diff options
| author | Matt Arsenault <Matthew.Arsenault@amd.com> | 2019-06-14 21:52:26 +0000 |
|---|---|---|
| committer | Matt Arsenault <Matthew.Arsenault@amd.com> | 2019-06-14 21:52:26 +0000 |
| commit | aa41e92e17a34d5cddfc5cf4b42bea9c5b82a22d (patch) | |
| tree | a8e610d7ea94ae9da85fc9d43609ba037ab9f202 /llvm/test/CodeGen/AMDGPU/call-waitcnt.ll | |
| parent | af857b93df36a983a27a82455c0dea39f53da5b1 (diff) | |
| download | bcm5719-llvm-aa41e92e17a34d5cddfc5cf4b42bea9c5b82a22d.tar.gz bcm5719-llvm-aa41e92e17a34d5cddfc5cf4b42bea9c5b82a22d.zip | |
AMDGPU: Avoid most waitcnts before calls
Currently you get extra waits, because waits are inserted for the
register dependencies of the call, and the function prolog waits on
everything.
Currently waits are still inserted on returns. It may make sense to
not do this, and wait in the caller instead.
llvm-svn: 363465
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/call-waitcnt.ll')
| -rw-r--r-- | llvm/test/CodeGen/AMDGPU/call-waitcnt.ll | 6 |
1 files changed, 1 insertions, 5 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/call-waitcnt.ll b/llvm/test/CodeGen/AMDGPU/call-waitcnt.ll index bb3888c0d76..0557c0fb613 100644 --- a/llvm/test/CodeGen/AMDGPU/call-waitcnt.ll +++ b/llvm/test/CodeGen/AMDGPU/call-waitcnt.ll @@ -17,7 +17,6 @@ define amdgpu_kernel void @call_memory_arg_load(i32 addrspace(3)* %ptr, i32) #0 ; GCN-NEXT: s_getpc_b64 s[6:7] ; GCN-NEXT: s_add_u32 s6, s6, func@rel32@lo+4 ; GCN-NEXT: s_addc_u32 s7, s7, func@rel32@hi+4 -; GCN-NEXT: s_waitcnt lgkmcnt(0) ; GCN-NEXT: s_swappc_b64 s[30:31], s[6:7] ; GCN-NEXT: s_endpgm %vgpr = load volatile i32, i32 addrspace(3)* %ptr @@ -67,7 +66,6 @@ define amdgpu_kernel void @call_no_wait_after_call(i32 addrspace(1)* %ptr, i32) ; GCN-NEXT: s_mov_b32 s32, s33 ; GCN-NEXT: v_mov_b32_e32 v32, 0 ; GCN-NEXT: s_swappc_b64 s[30:31], s[6:7] -; GCN-NEXT: s_waitcnt lgkmcnt(0) ; GCN-NEXT: v_mov_b32_e32 v0, s34 ; GCN-NEXT: v_mov_b32_e32 v1, s35 ; GCN-NEXT: global_store_dword v[0:1], v32, off @@ -91,7 +89,6 @@ define amdgpu_kernel void @call_no_wait_after_call_return_val(i32 addrspace(1)* ; GCN-NEXT: s_addc_u32 s7, s7, func.return@rel32@hi+4 ; GCN-NEXT: s_mov_b32 s32, s33 ; GCN-NEXT: s_swappc_b64 s[30:31], s[6:7] -; GCN-NEXT: s_waitcnt lgkmcnt(0) ; GCN-NEXT: v_mov_b32_e32 v1, s34 ; GCN-NEXT: v_mov_b32_e32 v2, s35 ; GCN-NEXT: global_store_dword v[1:2], v0, off @@ -138,7 +135,7 @@ define void @tailcall_got_load(i32 addrspace(1)* %ptr, i32) #0 { ret void } -; Need to wait for the address dependency +; No need to wait for the load. define void @tail_call_memory_arg_load(i32 addrspace(3)* %ptr, i32) #0 { ; GCN-LABEL: tail_call_memory_arg_load: ; GCN: ; %bb.0: @@ -147,7 +144,6 @@ define void @tail_call_memory_arg_load(i32 addrspace(3)* %ptr, i32) #0 { ; GCN-NEXT: s_add_u32 s6, s6, func@rel32@lo+4 ; GCN-NEXT: s_addc_u32 s7, s7, func@rel32@hi+4 ; GCN-NEXT: ds_read_b32 v0, v0 -; GCN-NEXT: s_waitcnt lgkmcnt(0) ; GCN-NEXT: s_setpc_b64 s[6:7] %vgpr = load volatile i32, i32 addrspace(3)* %ptr tail call void @func(i32 %vgpr) |
