[AMDGPU] Cluster export instructions in PostRA Scheduler (#141399)
DAG mutation needs to be applied post-RA to maintain order established during pre-RA scheduler.
This commit is contained in:
parent
b231e5ff50
commit
e6b43bdde3
@ -1129,6 +1129,7 @@ GCNTargetMachine::createPostMachineScheduler(MachineSchedContext *C) const {
|
||||
getOptLevel() >= CodeGenOptLevel::Less) &&
|
||||
EnableVOPD)
|
||||
DAG->addMutation(createVOPDPairingMutation());
|
||||
DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
|
||||
return DAG;
|
||||
}
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
@ -23,12 +23,12 @@ body: |
|
||||
; CHECK-NEXT: }
|
||||
; CHECK-NEXT: $vgpr6 = nofpexcept V_CVT_PKRTZ_F16_F32_e64 0, killed $vgpr6, 0, 1065353216, 0, 0, implicit $mode, implicit $exec
|
||||
; CHECK-NEXT: $vgpr5 = nofpexcept V_CVT_PKRTZ_F16_F32_e64 0, killed $vgpr5, 0, 1065353216, 0, 0, implicit $mode, implicit $exec
|
||||
; CHECK-NEXT: EXP 0, killed $vgpr0, killed $vgpr6, undef $vgpr0, undef $vgpr0, -1, 0, 3, implicit $exec
|
||||
; CHECK-NEXT: EXP 1, killed $vgpr1, killed $vgpr5, undef $vgpr0, undef $vgpr0, -1, 0, 3, implicit $exec
|
||||
; CHECK-NEXT: $vgpr7 = nnan nsz arcp contract afn reassoc nofpexcept V_MUL_F32_e32 killed $sgpr0, $vgpr2, implicit $mode, implicit $exec
|
||||
; CHECK-NEXT: $vgpr2 = nnan nsz arcp contract afn reassoc nofpexcept V_ADD_F32_e64 0, killed $vgpr2, 0, killed $sgpr1, 1, 0, implicit $mode, implicit $exec
|
||||
; CHECK-NEXT: $vgpr3 = nofpexcept V_CVT_PKRTZ_F16_F32_e32 killed $vgpr7, killed $vgpr3, implicit $mode, implicit $exec
|
||||
; CHECK-NEXT: $vgpr2 = nofpexcept V_CVT_PKRTZ_F16_F32_e32 killed $vgpr4, killed $vgpr2, implicit $mode, implicit $exec
|
||||
; CHECK-NEXT: EXP 0, killed $vgpr0, killed $vgpr6, undef $vgpr0, undef $vgpr0, -1, 0, 3, implicit $exec
|
||||
; CHECK-NEXT: EXP 1, killed $vgpr1, killed $vgpr5, undef $vgpr0, undef $vgpr0, -1, 0, 3, implicit $exec
|
||||
; CHECK-NEXT: EXP_DONE 2, killed $vgpr3, killed $vgpr2, undef $vgpr0, undef $vgpr0, -1, 0, 3, implicit $exec
|
||||
BUNDLE implicit-def $sgpr0, implicit-def $sgpr1, implicit $sgpr8_sgpr9_sgpr10_sgpr11 {
|
||||
$sgpr0 = S_BUFFER_LOAD_DWORD_IMM $sgpr8_sgpr9_sgpr10_sgpr11, 20, 0 :: (dereferenceable invariant load (s32))
|
||||
|
Loading…
x
Reference in New Issue
Block a user