1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-26 12:43:36 +01:00
llvm-mirror/test/CodeGen/AMDGPU/llvm.amdgcn.frexp.exp.f16.ll
Matt Arsenault 248de533d0 AMDGPU: Try to eliminate clearing of high bits of 16-bit instructions
These used to consistently be zeroed pre-gfx9, but gfx9 made the
situation complicated since now some still do and some don't. This
also manages to pick up a few cases that the pattern fails to optimize
away.

We handle some cases with instruction patterns, but some get
through. In particular this improves the integer cases.
2021-06-22 13:42:49 -04:00

49 lines
1.6 KiB
LLVM

; RUN: llc -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
declare i16 @llvm.amdgcn.frexp.exp.i16.f16(half %a)
; GCN-LABEL: {{^}}frexp_exp_f16
; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
; VI: v_frexp_exp_i16_f16_e32 v[[R_I16:[0-9]+]], v[[A_F16]]
; GCN: buffer_store_short v[[R_I16]]
define amdgpu_kernel void @frexp_exp_f16(
i16 addrspace(1)* %r,
half addrspace(1)* %a) {
entry:
%a.val = load half, half addrspace(1)* %a
%r.val = call i16 @llvm.amdgcn.frexp.exp.i16.f16(half %a.val)
store i16 %r.val, i16 addrspace(1)* %r
ret void
}
; GCN-LABEL: {{^}}frexp_exp_f16_sext
; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
; VI: v_frexp_exp_i16_f16_e32 v[[R_I16:[0-9]+]], v[[A_F16]]
; VI: v_bfe_i32 v[[R_I32:[0-9]+]], v[[R_I16]], 0, 16{{$}}
; GCN: buffer_store_dword v[[R_I32]]
define amdgpu_kernel void @frexp_exp_f16_sext(
i32 addrspace(1)* %r,
half addrspace(1)* %a) {
entry:
%a.val = load half, half addrspace(1)* %a
%r.val = call i16 @llvm.amdgcn.frexp.exp.i16.f16(half %a.val)
%r.val.sext = sext i16 %r.val to i32
store i32 %r.val.sext, i32 addrspace(1)* %r
ret void
}
; GCN-LABEL: {{^}}frexp_exp_f16_zext
; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
; VI: v_frexp_exp_i16_f16_e32 v[[R_I16:[0-9]+]], v[[A_F16]]
; GCN: buffer_store_dword v[[R_I16]]
define amdgpu_kernel void @frexp_exp_f16_zext(
i32 addrspace(1)* %r,
half addrspace(1)* %a) {
entry:
%a.val = load half, half addrspace(1)* %a
%r.val = call i16 @llvm.amdgcn.frexp.exp.i16.f16(half %a.val)
%r.val.zext = zext i16 %r.val to i32
store i32 %r.val.zext, i32 addrspace(1)* %r
ret void
}