mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-11-23 11:13:28 +01:00
c22abcde1d
Use the 64-bit SGPR base with a 0 offset, since it's 1 fewer instruction to materialize the 0 vs. the 64-bit copy.
39 lines
1.5 KiB
LLVM
39 lines
1.5 KiB
LLVM
; RUN: llc -march=amdgcn -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX8 %s
|
|
; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 %s
|
|
; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX10 %s
|
|
|
|
; FIXME: GFX9 should be producing v_mad_u16 instead of v_mad_legacy_u16.
|
|
|
|
; GCN-LABEL: {{^}}mad_u16
|
|
; GCN: {{flat|global}}_load_ushort v[[A:[0-9]+]]
|
|
; GCN: {{flat|global}}_load_ushort v[[B:[0-9]+]]
|
|
; GCN: {{flat|global}}_load_ushort v[[C:[0-9]+]]
|
|
; GFX8: v_mad_u16 v[[R:[0-9]+]], v[[A]], v[[B]], v[[C]]
|
|
; GFX9: v_mad_legacy_u16 v[[R:[0-9]+]], v[[A]], v[[B]], v[[C]]
|
|
; GFX10: v_mad_u16 v[[R:[0-9]+]], v[[A]], v[[B]], v[[C]]
|
|
; GCN: {{flat|global}}_store_short v{{.+}}, v[[R]]
|
|
; GCN: s_endpgm
|
|
define amdgpu_kernel void @mad_u16(
|
|
i16 addrspace(1)* %r,
|
|
i16 addrspace(1)* %a,
|
|
i16 addrspace(1)* %b,
|
|
i16 addrspace(1)* %c) {
|
|
entry:
|
|
%tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
%a.gep = getelementptr inbounds i16, i16 addrspace(1)* %a, i32 %tid
|
|
%b.gep = getelementptr inbounds i16, i16 addrspace(1)* %b, i32 %tid
|
|
%c.gep = getelementptr inbounds i16, i16 addrspace(1)* %c, i32 %tid
|
|
|
|
%a.val = load volatile i16, i16 addrspace(1)* %a.gep
|
|
%b.val = load volatile i16, i16 addrspace(1)* %b.gep
|
|
%c.val = load volatile i16, i16 addrspace(1)* %c.gep
|
|
|
|
%m.val = mul i16 %a.val, %b.val
|
|
%r.val = add i16 %m.val, %c.val
|
|
|
|
store i16 %r.val, i16 addrspace(1)* %r
|
|
ret void
|
|
}
|
|
|
|
declare i32 @llvm.amdgcn.workitem.id.x()
|