AMDGPU: Match store d16_hi instructions

llvm-svn: 313712
This commit is contained in:
Matt Arsenault 2017-09-20 03:20:09 +00:00
parent 09613b122e
commit fcc213fab7
7 changed files with 679 additions and 25 deletions

View File

@ -222,23 +222,43 @@ def COND_NULL : PatLeaf <
// Load/Store Pattern Fragments
//===----------------------------------------------------------------------===//
class PrivateMemOp <dag ops, dag frag> : PatFrag <ops, frag, [{
return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUASI.PRIVATE_ADDRESS;
}]>;
class LoadFrag <SDPatternOperator op> : PatFrag<(ops node:$ptr), (op node:$ptr)>;
class PrivateLoad <SDPatternOperator op> : PrivateMemOp <
(ops node:$ptr), (op node:$ptr)
>;
class PrivateStore <SDPatternOperator op> : PrivateMemOp <
class StoreFrag<SDPatternOperator op> : PatFrag <
(ops node:$value, node:$ptr), (op node:$value, node:$ptr)
>;
def load_private : PrivateLoad <load>;
class StoreHi16<SDPatternOperator op> : PatFrag <
(ops node:$value, node:$ptr), (op (srl node:$value, (i32 16)), node:$ptr)
>;
class PrivateAddress : CodePatPred<[{
return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUASI.PRIVATE_ADDRESS;
}]>;
class LocalAddress : CodePatPred<[{
return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUASI.LOCAL_ADDRESS;
}]>;
class GlobalAddress : CodePatPred<[{
return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUASI.GLOBAL_ADDRESS;
}]>;
class FlatLoadAddress : CodePatPred<[{
const auto AS = cast<MemSDNode>(N)->getAddressSpace();
return AS == AMDGPUASI.FLAT_ADDRESS ||
AS == AMDGPUASI.GLOBAL_ADDRESS;
}]>;
def load_private : LoadFrag <load>, PrivateAddress;
def truncstorei8_private : StoreFrag<truncstorei8>, PrivateAddress;
def truncstorei16_private : StoreFrag <truncstorei16>, PrivateAddress;
def store_private : StoreFrag <store>, PrivateAddress;
def store_private_hi16 : StoreHi16 <truncstorei16>, PrivateAddress;
def truncstorei8_private_hi16 : StoreHi16<truncstorei8>, PrivateAddress;
def truncstorei8_private : PrivateStore <truncstorei8>;
def truncstorei16_private : PrivateStore <truncstorei16>;
def store_private : PrivateStore <store>;
class GlobalMemOp <dag ops, dag frag> : PatFrag <ops, frag, [{
return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUASI.GLOBAL_ADDRESS;
@ -315,8 +335,12 @@ def sextloadi8_constant : ConstantLoad <sextloadi8>;
def az_extloadi8_local : LocalLoad <az_extloadi8>;
def sextloadi8_local : LocalLoad <sextloadi8>;
def extloadi8_private : PrivateLoad <az_extloadi8>;
def sextloadi8_private : PrivateLoad <sextloadi8>;
def store_local_hi16 : StoreHi16 <truncstorei16>, LocalAddress;
def truncstorei8_local_hi16 : StoreHi16<truncstorei8>, LocalAddress;
def extloadi8_private : LoadFrag <az_extloadi8>, PrivateAddress;
def sextloadi8_private : LoadFrag <sextloadi8>, PrivateAddress;
def az_extloadi16 : PatFrag<(ops node:$ptr), (az_extload node:$ptr), [{
return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i16;
@ -331,8 +355,8 @@ def sextloadi16_constant : ConstantLoad <sextloadi16>;
def az_extloadi16_local : LocalLoad <az_extloadi16>;
def sextloadi16_local : LocalLoad <sextloadi16>;
def extloadi16_private : PrivateLoad <az_extloadi16>;
def sextloadi16_private : PrivateLoad <sextloadi16>;
def extloadi16_private : LoadFrag <az_extloadi16>, PrivateAddress;
def sextloadi16_private : LoadFrag <sextloadi16>, PrivateAddress;
def az_extloadi32 : PatFrag<(ops node:$ptr), (az_extload node:$ptr), [{
return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i32;
@ -347,9 +371,12 @@ def az_extloadi32_constant : ConstantLoad <az_extloadi32>;
def truncstorei8_global : GlobalStore <truncstorei8>;
def truncstorei16_global : GlobalStore <truncstorei16>;
def local_store : LocalStore <store>;
def truncstorei8_local : LocalStore <truncstorei8>;
def truncstorei16_local : LocalStore <truncstorei16>;
def truncstorei8_global_hi16 : StoreHi16 <truncstorei8>, GlobalAddress;
def truncstorei16_global_hi16 : StoreHi16 <truncstorei16>, GlobalAddress;
def local_store : StoreFrag <store>, LocalAddress;
def truncstorei8_local : StoreFrag <truncstorei8>, LocalAddress;
def truncstorei16_local : StoreFrag <truncstorei16>, LocalAddress;
def local_load : LocalLoad <load>;

