R600: Correctly set the src value offset for scalarized kernel args

This for some reason fixes v1i64 kernel arguments on pre-SI. This
currently breaks some other cases in the kernel-args.ll test for R600,
but I'm not particularly confident in the new output. VTX_READ_* are not
used for some of the scalarized cases, and the code reading from the
constant buffer doesn't make much sense to me.

llvm-svn: 215564
This commit is contained in:
Matt Arsenault 2014-08-13 18:14:11 +00:00
parent 5aa2194ea5
commit 74ef277774
2 changed files with 47 additions and 11 deletions

View File

@ -1705,8 +1705,13 @@ SDValue R600TargetLowering::LowerFormalArguments(
for (unsigned i = 0, e = Ins.size(); i < e; ++i) { for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
CCValAssign &VA = ArgLocs[i]; CCValAssign &VA = ArgLocs[i];
EVT VT = Ins[i].VT; const ISD::InputArg &In = Ins[i];
EVT MemVT = LocalIns[i].VT; EVT VT = In.VT;
EVT MemVT = VA.getLocVT();
if (!VT.isVector() && MemVT.isVector()) {
// Get load source type if scalarized.
MemVT = MemVT.getVectorElementType();
}
if (ShaderType != ShaderType::COMPUTE) { if (ShaderType != ShaderType::COMPUTE) {
unsigned Reg = MF.addLiveIn(VA.getLocReg(), &AMDGPU::R600_Reg128RegClass); unsigned Reg = MF.addLiveIn(VA.getLocReg(), &AMDGPU::R600_Reg128RegClass);
@ -1716,7 +1721,7 @@ SDValue R600TargetLowering::LowerFormalArguments(
} }
PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()), PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
AMDGPUAS::CONSTANT_BUFFER_0); AMDGPUAS::CONSTANT_BUFFER_0);
// i64 isn't a legal type, so the register type used ends up as i32, which // i64 isn't a legal type, so the register type used ends up as i32, which
// isn't expected here. It attempts to create this sextload, but it ends up // isn't expected here. It attempts to create this sextload, but it ends up
@ -1725,15 +1730,28 @@ SDValue R600TargetLowering::LowerFormalArguments(
// The first 36 bytes of the input buffer contains information about // The first 36 bytes of the input buffer contains information about
// thread group and global sizes. // thread group and global sizes.
ISD::LoadExtType Ext = ISD::NON_EXTLOAD;
if (MemVT.getScalarSizeInBits() != VT.getScalarSizeInBits()) {
// FIXME: This should really check the extload type, but the handling of
// extload vector parameters seems to be broken.
// FIXME: This should really check the extload type, but the handling of // Ext = In.Flags.isSExt() ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
// extload vecto parameters seems to be broken. Ext = ISD::SEXTLOAD;
//ISD::LoadExtType Ext = Ins[i].Flags.isSExt() ? ISD::SEXTLOAD : ISD::ZEXTLOAD; }
ISD::LoadExtType Ext = ISD::SEXTLOAD;
SDValue Arg = DAG.getExtLoad(Ext, DL, VT, Chain, // Compute the offset from the value.
DAG.getConstant(36 + VA.getLocMemOffset(), MVT::i32), // XXX - I think PartOffset should give you this, but it seems to give the
MachinePointerInfo(UndefValue::get(PtrTy)), // size of the register which isn't useful.
MemVT, false, false, false, 4);
unsigned ValBase = ArgLocs[In.OrigArgIndex].getLocMemOffset();
unsigned PartOffset = VA.getLocMemOffset();
MachinePointerInfo PtrInfo(UndefValue::get(PtrTy), PartOffset - ValBase);
SDValue Arg = DAG.getLoad(ISD::UNINDEXED, Ext, VT, DL, Chain,
DAG.getConstant(36 + PartOffset, MVT::i32),
DAG.getUNDEF(MVT::i32),
PtrInfo,
MemVT, false, true, true, 4);
// 4 is the preferred alignment for the CONSTANT memory space. // 4 is the preferred alignment for the CONSTANT memory space.
InVals.push_back(Arg); InVals.push_back(Arg);

View File

@ -453,3 +453,21 @@ entry:
store <16 x float> %in, <16 x float> addrspace(1)* %out, align 4 store <16 x float> %in, <16 x float> addrspace(1)* %out, align 4
ret void ret void
} }
; FUNC-LABEL: @kernel_arg_i64
; SI: S_LOAD_DWORDX2
; SI: S_LOAD_DWORDX2
; SI: BUFFER_STORE_DWORDX2
define void @kernel_arg_i64(i64 addrspace(1)* %out, i64 %a) nounwind {
store i64 %a, i64 addrspace(1)* %out, align 8
ret void
}
; XFUNC-LABEL: @kernel_arg_v1i64
; XSI: S_LOAD_DWORDX2
; XSI: S_LOAD_DWORDX2
; XSI: BUFFER_STORE_DWORDX2
; define void @kernel_arg_v1i64(<1 x i64> addrspace(1)* %out, <1 x i64> %a) nounwind {
; store <1 x i64> %a, <1 x i64> addrspace(1)* %out, align 8
; ret void
; }