mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-11-26 12:43:36 +01:00
4b0ec23e84
This is to allow 64 bit constant rematerialization. If a constant is split into two separate moves initializing sub0 and sub1 like now RA cannot rematerizalize a 64 bit register. This gives 10-20% uplift in a set of huge apps heavily using double precession math. Fixes: SWDEV-292645 Differential Revision: https://reviews.llvm.org/D104874
104 lines
4.3 KiB
LLVM
104 lines
4.3 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -verify-machineinstrs < %s | FileCheck %s -check-prefix=GCN
|
|
; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1031 -verify-machineinstrs < %s | FileCheck %s -check-prefix=GCN
|
|
|
|
define i32 @global_atomic_csub(i32 addrspace(1)* %ptr, i32 %data) {
|
|
; GCN-LABEL: global_atomic_csub:
|
|
; GCN: ; %bb.0:
|
|
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GCN-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GCN-NEXT: global_atomic_csub v0, v[0:1], v2, off glc
|
|
; GCN-NEXT: s_waitcnt vmcnt(0)
|
|
; GCN-NEXT: s_setpc_b64 s[30:31]
|
|
%ret = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* %ptr, i32 %data)
|
|
ret i32 %ret
|
|
}
|
|
|
|
define i32 @global_atomic_csub_offset(i32 addrspace(1)* %ptr, i32 %data) {
|
|
; GCN-LABEL: global_atomic_csub_offset:
|
|
; GCN: ; %bb.0:
|
|
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GCN-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GCN-NEXT: s_mov_b64 s[4:5], 0x1000
|
|
; GCN-NEXT: v_mov_b32_e32 v3, s4
|
|
; GCN-NEXT: v_mov_b32_e32 v4, s5
|
|
; GCN-NEXT: v_add_co_u32 v0, vcc_lo, v0, v3
|
|
; GCN-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, v1, v4, vcc_lo
|
|
; GCN-NEXT: global_atomic_csub v0, v[0:1], v2, off glc
|
|
; GCN-NEXT: s_waitcnt vmcnt(0)
|
|
; GCN-NEXT: s_setpc_b64 s[30:31]
|
|
%gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 1024
|
|
%ret = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* %gep, i32 %data)
|
|
ret i32 %ret
|
|
}
|
|
|
|
define void @global_atomic_csub_nortn(i32 addrspace(1)* %ptr, i32 %data) {
|
|
; GCN-LABEL: global_atomic_csub_nortn:
|
|
; GCN: ; %bb.0:
|
|
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GCN-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GCN-NEXT: global_atomic_csub v0, v[0:1], v2, off glc
|
|
; GCN-NEXT: s_waitcnt vmcnt(0)
|
|
; GCN-NEXT: s_setpc_b64 s[30:31]
|
|
%ret = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* %ptr, i32 %data)
|
|
ret void
|
|
}
|
|
|
|
define void @global_atomic_csub_offset_nortn(i32 addrspace(1)* %ptr, i32 %data) {
|
|
; GCN-LABEL: global_atomic_csub_offset_nortn:
|
|
; GCN: ; %bb.0:
|
|
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GCN-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GCN-NEXT: s_mov_b64 s[4:5], 0x1000
|
|
; GCN-NEXT: v_mov_b32_e32 v3, s4
|
|
; GCN-NEXT: v_mov_b32_e32 v4, s5
|
|
; GCN-NEXT: v_add_co_u32 v0, vcc_lo, v0, v3
|
|
; GCN-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, v1, v4, vcc_lo
|
|
; GCN-NEXT: global_atomic_csub v0, v[0:1], v2, off glc
|
|
; GCN-NEXT: s_waitcnt vmcnt(0)
|
|
; GCN-NEXT: s_setpc_b64 s[30:31]
|
|
%gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 1024
|
|
%ret = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* %gep, i32 %data)
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @global_atomic_csub_sgpr_base_offset(i32 addrspace(1)* %ptr, i32 %data) {
|
|
; GCN-LABEL: global_atomic_csub_sgpr_base_offset:
|
|
; GCN: ; %bb.0:
|
|
; GCN-NEXT: s_clause 0x1
|
|
; GCN-NEXT: s_load_dword s2, s[4:5], 0x8
|
|
; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
|
|
; GCN-NEXT: v_mov_b32_e32 v1, 0x1000
|
|
; GCN-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GCN-NEXT: v_mov_b32_e32 v0, s2
|
|
; GCN-NEXT: global_atomic_csub v0, v1, v0, s[0:1] glc
|
|
; GCN-NEXT: s_waitcnt vmcnt(0)
|
|
; GCN-NEXT: global_store_dword v[0:1], v0, off
|
|
; GCN-NEXT: s_endpgm
|
|
%gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 1024
|
|
%ret = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* %gep, i32 %data)
|
|
store i32 %ret, i32 addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @global_atomic_csub_sgpr_base_offset_nortn(i32 addrspace(1)* %ptr, i32 %data) {
|
|
; GCN-LABEL: global_atomic_csub_sgpr_base_offset_nortn:
|
|
; GCN: ; %bb.0:
|
|
; GCN-NEXT: s_clause 0x1
|
|
; GCN-NEXT: s_load_dword s2, s[4:5], 0x8
|
|
; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
|
|
; GCN-NEXT: v_mov_b32_e32 v1, 0x1000
|
|
; GCN-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GCN-NEXT: v_mov_b32_e32 v0, s2
|
|
; GCN-NEXT: global_atomic_csub v0, v1, v0, s[0:1] glc
|
|
; GCN-NEXT: s_endpgm
|
|
%gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 1024
|
|
%ret = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* %gep, i32 %data)
|
|
ret void
|
|
}
|
|
|
|
declare i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* nocapture, i32) #1
|
|
|
|
attributes #0 = { nounwind willreturn }
|
|
attributes #1 = { argmemonly nounwind }
|