View File

@ -1264,6 +1264,16 @@ defm : MUBUFScratchStorePat <BUFFER_STORE_DWORD_OFFEN, BUFFER_STORE_DWORD_OFFSET
defm : MUBUFScratchStorePat <BUFFER_STORE_DWORDX2_OFFEN, BUFFER_STORE_DWORDX2_OFFSET, v2i32, store_private>;
defm : MUBUFScratchStorePat <BUFFER_STORE_DWORDX4_OFFEN, BUFFER_STORE_DWORDX4_OFFSET, v4i32, store_private>;
let Predicates = [HasD16LoadStore] in {
// Hiding the extract high pattern in the PatFrag seems to not
// automatically increase the complexity.
let AddedComplexity = 1 in {
defm : MUBUFScratchStorePat <BUFFER_STORE_SHORT_D16_HI_OFFEN, BUFFER_STORE_SHORT_D16_HI_OFFSET, i32, store_private_hi16>;
defm : MUBUFScratchStorePat <BUFFER_STORE_BYTE_D16_HI_OFFEN, BUFFER_STORE_BYTE_D16_HI_OFFSET, i32, truncstorei8_private_hi16>;
}
}
//===----------------------------------------------------------------------===//
// MTBUF Patterns
//===----------------------------------------------------------------------===//

View File

@ -576,6 +576,11 @@ def : DSWritePat <DS_WRITE_B8, i16, si_truncstore_local_i8>;
def : DSWritePat <DS_WRITE_B16, i16, si_store_local>;
def : DSWritePat <DS_WRITE_B32, i32, si_store_local>;
let Predicates = [HasD16LoadStore] in {
def : DSWritePat <DS_WRITE_B16_D16_HI, i32, store_local_hi16>;
def : DSWritePat <DS_WRITE_B8_D16_HI, i32, truncstorei8_local_hi16>;
}
let AddedComplexity = 100 in {
def : DSWritePat <DS_WRITE_B64, v2i32, si_store_local_align8>;

View File

@ -645,6 +645,10 @@ def flat_store : flat_st <store>;
def flat_truncstorei8 : flat_st <truncstorei8>;
def flat_truncstorei16 : flat_st <truncstorei16>;
def flat_truncstorei8_hi16 : StoreHi16<truncstorei8>, FlatLoadAddress;
def flat_truncstorei16_hi16 : StoreHi16<truncstorei16>, FlatLoadAddress;
// Patterns for global loads with no offset.
class FlatLoadPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : Pat <
(vt (node (FLATOffset i64:$vaddr, i16:$offset, i1:$slc))),
@ -752,6 +756,12 @@ def : FlatAtomicPat <FLAT_ATOMIC_XOR_X2_RTN, atomic_xor_global, i64>;
def : FlatStorePat <FLAT_STORE_BYTE, flat_truncstorei8, i16>;
def : FlatStorePat <FLAT_STORE_SHORT, flat_store, i16>;
let Predicates = [HasD16LoadStore] in {
def : FlatStorePat <FLAT_STORE_SHORT_D16_HI, flat_truncstorei16_hi16, i32>;
def : FlatStorePat <FLAT_STORE_BYTE_D16_HI, flat_truncstorei8_hi16, i32>;
}
} // End Predicates = [HasFlatAddressSpace]
let Predicates = [HasFlatGlobalInsts], AddedComplexity = 10 in {
@ -779,6 +789,13 @@ def : FlatStoreSignedPat <GLOBAL_STORE_DWORD, global_store, i32>;
def : FlatStoreSignedPat <GLOBAL_STORE_DWORDX2, global_store, v2i32>;
def : FlatStoreSignedPat <GLOBAL_STORE_DWORDX4, global_store, v4i32>;
let Predicates = [HasD16LoadStore] in {
def : FlatStoreSignedPat <GLOBAL_STORE_SHORT_D16_HI, truncstorei16_global_hi16, i32>;
def : FlatStoreSignedPat <GLOBAL_STORE_BYTE_D16_HI, truncstorei8_global_hi16, i32>;
}
def : FlatStoreSignedAtomicPat <GLOBAL_STORE_DWORD, global_store_atomic, i32>;
def : FlatStoreSignedAtomicPat <GLOBAL_STORE_DWORDX2, global_store_atomic, i64>;

