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

AMDGPU/SI: Promote i16 = fp_[us]int f32 for VI 

Summary: This fixes a regression caused by r286464.

Reviewers: arsenm

Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, llvm-commits, tony-tye

Differential Revision: https://reviews.llvm.org/D26570

llvm-svn: 286687
This commit is contained in:
Tom Stellard 2016-11-12 00:19:11 +00:00
parent 2f94a7a1fa
commit b4c8e8e30b
3 changed files with 43 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
llvm/lib/Target/AMDGPU/SIISelLowering.cpp
View File

@ -271,6 +271,12 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
AddPromotedToType(ISD::FP16_TO_FP, MVT::i16, MVT::i32);
setOperationAction(ISD::FP_TO_FP16, MVT::i16, Promote);
AddPromotedToType(ISD::FP_TO_FP16, MVT::i16, MVT::i32);
setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
AddPromotedToType(ISD::FP_TO_SINT, MVT::i16, MVT::i32);
setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
AddPromotedToType(ISD::FP_TO_UINT, MVT::i16, MVT::i32);
}
setTargetDAGCombine(ISD::FADD);

9
llvm/test/CodeGen/AMDGPU/fp_to_sint.ll
View File

@ -248,5 +248,14 @@ define void @fp_to_uint_fabs_f32_to_i1(i1 addrspace(1)* %out, float %in) #0 {
ret void
}
; FUNC-LABEL: {{^}}fp_to_sint_f32_i16:
; SI: v_cvt_i32_f32_e32 [[VAL:v[0-9]+]], s{{[0-9]+}}
; SI: buffer_store_short [[VAL]]
define void @fp_to_sint_f32_i16(i16 addrspace(1)* %out, float %in) #0 {
%sint = fptosi float %in to i16
store i16 %sint, i16 addrspace(1)* %out
ret void
}
attributes #0 = { nounwind }
attributes #1 = { nounwind readnone }

