2012-12-12 05:25:42 +08:00
|
|
|
//===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
|
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
/// \file
|
2018-05-01 23:54:18 +08:00
|
|
|
/// This is the parent TargetLowering class for hardware code gen
|
2012-12-12 05:25:42 +08:00
|
|
|
/// targets.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2017-11-27 21:26:38 +08:00
|
|
|
#define AMDGPU_LOG2E_F 1.44269504088896340735992468100189214f
|
|
|
|
#define AMDGPU_LN2_F 0.693147180559945309417232121458176568f
|
|
|
|
#define AMDGPU_LN10_F 2.30258509299404568401799145468436421f
|
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
#include "AMDGPUISelLowering.h"
|
2013-06-04 01:40:11 +08:00
|
|
|
#include "AMDGPU.h"
|
Re-commit AMDGPU/GlobalISel: Add support for simple shaders
Fix build when global-isel is disabled and fix a warning.
Summary: We can select constant/global G_LOAD, global G_STORE, and G_GEP.
Reviewers: qcolombet, MatzeB, t.p.northover, ab, arsenm
Subscribers: mehdi_amini, vkalintiris, kzhuravl, wdng, nhaehnle, mgorny, yaxunl, tony-tye, modocache, llvm-commits, dberris
Differential Revision: https://reviews.llvm.org/D26730
llvm-svn: 293551
2017-01-31 05:56:46 +08:00
|
|
|
#include "AMDGPUCallLowering.h"
|
2013-11-14 07:36:50 +08:00
|
|
|
#include "AMDGPUFrameLowering.h"
|
2014-06-24 02:00:31 +08:00
|
|
|
#include "AMDGPUIntrinsicInfo.h"
|
2013-03-07 17:03:52 +08:00
|
|
|
#include "AMDGPURegisterInfo.h"
|
|
|
|
#include "AMDGPUSubtarget.h"
|
2017-08-02 03:54:18 +08:00
|
|
|
#include "AMDGPUTargetMachine.h"
|
2018-03-05 23:12:21 +08:00
|
|
|
#include "Utils/AMDGPUBaseInfo.h"
|
2013-07-23 09:48:05 +08:00
|
|
|
#include "R600MachineFunctionInfo.h"
|
2017-06-06 19:49:48 +08:00
|
|
|
#include "SIInstrInfo.h"
|
2013-06-04 01:40:11 +08:00
|
|
|
#include "SIMachineFunctionInfo.h"
|
AMDGPU: Remove #include "MCTargetDesc/AMDGPUMCTargetDesc.h" from common headers
Summary:
MCTargetDesc/AMDGPUMCTargetDesc.h contains enums for all the instuction
and register defintions, which are huge so we only want to include
them where needed.
This will also make it easier if we want to split the R600 and GCN
definitions into separate tablegenerated files.
I was unable to remove AMDGPUMCTargetDesc.h from SIMachineFunctionInfo.h
because it uses some enums from the header to initialize default values
for the SIMachineFunction class, so I ended up having to remove includes of
SIMachineFunctionInfo.h from headers too.
Reviewers: arsenm, nhaehnle
Reviewed By: nhaehnle
Subscribers: MatzeB, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D46272
llvm-svn: 332930
2018-05-22 10:03:23 +08:00
|
|
|
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
|
2013-03-07 17:03:52 +08:00
|
|
|
#include "llvm/CodeGen/CallingConvLower.h"
|
2012-12-12 05:25:42 +08:00
|
|
|
#include "llvm/CodeGen/MachineFunction.h"
|
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
|
|
|
#include "llvm/CodeGen/SelectionDAG.h"
|
|
|
|
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
|
2013-06-28 23:47:08 +08:00
|
|
|
#include "llvm/IR/DataLayout.h"
|
2016-02-02 21:52:43 +08:00
|
|
|
#include "llvm/IR/DiagnosticInfo.h"
|
2017-04-28 13:31:46 +08:00
|
|
|
#include "llvm/Support/KnownBits.h"
|
2012-12-12 05:25:42 +08:00
|
|
|
using namespace llvm;
|
2014-04-23 00:42:00 +08:00
|
|
|
|
2016-06-18 13:15:53 +08:00
|
|
|
static bool allocateKernArg(unsigned ValNo, MVT ValVT, MVT LocVT,