View File

@ -1,6 +1,6 @@
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=kaveri -mtriple=amdgcn--amdhsa -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=CIVI -check-prefix=GCN %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tonga -mtriple=amdgcn--amdhsa -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=CIVI -check-prefix=GCN -check-prefix=GFX89 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx901 -mtriple=amdgcn--amdhsa -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 -check-prefix=GCN -check-prefix=GFX89 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=kaveri -mtriple=amdgcn--amdhsa -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CI,CIVI %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tonga -mtriple=amdgcn--amdhsa -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI,CIVI,GFX89 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx901 -mtriple=amdgcn--amdhsa -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89,GFX9 %s
; FIXME: Should be able to do scalar op
; GCN-LABEL: {{^}}s_fneg_f16:
@ -154,7 +154,8 @@ define amdgpu_kernel void @v_extract_fneg_fold_v2f16(<2 x half> addrspace(1)* %i
; GCN-LABEL: {{^}}v_extract_fneg_no_fold_v2f16:
; GCN: {{flat|global}}_load_dword [[VAL:v[0-9]+]]
; GCN: v_xor_b32_e32 [[NEG:v[0-9]+]], 0x80008000, [[VAL]]
; GCN: v_lshrrev_b32_e32 [[ELT1:v[0-9]+]], 16, [[NEG]]
; CIVI: v_lshrrev_b32_e32 [[ELT1:v[0-9]+]], 16, [[NEG]]
; GFX9: global_store_short_d16_hi v{{\[[0-9]+:[0-9]+\]}}, [[NEG]], off
define amdgpu_kernel void @v_extract_fneg_no_fold_v2f16(<2 x half> addrspace(1)* %in) #0 {
%val = load <2 x half>, <2 x half> addrspace(1)* %in
%fneg = fsub <2 x half> <half -0.0, half -0.0>, %val

View File

@ -85,11 +85,11 @@ entry:
}
; FUNC-LABEL: {{^}}store_i24:
; GCN: s_lshr_b32 s{{[0-9]+}}, s{{[0-9]+}}, 16
; SIVI: s_lshr_b32 s{{[0-9]+}}, s{{[0-9]+}}, 16
; SIVI-DAG: buffer_store_byte
; SIVI-DAG: buffer_store_short
; GFX9-DAG: global_store_byte
; GFX9-DAG: global_store_byte_d16_hi v{{\[[0-9]:[0-9]+\]}}, v{{[0-9]+}}, off offset:2
; GFX9-DAG: global_store_short
; EG: MEM_RAT MSKOR