43
llvm/test/CodeGen/AMDGPU/fp_to_uint.ll
View File

@ -1,5 +1,5 @@
; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s -check-prefix=SI -check-prefix=FUNC
; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck %s -check-prefix=SI -check-prefix=FUNC
; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s -check-prefixes=GCN,FUNC,SI
; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck %s -check-prefixes=GCN,FUNC,VI
; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck %s -check-prefix=EG -check-prefix=FUNC
declare float @llvm.fabs.f32(float) #1
@ -7,8 +7,8 @@ declare float @llvm.fabs.f32(float) #1
; FUNC-LABEL: {{^}}fp_to_uint_f32_to_i32:
; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
; SI: v_cvt_u32_f32_e32
; SI: s_endpgm
; GCN: v_cvt_u32_f32_e32
; GCN: s_endpgm
define void @fp_to_uint_f32_to_i32 (i32 addrspace(1)* %out, float %in) {
%conv = fptoui float %in to i32
store i32 %conv, i32 addrspace(1)* %out
@ -19,8 +19,8 @@ define void @fp_to_uint_f32_to_i32 (i32 addrspace(1)* %out, float %in) {
; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
; SI: v_cvt_u32_f32_e32
; SI: v_cvt_u32_f32_e32
; GCN: v_cvt_u32_f32_e32
; GCN: v_cvt_u32_f32_e32
define void @fp_to_uint_v2f32_to_v2i32(<2 x i32> addrspace(1)* %out, <2 x float> %in) {
%result = fptoui <2 x float> %in to <2 x i32>
store <2 x i32> %result, <2 x i32> addrspace(1)* %out
@ -32,10 +32,10 @@ define void @fp_to_uint_v2f32_to_v2i32(<2 x i32> addrspace(1)* %out, <2 x float>
; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
; EG: FLT_TO_UINT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW]}}
; SI: v_cvt_u32_f32_e32
; SI: v_cvt_u32_f32_e32
; SI: v_cvt_u32_f32_e32
; SI: v_cvt_u32_f32_e32
; GCN: v_cvt_u32_f32_e32
; GCN: v_cvt_u32_f32_e32
; GCN: v_cvt_u32_f32_e32
; GCN: v_cvt_u32_f32_e32
define void @fp_to_uint_v4f32_to_v4i32(<4 x i32> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
%value = load <4 x float>, <4 x float> addrspace(1) * %in
@ -67,7 +67,7 @@ define void @fp_to_uint_v4f32_to_v4i32(<4 x i32> addrspace(1)* %out, <4 x float>
; EG-DAG: CNDE_INT
; EG-DAG: CNDE_INT
; SI: s_endpgm
; GCN: s_endpgm
define void @fp_to_uint_f32_to_i64(i64 addrspace(1)* %out, float %x) {
%conv = fptoui float %x to i64
store i64 %conv, i64 addrspace(1)* %out
@ -118,7 +118,7 @@ define void @fp_to_uint_f32_to_i64(i64 addrspace(1)* %out, float %x) {
; EG-DAG: CNDE_INT
; EG-DAG: CNDE_INT
; SI: s_endpgm
; GCN: s_endpgm
define void @fp_to_uint_v2f32_to_v2i64(<2 x i64> addrspace(1)* %out, <2 x float> %x) {
%conv = fptoui <2 x float> %x to <2 x i64>
store <2 x i64> %conv, <2 x i64> addrspace(1)* %out
@ -211,7 +211,7 @@ define void @fp_to_uint_v2f32_to_v2i64(<2 x i64> addrspace(1)* %out, <2 x float>
; EG-DAG: CNDE_INT
; EG-DAG: CNDE_INT
; SI: s_endpgm
; GCN: s_endpgm
define void @fp_to_uint_v4f32_to_v4i64(<4 x i64> addrspace(1)* %out, <4 x float> %x) {
%conv = fptoui <4 x float> %x to <4 x i64>
store <4 x i64> %conv, <4 x i64> addrspace(1)* %out
@ -220,7 +220,7 @@ define void @fp_to_uint_v4f32_to_v4i64(<4 x i64> addrspace(1)* %out, <4 x float>
; FUNC-LABEL: {{^}}fp_to_uint_f32_to_i1:
; SI: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1.0, s{{[0-9]+}}
; GCN: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1.0, s{{[0-9]+}}
; EG: AND_INT
; EG: SETE_DX10 {{[*]?}} T{{[0-9]+}}.{{[XYZW]}}, KC0[2].Z, 1.0,
@ -231,7 +231,7 @@ define void @fp_to_uint_f32_to_i1(i1 addrspace(1)* %out, float %in) #0 {
}
; FUNC-LABEL: {{^}}fp_to_uint_fabs_f32_to_i1:
; SI: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1.0, |s{{[0-9]+}}|
; GCN: v_cmp_eq_f32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1.0, |s{{[0-9]+}}|
define void @fp_to_uint_fabs_f32_to_i1(i1 addrspace(1)* %out, float %in) #0 {
%in.fabs = call float @llvm.fabs.f32(float %in)
%conv = fptoui float %in.fabs to i1
@ -239,5 +239,18 @@ define void @fp_to_uint_fabs_f32_to_i1(i1 addrspace(1)* %out, float %in) #0 {
ret void
}
; FUNC-LABEL: {{^}}fp_to_uint_f32_to_i16:
; The reason different instructions are used on SI and VI is because for
; SI fp_to_uint is legalized by the type legalizer and for VI it is
; legalized by the dag legalizer and they legalize fp_to_uint differently.
; SI: v_cvt_u32_f32_e32 [[VAL:v[0-9]+]], s{{[0-9]+}}
; VI: v_cvt_i32_f32_e32 [[VAL:v[0-9]+]], s{{[0-9]+}}
; GCN: buffer_store_short [[VAL]]
define void @fp_to_uint_f32_to_i16(i16 addrspace(1)* %out, float %in) #0 {
%uint = fptoui float %in to i16
store i16 %uint, i16 addrspace(1)* %out
ret void
}
attributes #0 = { nounwind }
attributes #1 = { nounwind readnone }