|
|
|
|
CCValAssign::LocInfo LocInfo,
|
|
|
|
ISD::ArgFlagsTy ArgFlags, CCState &State) {
|
|
|
|
MachineFunction &MF = State.getMachineFunction();
|
|
|
|
AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
|
2013-10-23 08:44:32 +08:00
|
|
|
|
2016-09-17 05:53:00 +08:00
|
|
|
uint64_t Offset = MFI->allocateKernArg(LocVT.getStoreSize(),
|
2016-06-18 13:15:53 +08:00
|
|
|
ArgFlags.getOrigAlign());
|
|
|
|
State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, Offset, LocVT, LocInfo));
|
2013-10-23 08:44:32 +08:00
|
|
|
return true;
|
|
|
|
}
|
2012-12-12 05:25:42 +08:00
|
|
|
|
2017-04-07 01:37:27 +08:00
|
|
|
static bool allocateCCRegs(unsigned ValNo, MVT ValVT, MVT LocVT,
|
|
|
|
CCValAssign::LocInfo LocInfo,
|
|
|
|
ISD::ArgFlagsTy ArgFlags, CCState &State,
|
|
|
|
const TargetRegisterClass *RC,
|
|
|
|
unsigned NumRegs) {
|
|
|
|
ArrayRef<MCPhysReg> RegList = makeArrayRef(RC->begin(), NumRegs);
|
|
|
|
unsigned RegResult = State.AllocateReg(RegList);
|
|
|
|
if (RegResult == AMDGPU::NoRegister)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
State.addLoc(CCValAssign::getReg(ValNo, ValVT, RegResult, LocVT, LocInfo));
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool allocateSGPRTuple(unsigned ValNo, MVT ValVT, MVT LocVT,
|
|
|
|
CCValAssign::LocInfo LocInfo,
|
|
|
|
ISD::ArgFlagsTy ArgFlags, CCState &State) {
|
|
|
|
switch (LocVT.SimpleTy) {
|
|
|
|
case MVT::i64:
|
|
|
|
case MVT::f64:
|
|
|
|
case MVT::v2i32:
|
2018-06-15 23:15:46 +08:00
|
|
|
case MVT::v2f32:
|
|
|
|
case MVT::v4i16:
|
|
|
|
case MVT::v4f16: {
|
2017-04-07 01:37:27 +08:00
|
|
|
// Up to SGPR0-SGPR39
|
|
|
|
return allocateCCRegs(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State,
|
|
|
|
&AMDGPU::SGPR_64RegClass, 20);
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-05-18 05:56:25 +08:00
|
|
|
// Allocate up to VGPR31.
|
|
|
|
//
|
|
|
|
// TODO: Since there are no VGPR alignent requirements would it be better to
|
|
|
|
// split into individual scalar registers?
|
|
|
|
static bool allocateVGPRTuple(unsigned ValNo, MVT ValVT, MVT LocVT,
|
|
|
|
CCValAssign::LocInfo LocInfo,
|
|
|
|
ISD::ArgFlagsTy ArgFlags, CCState &State) {
|
|
|
|
switch (LocVT.SimpleTy) {
|
|
|
|
case MVT::i64:
|
|
|
|
case MVT::f64:
|
|
|
|
case MVT::v2i32:
|
2018-06-15 23:15:46 +08:00
|
|
|
case MVT::v2f32:
|
|
|
|
case MVT::v4i16:
|
|
|
|
case MVT::v4f16: {
|
2017-05-18 05:56:25 +08:00
|
|
|
return allocateCCRegs(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State,
|
|
|
|
&AMDGPU::VReg_64RegClass, 31);
|
|
|
|
}
|
|
|
|
case MVT::v4i32:
|
|
|
|
case MVT::v4f32:
|
|
|
|
case MVT::v2i64:
|
|
|
|
case MVT::v2f64: {
|
|
|
|
return allocateCCRegs(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State,
|
|
|
|
&AMDGPU::VReg_128RegClass, 29);
|
|
|
|
}
|
|
|
|
case MVT::v8i32:
|
|
|
|
case MVT::v8f32: {
|
|
|
|
return allocateCCRegs(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State,
|
|
|
|
&AMDGPU::VReg_256RegClass, 25);
|
|
|
|
|
|
|
|
}
|
|
|
|
case MVT::v16i32:
|
|
|
|
case MVT::v16f32: {
|
|
|
|
return allocateCCRegs(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State,
|
|
|
|
&AMDGPU::VReg_512RegClass, 17);
|
|
|
|
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-03-07 17:03:52 +08:00
|
|
|
#include "AMDGPUGenCallingConv.inc"
|
|
|
|
|
2014-06-11 11:29:54 +08:00
|
|
|
// Find a larger type to do a load / store of a vector with.