View File

@ -0,0 +1,594 @@
; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9 %s
; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,VI %s
; GCN-LABEL: {{^}}store_global_hi_v2i16:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_short_d16_hi v[0:1], v2, off
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_v2i16(i16 addrspace(1)* %out, i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
store i16 %hi, i16 addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_global_hi_v2f16:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_short_d16_hi v[0:1], v2, off
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_v2f16(half addrspace(1)* %out, i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x half>
%hi = extractelement <2 x half> %value, i32 1
store half %hi, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_global_hi_i32_shift:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_short_d16_hi v[0:1], v2, off
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_i32_shift(i16 addrspace(1)* %out, i32 %value) #0 {
entry:
%hi32 = lshr i32 %value, 16
%hi = trunc i32 %hi32 to i16
store i16 %hi, i16 addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_global_hi_v2i16_i8:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_byte_d16_hi v[0:1], v2, off
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_byte v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_v2i16_i8(i8 addrspace(1)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
store i8 %trunc, i8 addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_global_hi_i8_shift:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_byte_d16_hi v[0:1], v2, off
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_byte v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_i8_shift(i8 addrspace(1)* %out, i32 %value) #0 {
entry:
%hi32 = lshr i32 %value, 16
%hi = trunc i32 %hi32 to i8
store i8 %hi, i8 addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_global_hi_v2i16_max_offset:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_short_d16_hi v[0:1], v2, off offset:4094
; VI-DAG: v_add_i32_e32
; VI-DAG: v_addc_u32_e32
; VI-DAG: v_lshrrev_b32_e32 v2, 16, v2
; VI: flat_store_short v[0:1], v2{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_v2i16_max_offset(i16 addrspace(1)* %out, i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%gep = getelementptr inbounds i16, i16 addrspace(1)* %out, i64 2047
store i16 %hi, i16 addrspace(1)* %gep
ret void
}
; GCN-LABEL: {{^}}store_global_hi_v2i16_min_offset:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_short_d16_hi v[0:1], v2, off offset:-4096{{$}}
; VI-DAG: v_add_i32_e32
; VI-DAG: v_addc_u32_e32
; VI-DAG: v_lshrrev_b32_e32 v2, 16, v2
; VI: flat_store_short v[0:1], v{{[0-9]$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_v2i16_min_offset(i16 addrspace(1)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%gep = getelementptr inbounds i16, i16 addrspace(1)* %out, i64 -2048
store i16 %hi, i16 addrspace(1)* %gep
ret void
}
; GCN-LABEL: {{^}}store_global_hi_v2i16_i8_max_offset:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_byte_d16_hi v[0:1], v2, off offset:4095
; VI-DAG: v_add_i32_e32
; VI-DAG: v_addc_u32_e32
; VI-DAG: v_lshrrev_b32_e32 v2, 16, v2
; VI: flat_store_byte v[0:1], v{{[0-9]$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_v2i16_i8_max_offset(i8 addrspace(1)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
%gep = getelementptr inbounds i8, i8 addrspace(1)* %out, i64 4095
store i8 %trunc, i8 addrspace(1)* %gep
ret void
}
; GCN-LABEL: {{^}}store_global_hi_v2i16_i8_min_offset:
; GCN: s_waitcnt
; GFX9-NEXT: global_store_byte_d16_hi v[0:1], v2, off offset:-4095
; VI-DAG: v_add_i32_e32
; VI-DAG: v_addc_u32_e32
; VI-DAG: v_lshrrev_b32_e32 v2, 16, v2
; VI: flat_store_byte v[0:1], v{{[0-9]$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_global_hi_v2i16_i8_min_offset(i8 addrspace(1)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
%gep = getelementptr inbounds i8, i8 addrspace(1)* %out, i64 -4095
store i8 %trunc, i8 addrspace(1)* %gep
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_v2i16:
; GCN: s_waitcnt
; GFX9-NEXT: flat_store_short_d16_hi v[0:1], v2{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_v2i16(i16 addrspace(4)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
store i16 %hi, i16 addrspace(4)* %out
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_v2f16:
; GCN: s_waitcnt
