llvm/test/CodeGen/AMDGPU/global-atomic-fadd.f64.ll

; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
; RUN: llc -mtriple=amdgcn -mcpu=gfx90a -stop-after=amdgpu-isel -amdgpu-atomic-optimizer-strategy=None < %s | FileCheck -check-prefixes=GFX90A_GFX942,GFX90A %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx942 -stop-after=amdgpu-isel -amdgpu-atomic-optimizer-strategy=None < %s | FileCheck -check-prefixes=GFX90A_GFX942,GFX942 %s

define amdgpu_ps void @global_atomic_fadd_f64_no_rtn_atomicrmw(ptr addrspace(1) %ptr, double %data) {
  ; GFX90A_GFX942-LABEL: name: global_atomic_fadd_f64_no_rtn_atomicrmw
  ; GFX90A_GFX942: bb.0 (%ir-block.0):
  ; GFX90A_GFX942-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
  ; GFX90A_GFX942-NEXT: {{  $}}
  ; GFX90A_GFX942-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
  ; GFX90A_GFX942-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
  ; GFX90A_GFX942-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
  ; GFX90A_GFX942-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
  ; GFX90A_GFX942-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
  ; GFX90A_GFX942-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
  ; GFX90A_GFX942-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
  ; GFX90A_GFX942-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
  ; GFX90A_GFX942-NEXT:   GLOBAL_ATOMIC_ADD_F64 killed [[COPY4]], killed [[COPY5]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
  ; GFX90A_GFX942-NEXT:   S_ENDPGM 0
  %ret = atomicrmw fadd ptr addrspace(1) %ptr, double %data syncscope("wavefront") monotonic, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_ps double @global_atomic_fadd_f64_rtn_atomicrmw(ptr addrspace(1) %ptr, double %data) {
  ; GFX90A-LABEL: name: global_atomic_fadd_f64_rtn_atomicrmw
  ; GFX90A: bb.0 (%ir-block.0):
  ; GFX90A-NEXT:   successors: %bb.1(0x80000000)
  ; GFX90A-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
  ; GFX90A-NEXT: {{  $}}
  ; GFX90A-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
  ; GFX90A-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
  ; GFX90A-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
  ; GFX90A-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
  ; GFX90A-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
  ; GFX90A-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
  ; GFX90A-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
  ; GFX90A-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
  ; GFX90A-NEXT:   [[COPY6:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
  ; GFX90A-NEXT:   [[GLOBAL_LOAD_DWORDX2_:%[0-9]+]]:vreg_64_align2 = GLOBAL_LOAD_DWORDX2 [[COPY6]], 0, 0, implicit $exec :: (load (s64) from %ir.ptr, addrspace 1)
  ; GFX90A-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
  ; GFX90A-NEXT: {{  $}}
  ; GFX90A-NEXT: bb.1.atomicrmw.start:
  ; GFX90A-NEXT:   successors: %bb.2(0x04000000), %bb.1(0x7c000000)
  ; GFX90A-NEXT: {{  $}}
  ; GFX90A-NEXT:   [[PHI:%[0-9]+]]:sreg_64 = PHI [[S_MOV_B64_]], %bb.0, %4, %bb.1
  ; GFX90A-NEXT:   [[PHI1:%[0-9]+]]:vreg_64_align2 = PHI [[GLOBAL_LOAD_DWORDX2_]], %bb.0, %3, %bb.1
  ; GFX90A-NEXT:   [[V_ADD_F64_e64_:%[0-9]+]]:vreg_64_align2 = nofpexcept V_ADD_F64_e64 0, [[PHI1]], 0, [[COPY4]], 0, 0, implicit $mode, implicit $exec
  ; GFX90A-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_]].sub1
  ; GFX90A-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_]].sub0
  ; GFX90A-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[PHI1]].sub1
  ; GFX90A-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[PHI1]].sub0
  ; GFX90A-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:sgpr_128 = REG_SEQUENCE killed [[COPY8]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, killed [[COPY10]], %subreg.sub2, killed [[COPY9]], %subreg.sub3
  ; GFX90A-NEXT:   [[COPY11:%[0-9]+]]:vreg_128_align2 = COPY [[REG_SEQUENCE2]]
  ; GFX90A-NEXT:   [[GLOBAL_ATOMIC_CMPSWAP_X2_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_CMPSWAP_X2_RTN [[COPY5]], killed [[COPY11]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic monotonic (s64) on %ir.ptr, addrspace 1)
  ; GFX90A-NEXT:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U64_e64 [[GLOBAL_ATOMIC_CMPSWAP_X2_RTN]], [[PHI1]], implicit $exec
  ; GFX90A-NEXT:   [[SI_IF_BREAK:%[0-9]+]]:sreg_64 = SI_IF_BREAK killed [[V_CMP_EQ_U64_e64_]], [[PHI]], implicit-def dead $scc
  ; GFX90A-NEXT:   SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
  ; GFX90A-NEXT:   S_BRANCH %bb.2
  ; GFX90A-NEXT: {{  $}}
  ; GFX90A-NEXT: bb.2.atomicrmw.end:
  ; GFX90A-NEXT:   [[PHI2:%[0-9]+]]:vreg_64_align2 = PHI [[GLOBAL_ATOMIC_CMPSWAP_X2_RTN]], %bb.1
  ; GFX90A-NEXT:   [[PHI3:%[0-9]+]]:sreg_64 = PHI [[SI_IF_BREAK]], %bb.1
  ; GFX90A-NEXT:   SI_END_CF [[PHI3]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
  ; GFX90A-NEXT:   [[COPY12:%[0-9]+]]:vgpr_32 = COPY [[PHI2]].sub0
  ; GFX90A-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY12]], implicit $exec
  ; GFX90A-NEXT:   [[COPY13:%[0-9]+]]:vgpr_32 = COPY [[PHI2]].sub1
  ; GFX90A-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY13]], implicit $exec
  ; GFX90A-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
  ; GFX90A-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
  ; GFX90A-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
  ;
  ; GFX942-LABEL: name: global_atomic_fadd_f64_rtn_atomicrmw
  ; GFX942: bb.0 (%ir-block.0):
  ; GFX942-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
  ; GFX942-NEXT: {{  $}}
  ; GFX942-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
  ; GFX942-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
  ; GFX942-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
  ; GFX942-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
  ; GFX942-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
  ; GFX942-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
  ; GFX942-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
  ; GFX942-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
  ; GFX942-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_RTN killed [[COPY4]], killed [[COPY5]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
  ; GFX942-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub0
  ; GFX942-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY6]], implicit $exec
  ; GFX942-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub1
  ; GFX942-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY7]], implicit $exec
  ; GFX942-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
  ; GFX942-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
  ; GFX942-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
  %ret = atomicrmw fadd ptr addrspace(1) %ptr, double %data syncscope("wavefront") monotonic
  ret double %ret
}