|
|
|
|
EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
|
|
|
|
unsigned StoreSize = VT.getStoreSizeInBits();
|
|
|
|
if (StoreSize <= 32)
|
|
|
|
return EVT::getIntegerVT(Ctx, StoreSize);
|
|
|
|
|
|
|
|
assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
|
|
|
|
return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
|
|
|
|
}
|
|
|
|
|
2017-11-07 01:04:37 +08:00
|
|
|
unsigned AMDGPUTargetLowering::numBitsUnsigned(SDValue Op, SelectionDAG &DAG) {
|
|
|
|
KnownBits Known;
|
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
DAG.computeKnownBits(Op, Known);
|
|
|
|
|
|
|
|
return VT.getSizeInBits() - Known.countMinLeadingZeros();
|
|
|
|
}
|
|
|
|
|
|
|
|
unsigned AMDGPUTargetLowering::numBitsSigned(SDValue Op, SelectionDAG &DAG) {
|
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
|
|
|
|
// In order for this to be a signed 24-bit value, bit 23, must
|
|
|
|
// be a sign bit.
|
|
|
|
return VT.getSizeInBits() - DAG.ComputeNumSignBits(Op);
|
|
|
|
}
|
|
|
|
|
2016-06-24 14:30:11 +08:00
|
|
|
AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
|
2015-01-31 07:24:40 +08:00
|
|
|
const AMDGPUSubtarget &STI)
|
|
|
|
: TargetLowering(TM), Subtarget(&STI) {
|
2017-03-27 22:04:01 +08:00
|
|
|
AMDGPUASI = AMDGPU::getAMDGPUAS(TM);
|
2016-05-21 10:27:49 +08:00
|
|
|
// Lower floating point store/load to integer store/load to reduce the number
|
|
|
|
// of patterns in tablegen.
|
|
|
|
setOperationAction(ISD::LOAD, MVT::f32, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
|
2014-06-24 02:00:44 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
|
2014-06-24 02:00:44 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
|
2015-08-27 02:37:13 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
|
2012-12-12 05:25:42 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
|
2012-12-12 05:25:42 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::LOAD, MVT::i64, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32);
|
2014-09-11 05:44:27 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::LOAD, MVT::f64, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32);
|
2015-02-21 06:10:41 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
|
|
|
|
AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v4i32);
|
|
|
|
|
|
|
|
// There are no 64-bit extloads. These should be done as a 32-bit extload and
|
|
|
|
// an extension to 64-bit.
|
|
|
|
for (MVT VT : MVT::integer_valuetypes()) {
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (MVT VT : MVT::integer_valuetypes()) {
|
|
|
|
if (VT == MVT::i64)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal);
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal);
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
|
|
|
|
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal);
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal);
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
|
|
|
|
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (MVT VT : MVT::integer_vector_valuetypes()) {
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
|
|
|
|
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
|
|
|
|
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
|
|
|
|
}
|
|
|
|
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand);
|
|
|
|
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f32, Expand);
|
|
|
|
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
|
|
|
|
setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
|
2015-02-21 06:10:45 +08:00
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
setOperationAction(ISD::STORE, MVT::f32, Promote);
|
|
|
|
AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
|
|
|
|
|
2013-07-19 05:43:42 +08:00
|
|
|
setOperationAction(ISD::STORE, MVT::v2f32, Promote);
|
|
|
|
AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
|
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
setOperationAction(ISD::STORE, MVT::v4f32, Promote);
|
|
|
|
AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
|
|
|
|
|
2013-10-23 08:44:32 +08:00
|
|
|
setOperationAction(ISD::STORE, MVT::v8f32, Promote);
|
|
|
|
AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
|
|
|
|
|
|
|
|
setOperationAction(ISD::STORE, MVT::v16f32, Promote);
|
|
|
|
AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
|
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::STORE, MVT::i64, Promote);
|
|
|
|
AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
|
|
|
|
|
|
|
|
setOperationAction(ISD::STORE, MVT::v2i64, Promote);
|
|
|
|
AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32);
|
|
|
|
|
2013-07-13 02:14:56 +08:00
|
|
|
setOperationAction(ISD::STORE, MVT::f64, Promote);
|
2016-05-21 10:27:49 +08:00
|
|
|
AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32);
|
2013-07-13 02:14:56 +08:00
|
|
|
|
2014-05-09 02:01:56 +08:00
|
|
|
setOperationAction(ISD::STORE, MVT::v2f64, Promote);
|
2016-05-21 10:27:49 +08:00
|
|
|
AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v4i32);