; GFX9-NEXT: flat_store_short_d16_hi v[0:1], v2{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_v2f16(half addrspace(4)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x half>
%hi = extractelement <2 x half> %value, i32 1
store half %hi, half addrspace(4)* %out
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_i32_shift:
; GCN: s_waitcnt
; GFX9-NEXT: flat_store_short_d16_hi v[0:1], v2{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_i32_shift(i16 addrspace(4)* %out, i32 %value) #0 {
entry:
%hi32 = lshr i32 %value, 16
%hi = trunc i32 %hi32 to i16
store i16 %hi, i16 addrspace(4)* %out
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_v2i16_i8:
; GCN: s_waitcnt
; GFX9-NEXT: flat_store_byte_d16_hi v[0:1], v2{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_byte v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_v2i16_i8(i8 addrspace(4)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
store i8 %trunc, i8 addrspace(4)* %out
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_i8_shift:
; GCN: s_waitcnt
; GFX9-NEXT: flat_store_byte_d16_hi v[0:1], v2{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT: flat_store_byte v[0:1], v2
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_i8_shift(i8 addrspace(4)* %out, i32 %value) #0 {
entry:
%hi32 = lshr i32 %value, 16
%hi = trunc i32 %hi32 to i8
store i8 %hi, i8 addrspace(4)* %out
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_v2i16_max_offset:
; GCN: s_waitcnt
; GFX9-NEXT: flat_store_short_d16_hi v[0:1], v2 offset:4094{{$}}
; VI-DAG: v_add_i32_e32
; VI-DAG: v_addc_u32_e32
; VI-DAG: v_lshrrev_b32_e32 v2, 16, v2
; VI: flat_store_short v[0:1], v2{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_v2i16_max_offset(i16 addrspace(4)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%gep = getelementptr inbounds i16, i16 addrspace(4)* %out, i64 2047
store i16 %hi, i16 addrspace(4)* %gep
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_v2i16_neg_offset:
; GCN: s_waitcnt
; GCN: v_add_i32_e32
; GCN: v_addc_u32_e32
; GFX9-NEXT: flat_store_short_d16_hi v[0:1], v2{{$}}
; VI: flat_store_short v[0:1], v2{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_v2i16_neg_offset(i16 addrspace(4)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%gep = getelementptr inbounds i16, i16 addrspace(4)* %out, i64 -1023
store i16 %hi, i16 addrspace(4)* %gep
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_v2i16_i8_max_offset:
; GCN: s_waitcnt
; GFX9-NEXT: flat_store_byte_d16_hi v[0:1], v2 offset:4095{{$}}
; VI-DAG: v_lshrrev_b32_e32 v2, 16, v2
; VI-DAG: v_add_i32_e32
; VI-DAG: v_addc_u32_e32
; VI: flat_store_byte v[0:1], v2{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_v2i16_i8_max_offset(i8 addrspace(4)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
%gep = getelementptr inbounds i8, i8 addrspace(4)* %out, i64 4095
store i8 %trunc, i8 addrspace(4)* %gep
ret void
}
; GCN-LABEL: {{^}}store_flat_hi_v2i16_i8_neg_offset:
; GCN: s_waitcnt
; GCN-DAG: v_add_i32_e32
; GCN-DAG: v_addc_u32_e32
; GFX9-NEXT: flat_store_byte_d16_hi v[0:1], v2{{$}}
; VI-DAG: v_lshrrev_b32_e32 v2, 16, v2
; VI: flat_store_byte v[0:1], v2{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_flat_hi_v2i16_i8_neg_offset(i8 addrspace(4)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
%gep = getelementptr inbounds i8, i8 addrspace(4)* %out, i64 -4095
store i8 %trunc, i8 addrspace(4)* %gep
ret void
}
; GCN-LABEL: {{^}}store_private_hi_v2i16:
; GCN: s_waitcnt
; GFX9-NEXT: buffer_store_short_d16_hi v1, v0, s[0:3], s4 offen{{$}}
; VI: v_lshrrev_b32_e32 v1, 16, v1
; VI: buffer_store_short v1, v0, s[0:3], s4 offen{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_private_hi_v2i16(i16* %out, i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
store i16 %hi, i16* %out
ret void
}
; GCN-LABEL: {{^}}store_private_hi_v2f16:
; GCN: s_waitcnt
; GFX9-NEXT: buffer_store_short_d16_hi v1, v0, s[0:3], s4 offen{{$}}
; VI: v_lshrrev_b32_e32 v1, 16, v1
; VI: buffer_store_short v1, v0, s[0:3], s4 offen{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_private_hi_v2f16(half* %out, i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x half>
%hi = extractelement <2 x half> %value, i32 1
store half %hi, half* %out
ret void
}
; GCN-LABEL: {{^}}store_private_hi_i32_shift:
; GCN: s_waitcnt
; GFX9-NEXT: buffer_store_short_d16_hi v1, v0, s[0:3], s4 offen{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT: buffer_store_short v1, v0, s[0:3], s4 offen{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_private_hi_i32_shift(i16* %out, i32 %value) #0 {
