llvm-for-llvmta/test/CodeGen/AMDGPU/scalar-store-cache-flush.mir

188 lines
3.5 KiB
YAML

# RUN: llc -march=amdgcn -mcpu=fiji -run-pass si-insert-waitcnts %s -o - | FileCheck %s
--- |
define amdgpu_kernel void @basic_insert_dcache_wb() {
ret void
}
define amdgpu_kernel void @explicit_flush_after() {
ret void
}
define amdgpu_kernel void @explicit_flush_before() {
ret void
}
define amdgpu_kernel void @no_scalar_store() {
ret void
}
define amdgpu_kernel void @multi_block_store() {
bb0:
br i1 undef, label %bb1, label %bb2
bb1:
ret void
bb2:
ret void
}
define amdgpu_kernel void @one_block_store() {
bb0:
br i1 undef, label %bb1, label %bb2
bb1:
ret void
bb2:
ret void
}
define amdgpu_ps float @si_return() {
ret float undef
}
...
---
# CHECK-LABEL: name: basic_insert_dcache_wb
# CHECK: bb.0:
# CHECK-NEXT: S_STORE_DWORD
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: S_ENDPGM 0
name: basic_insert_dcache_wb
tracksRegLiveness: false
machineFunctionInfo:
isEntryFunction: true
body: |
bb.0:
S_STORE_DWORD_SGPR undef $sgpr2, undef $sgpr0_sgpr1, undef $m0, 0, 0
S_ENDPGM 0
...
---
# Already has an explicitly requested flush after the last store.
# CHECK-LABEL: name: explicit_flush_after
# CHECK: bb.0:
# CHECK-NEXT: S_STORE_DWORD
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: S_ENDPGM 0
name: explicit_flush_after
tracksRegLiveness: false
machineFunctionInfo:
isEntryFunction: true
body: |
bb.0:
S_STORE_DWORD_SGPR undef $sgpr2, undef $sgpr0_sgpr1, undef $m0, 0, 0
S_DCACHE_WB
S_ENDPGM 0
...
---
# Already has an explicitly requested flush before the last store.
# CHECK-LABEL: name: explicit_flush_before
# CHECK: bb.0:
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: S_STORE_DWORD
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: S_ENDPGM 0
name: explicit_flush_before
tracksRegLiveness: false
machineFunctionInfo:
isEntryFunction: true
body: |
bb.0:
S_DCACHE_WB
S_STORE_DWORD_SGPR undef $sgpr2, undef $sgpr0_sgpr1, undef $m0, 0, 0
S_ENDPGM 0
...
---
# CHECK-LABEL: no_scalar_store
# CHECK: bb.0
# CHECK-NEXT: S_ENDPGM 0
name: no_scalar_store
tracksRegLiveness: false
machineFunctionInfo:
isEntryFunction: true
body: |
bb.0:
S_ENDPGM 0
...
# CHECK-LABEL: name: multi_block_store
# CHECK: bb.0:
# CHECK-NEXT: S_STORE_DWORD
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: S_ENDPGM 0
# CHECK: bb.1:
# CHECK-NEXT: S_STORE_DWORD
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: S_ENDPGM 0
name: multi_block_store
tracksRegLiveness: false
machineFunctionInfo:
isEntryFunction: true
body: |
bb.0:
S_STORE_DWORD_SGPR undef $sgpr2, undef $sgpr0_sgpr1, undef $m0, 0, 0
S_ENDPGM 0
bb.1:
S_STORE_DWORD_SGPR undef $sgpr4, undef $sgpr6_sgpr7, undef $m0, 0, 0
S_ENDPGM 0
...
...
# This one should be able to omit the flush in the storeless block but
# this isn't handled now.
# CHECK-LABEL: name: one_block_store
# CHECK: bb.0:
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: S_ENDPGM 0
# CHECK: bb.1:
# CHECK-NEXT: S_STORE_DWORD
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: S_ENDPGM 0
name: one_block_store
tracksRegLiveness: false
machineFunctionInfo:
isEntryFunction: true
body: |
bb.0:
S_ENDPGM 0
bb.1:
S_STORE_DWORD_SGPR undef $sgpr4, undef $sgpr6_sgpr7, undef $m0, 0, 0
S_ENDPGM 0
...
---
# CHECK-LABEL: name: si_return
# CHECK: bb.0:
# CHECK-NEXT: S_STORE_DWORD
# CHECK-NEXT: S_WAITCNT
# CHECK-NEXT: S_DCACHE_WB
# CHECK-NEXT: SI_RETURN
name: si_return
tracksRegLiveness: false
machineFunctionInfo:
isEntryFunction: true
body: |
bb.0:
S_STORE_DWORD_SGPR undef $sgpr2, undef $sgpr0_sgpr1, undef $m0, 0, 0
SI_RETURN_TO_EPILOG undef $vgpr0
...