb.0, %4, %bb.1
58+
; GFX940-NEXT: [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[FLAT_LOAD_DWORD]], %bb.0, %3, %bb.1
59+
; GFX940-NEXT: [[V_PK_ADD_F16_:%[0-9]+]]:vgpr_32 = nofpexcept V_PK_ADD_F16 8, [[PHI1]], 8, [[COPY]], 0, 0, 0, 0, 0, implicit $mode, implicit $exec
60+
; GFX940-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE killed [[V_PK_ADD_F16_]], %subreg.sub0, [[PHI1]], %subreg.sub1
61+
; GFX940-NEXT: [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
62+
; GFX940-NEXT: [[FLAT_ATOMIC_CMPSWAP_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_CMPSWAP_RTN [[COPY3]], killed [[COPY5]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("agent") seq_cst seq_cst (s32) on %ir.ptr)
63+
; GFX940-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 [[FLAT_ATOMIC_CMPSWAP_RTN]], [[PHI1]], implicit $exec
64+
; GFX940-NEXT: [[SI_IF_BREAK:%[0-9]+]]:sreg_64 = SI_IF_BREAK killed [[V_CMP_EQ_U32_e64_]], [[PHI]], implicit-def dead $scc
65+
; GFX940-NEXT: SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
66+
; GFX940-NEXT: S_BRANCH %bb.2
67+
; GFX940-NEXT: {{ $}}
68+
; GFX940-NEXT: bb.2.atomicrmw.end:
69+
; GFX940-NEXT: [[PHI2:%[0-9]+]]:vgpr_32 = PHI [[FLAT_ATOMIC_CMPSWAP_RTN]], %bb.1
70+
; GFX940-NEXT: [[PHI3:%[0-9]+]]:sreg_64 = PHI [[SI_IF_BREAK]], %bb.1
71+
; GFX940-NEXT: SI_END_CF [[PHI3]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
72+
; GFX940-NEXT: $vgpr0 = COPY [[PHI2]]
73+
; GFX940-NEXT: SI_RETURN implicit $vgpr0
74+
%result = atomicrmw fadd ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
75+
ret <2 x half> %result
76+
}
77+
78+
define void @flat_agent_atomic_fadd_noret_v2f16(ptr %ptr, <2 x half> %val) {
79+
; GFX940-LABEL: name: flat_agent_atomic_fadd_noret_v2f16
80+
; GFX940: bb.0 (%ir-block.0):
81+
; GFX940-NEXT: successors: %bb.1(0x80000000)
82+
; GFX940-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2
83+
; GFX940-NEXT: {{ $}}
84+
; GFX940-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
85+
; GFX940-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
86+
; GFX940-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
87+
; GFX940-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
88+
; GFX940-NEXT: [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
89+
; GFX940-NEXT: [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
90+
; GFX940-NEXT: [[FLAT_LOAD_DWORD:%[0-9]+]]:vgpr_32 = FLAT_LOAD_DWORD [[COPY4]], 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.ptr)
91+
; GFX940-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
92+
; GFX940-NEXT: {{ $}}
93+
; GFX940-NEXT: bb.1.atomicrmw.start:
94+
; GFX940-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
95+
; GFX940-NEXT: {{ $}}
96+
; GFX940-NEXT: [[PHI:%[0-9]+]]:sreg_64 = PHI [[S_MOV_B64_]], %bb.0, %4, %bb.1
97+
; GFX940-NEXT: [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[FLAT_LOAD_DWORD]], %bb.0, %3, %bb.1
98+
; GFX940-NEXT: [[V_PK_ADD_F16_:%[0-9]+]]:vgpr_32 = nofpexcept V_PK_ADD_F16 8, [[PHI1]], 8, [[COPY]], 0, 0, 0, 0, 0, implicit $mode, implicit $exec
99+
; GFX940-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE killed [[V_PK_ADD_F16_]], %subreg.sub0, [[PHI1]], %subreg.sub1
100+
; GFX940-NEXT: [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
101+
; GFX940-NEXT: [[FLAT_ATOMIC_CMPSWAP_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_CMPSWAP_RTN [[COPY3]], killed [[COPY5]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("agent") seq_cst seq_cst (s32) on %ir.ptr)
102+
; GFX940-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 [[FLAT_ATOMIC_CMPSWAP_RTN]], [[PHI1]], implicit $exec
103+
; GFX940-NEXT: [[SI_IF_BREAK:%[0-9]+]]:sreg_64 = SI_IF_BREAK killed [[V_CMP_EQ_U32_e64_]], [[PHI]], implicit-def dead $scc
104+
; GFX940-NEXT: SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
105+
; GFX940-NEXT: S_BRANCH %bb.2
106+
; GFX940-NEXT: {{ $}}
107+
; GFX940-NEXT: bb.2.atomicrmw.end:
108+
; GFX940-NEXT: [[PHI2:%[0-9]+]]:sreg_64 = PHI [[SI_IF_BREAK]], %bb.1
109+
; GFX940-NEXT: SI_END_CF [[PHI2]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
110+
; GFX940-NEXT: SI_RETURN
111+
%result = atomicrmw fadd ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
112+
ret void
113+
}

0 commit comments

Comments
 (0)