define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_atomicrmw(ptr addrspace(1) inreg %ptr, double %data) {
  ; GFX90A_GFX942-LABEL: name: global_atomic_fadd_f64_saddr_no_rtn_atomicrmw
  ; GFX90A_GFX942: bb.0 (%ir-block.0):
  ; GFX90A_GFX942-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
  ; GFX90A_GFX942-NEXT: {{  $}}
  ; GFX90A_GFX942-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
  ; GFX90A_GFX942-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
  ; GFX90A_GFX942-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1
  ; GFX90A_GFX942-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0
  ; GFX90A_GFX942-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
  ; GFX90A_GFX942-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
  ; GFX90A_GFX942-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
  ; GFX90A_GFX942-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
  ; GFX90A_GFX942-NEXT:   GLOBAL_ATOMIC_ADD_F64_SADDR killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
  ; GFX90A_GFX942-NEXT:   S_ENDPGM 0
  %ret = atomicrmw fadd ptr addrspace(1) %ptr, double %data syncscope("wavefront") monotonic, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_atomicrmw(ptr addrspace(1) inreg %ptr, double %data) {
  ; GFX90A_GFX942-LABEL: name: global_atomic_fadd_f64_saddr_rtn_atomicrmw
  ; GFX90A_GFX942: bb.0 (%ir-block.0):
  ; GFX90A_GFX942-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
  ; GFX90A_GFX942-NEXT: {{  $}}
  ; GFX90A_GFX942-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
  ; GFX90A_GFX942-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
  ; GFX90A_GFX942-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1
  ; GFX90A_GFX942-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0
  ; GFX90A_GFX942-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
  ; GFX90A_GFX942-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
  ; GFX90A_GFX942-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
  ; GFX90A_GFX942-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
  ; GFX90A_GFX942-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_SADDR_RTN killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
  ; GFX90A_GFX942-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub0
  ; GFX90A_GFX942-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY5]], implicit $exec
  ; GFX90A_GFX942-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub1
  ; GFX90A_GFX942-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY6]], implicit $exec
  ; GFX90A_GFX942-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
  ; GFX90A_GFX942-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
  ; GFX90A_GFX942-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
  %ret = atomicrmw fadd ptr addrspace(1) %ptr, double %data syncscope("wavefront") monotonic, !amdgpu.no.fine.grained.memory !0
  ret double %ret
}

!0 = !{}