entry:
%hi32 = lshr i32 %value, 16
%hi = trunc i32 %hi32 to i16
store i16 %hi, i16* %out
ret void
}
; GCN-LABEL: {{^}}store_private_hi_v2i16_i8:
; GCN: s_waitcnt
; GFX9-NEXT: buffer_store_byte_d16_hi v1, v0, s[0:3], s4 offen{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT: buffer_store_byte v1, v0, s[0:3], s4 offen{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_private_hi_v2i16_i8(i8* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
store i8 %trunc, i8* %out
ret void
}
; GCN-LABEL: {{^}}store_private_hi_i8_shift:
; GCN: s_waitcnt
; GFX9-NEXT: buffer_store_byte_d16_hi v1, v0, s[0:3], s4 offen{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT: buffer_store_byte v1, v0, s[0:3], s4 offen{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_private_hi_i8_shift(i8* %out, i32 %value) #0 {
entry:
%hi32 = lshr i32 %value, 16
%hi = trunc i32 %hi32 to i8
store i8 %hi, i8* %out
ret void
}
; GCN-LABEL: {{^}}store_private_hi_v2i16_max_offset:
; GCN: s_waitcnt
; GFX9-NEXT: buffer_store_short_d16_hi v1, v0, s[0:3], s4 offen offset:4094{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT: buffer_store_short v1, v0, s[0:3], s4 offen offset:4094{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_private_hi_v2i16_max_offset(i16* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%gep = getelementptr inbounds i16, i16* %out, i64 2047
store i16 %hi, i16* %gep
ret void
}
; GCN-LABEL: {{^}}store_private_hi_v2i16_nooff:
; GCN: s_waitcnt
; GFX9-NEXT: buffer_store_short_d16_hi v0, off, s[0:3], s4{{$}}
; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; VI-NEXT: buffer_store_short v0, off, s[0:3], s4{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_private_hi_v2i16_nooff(i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
store volatile i16 %hi, i16* null
ret void
}
; GCN-LABEL: {{^}}store_private_hi_v2i16_i8_nooff:
; GCN: s_waitcnt
; GFX9-NEXT: buffer_store_byte_d16_hi v0, off, s[0:3], s4{{$}}
; VI: v_lshrrev_b32_e32 v0, 16, v0
; VI: buffer_store_byte v0, off, s[0:3], s4{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_private_hi_v2i16_i8_nooff(i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
store volatile i8 %trunc, i8* null
ret void
}
; GCN-LABEL: {{^}}store_local_hi_v2i16:
; GCN: s_waitcnt
; GFX9-NEXT: ds_write_b16_d16_hi v0, v1{{$}}
; VI: v_lshrrev_b32_e32 v1, 16, v1
; VI: ds_write_b16 v0, v1
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_local_hi_v2i16(i16 addrspace(3)* %out, i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
store i16 %hi, i16 addrspace(3)* %out
ret void
}
; GCN-LABEL: {{^}}store_local_hi_v2f16:
; GCN: s_waitcnt
; GFX9-NEXT: ds_write_b16_d16_hi v0, v1{{$}}
; VI: v_lshrrev_b32_e32 v1, 16, v1
; VI: ds_write_b16 v0, v1
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_local_hi_v2f16(half addrspace(3)* %out, i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x half>
%hi = extractelement <2 x half> %value, i32 1
store half %hi, half addrspace(3)* %out
ret void
}
; GCN-LABEL: {{^}}store_local_hi_i32_shift:
; GCN: s_waitcnt
; GFX9-NEXT: ds_write_b16_d16_hi v0, v1{{$}}
; VI: v_lshrrev_b32_e32 v1, 16, v1
; VI: ds_write_b16 v0, v1
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_local_hi_i32_shift(i16 addrspace(3)* %out, i32 %value) #0 {
entry:
%hi32 = lshr i32 %value, 16
%hi = trunc i32 %hi32 to i16
store i16 %hi, i16 addrspace(3)* %out
ret void
}
; GCN-LABEL: {{^}}store_local_hi_v2i16_i8:
; GCN: s_waitcnt
; GFX9-NEXT: ds_write_b8_d16_hi v0, v1{{$}}
; VI: v_lshrrev_b32_e32 v1, 16, v1
; VI: ds_write_b8 v0, v1
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_local_hi_v2i16_i8(i8 addrspace(3)* %out, i32 %arg) #0 {
entry:
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%trunc = trunc i16 %hi to i8
store i8 %trunc, i8 addrspace(3)* %out
ret void
}
; GCN-LABEL: {{^}}store_local_hi_v2i16_max_offset:
; GCN: s_waitcnt
; GFX9-NEXT: ds_write_b16_d16_hi v0, v1 offset:65534{{$}}
; VI: v_lshrrev_b32_e32 v1, 16, v1
; VI: ds_write_b16 v0, v1 offset:65534{{$}}
; GCN-NEXT: s_waitcnt
; GCN-NEXT: s_setpc_b64
define void @store_local_hi_v2i16_max_offset(i16 addrspace(3)* %out, i32 %arg) #0 {
entry:
; FIXME: ABI for pre-gfx9
%value = bitcast i32 %arg to <2 x i16>
%hi = extractelement <2 x i16> %value, i32 1
%gep = getelementptr inbounds i16, i16 addrspace(3)* %out, i64 32767
store i16 %hi, i16 addrspace(3)* %gep
ret void
}
attributes #0 = { nounwind }