replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

R600: Match rcp node on pre-SI 

llvm-svn: 213844
This commit is contained in:
Matt Arsenault 2014-07-24 06:59:24 +00:00
parent 0daeb63f03
commit 9acb978105
4 changed files with 20 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
llvm/lib/Target/R600/CaymanInstructions.td
View File

@ -46,6 +46,8 @@ def SIN_cm : SIN_Common<0x8D>;
def COS_cm : COS_Common<0x8E>;
} // End isVector = 1
defm : RsqPat<RECIPSQRT_IEEE_cm, f32>;
def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;

1
llvm/lib/Target/R600/EvergreenInstructions.td
View File

@ -69,6 +69,7 @@ def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
defm : RsqPat<RECIPSQRT_IEEE_eg, f32>;
def SIN_eg : SIN_Common<0x8D>;
def COS_eg : COS_Common<0x8E>;

7
llvm/lib/Target/R600/R600Instructions.td
View File

@ -1068,7 +1068,7 @@ class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
}
class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
inst, "RECIP_IEEE", [(set f32:$dst, (AMDGPUrcp f32:$src0))]
> {
let Itinerary = TransALU;
}
@ -1114,6 +1114,7 @@ def FNEG_R600 : FNEG<R600_Reg32>;
// Helper patterns for complex intrinsics
//===----------------------------------------------------------------------===//
// FIXME: Should be predicated on unsafe fp math.
multiclass DIV_Common <InstR600 recip_ieee> {
def : Pat<
(int_AMDGPU_div f32:$src0, f32:$src1),
@ -1124,6 +1125,8 @@ def : Pat<
(fdiv f32:$src0, f32:$src1),
(MUL_IEEE $src0, (recip_ieee $src1))
>;
def : RcpPat<recip_ieee, f32>;
}
class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
@ -1180,6 +1183,8 @@ let Predicates = [isR600] in {
def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
defm : RsqPat<RECIPSQRT_IEEE_r600, f32>;
def : FROUNDPat <CNDGE_r600>;
def R600_ExportSwz : ExportSwzInst {

40
llvm/test/CodeGen/R600/llvm.AMDGPU.rcp.ll
View File

@ -1,65 +1,47 @@
; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI -check-prefix=FUNC %s
; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI -check-prefix=FUNC %s
; XUN: llc -march=r600 -mcpu=SI -mattr=+fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE-SPDENORM -check-prefix=SI -check-prefix=FUNC %s
; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG-SAFE -check-prefix=FUNC %s
; RUN: llc -march=r600 -mcpu=cayman -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
declare float @llvm.AMDGPU.rcp.f32(float) nounwind readnone
declare double @llvm.AMDGPU.rcp.f64(double) nounwind readnone
declare float @llvm.sqrt.f32(float) nounwind readnone
declare double @llvm.sqrt.f64(double) nounwind readnone
; FUNC-LABEL: @rcp_f32
; SI: V_RCP_F32_e32
; EG: RECIP_IEEE
define void @rcp_f32(float addrspace(1)* %out, float %src) nounwind {
%rcp = call float @llvm.AMDGPU.rcp.f32(float %src) nounwind readnone
store float %rcp, float addrspace(1)* %out, align 4
ret void
}
; FUNC-LABEL: @rcp_f64
; SI: V_RCP_F64_e32
define void @rcp_f64(double addrspace(1)* %out, double %src) nounwind {
%rcp = call double @llvm.AMDGPU.rcp.f64(double %src) nounwind readnone
store double %rcp, double addrspace(1)* %out, align 8
ret void
}
; FIXME: Evergreen only ever does unsafe fp math.
; FUNC-LABEL: @rcp_pat_f32
; SI-SAFE: V_RCP_F32_e32
; XSI-SAFE-SPDENORM-NOT: V_RCP_F32_e32
; EG: RECIP_IEEE
define void @rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind {
%rcp = fdiv float 1.0, %src
store float %rcp, float addrspace(1)* %out, align 4
ret void
}
; FUNC-LABEL: @rcp_pat_f64
; SI: V_RCP_F64_e32
define void @rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind {
%rcp = fdiv double 1.0, %src
store double %rcp, double addrspace(1)* %out, align 8
ret void
}
; FUNC-LABEL: @rsq_rcp_pat_f32
; SI-UNSAFE: V_RSQ_F32_e32
; SI-SAFE: V_SQRT_F32_e32
; SI-SAFE: V_RCP_F32_e32
; EG: RECIPSQRT_IEEE
define void @rsq_rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind {
%sqrt = call float @llvm.sqrt.f32(float %src) nounwind readnone
%rcp = call float @llvm.AMDGPU.rcp.f32(float %sqrt) nounwind readnone
store float %rcp, float addrspace(1)* %out, align 4
ret void
}
; FUNC-LABEL: @rsq_rcp_pat_f64
; SI-UNSAFE: V_RSQ_F64_e32
; SI-SAFE-NOT: V_RSQ_F64_e32
define void @rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind {
%sqrt = call double @llvm.sqrt.f64(double %src) nounwind readnone
%rcp = call double @llvm.AMDGPU.rcp.f64(double %sqrt) nounwind readnone
store double %rcp, double addrspace(1)* %out, align 8
ret void
}