|
2013-08-26 23:05:44 +08:00
|
|
|
|
2014-03-13 02:45:52 +08:00
|
|
|
setTruncStoreAction(MVT::i64, MVT::i1, Expand);
|
2016-05-21 10:27:49 +08:00
|
|
|
setTruncStoreAction(MVT::i64, MVT::i8, Expand);
|
|
|
|
setTruncStoreAction(MVT::i64, MVT::i16, Expand);
|
|
|
|
setTruncStoreAction(MVT::i64, MVT::i32, Expand);
|
|
|
|
|
2014-03-13 02:45:52 +08:00
|
|
|
setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
|
2016-05-21 10:27:49 +08:00
|
|
|
setTruncStoreAction(MVT::v2i64, MVT::v2i8, Expand);
|
|
|
|
setTruncStoreAction(MVT::v2i64, MVT::v2i16, Expand);
|
|
|
|
setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand);
|
2014-03-13 02:45:52 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setTruncStoreAction(MVT::f32, MVT::f16, Expand);
|
|
|
|
setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
|
|
|
|
setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
|
|
|
|
setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
|
2014-03-13 02:45:52 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setTruncStoreAction(MVT::f64, MVT::f16, Expand);
|
|
|
|
setTruncStoreAction(MVT::f64, MVT::f32, Expand);
|
2012-12-12 05:25:42 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
|
|
|
|
setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand);
|
2013-07-19 05:43:48 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setTruncStoreAction(MVT::v4f64, MVT::v4f32, Expand);
|
|
|
|
setTruncStoreAction(MVT::v4f64, MVT::v4f16, Expand);
|
2012-12-12 05:25:42 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setTruncStoreAction(MVT::v8f64, MVT::v8f32, Expand);
|
|
|
|
setTruncStoreAction(MVT::v8f64, MVT::v8f16, Expand);
|
2013-10-23 08:44:32 +08:00
|
|
|
|
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::Constant, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::Constant, MVT::i64, Legal);
|
|
|
|
setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
|
2013-07-13 02:14:56 +08:00
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::BR_JT, MVT::Other, Expand);
|
|
|
|
setOperationAction(ISD::BRIND, MVT::Other, Expand);
|
|
|
|
|
|
|
|
// This is totally unsupported, just custom lower to produce an error.
|
|
|
|
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
|
|
|
|
|
|
|
|
// Library functions. These default to Expand, but we have instructions
|
|
|
|
// for them.
|
|
|
|
setOperationAction(ISD::FCEIL, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FEXP2, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FPOW, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FLOG2, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FABS, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FRINT, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
|
|
|
|
|
|
|
|
setOperationAction(ISD::FROUND, MVT::f32, Custom);
|
|
|
|
setOperationAction(ISD::FROUND, MVT::f64, Custom);
|
|
|
|
|
2017-11-27 21:26:38 +08:00
|
|
|
setOperationAction(ISD::FLOG, MVT::f32, Custom);
|
|
|
|
setOperationAction(ISD::FLOG10, MVT::f32, Custom);
|
|
|
|
|
|
|
|
if (Subtarget->has16BitInsts()) {
|
|
|
|
setOperationAction(ISD::FLOG, MVT::f16, Custom);
|
|
|
|
setOperationAction(ISD::FLOG10, MVT::f16, Custom);
|
|
|
|
}
|
|
|
|
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
|
|
|
|
setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
|
|
|
|
|
|
|
|
setOperationAction(ISD::FREM, MVT::f32, Custom);
|
|
|
|
setOperationAction(ISD::FREM, MVT::f64, Custom);
|
|
|
|
|
|
|
|
// v_mad_f32 does not support denormals according to some sources.
|
|
|
|
if (!Subtarget->hasFP32Denormals())
|
|
|
|
setOperationAction(ISD::FMAD, MVT::f32, Legal);
|
|
|
|
|
|
|
|
// Expand to fneg + fadd.
|
|
|
|
setOperationAction(ISD::FSUB, MVT::f64, Expand);
|
2014-05-09 02:01:56 +08:00
|
|
|
|
2013-08-15 07:25:00 +08:00
|
|
|
setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
|
|
|
|
setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
|
2014-02-14 07:34:15 +08:00
|
|
|
setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
|
|
|
|
setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
|
2013-08-15 07:25:00 +08:00
|
|
|
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
|
2014-02-14 07:34:15 +08:00
|
|
|
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
|
|
|
|
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
|
|
|
|
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
|
|
|
|
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
|
|
|
|
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
|
2013-08-01 23:23:42 +08:00
|
|
|
|
2014-06-19 01:05:26 +08:00
|
|
|
if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
|
2014-06-19 01:05:30 +08:00
|
|
|
setOperationAction(ISD::FCEIL, MVT::f64, Custom);
|
|
|
|
setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
|
2014-06-19 01:05:26 +08:00
|
|
|
setOperationAction(ISD::FRINT, MVT::f64, Custom);
|
2014-06-19 01:05:30 +08:00
|
|
|
setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
|
2014-06-19 01:05:26 +08:00
|
|
|
}
|
|
|
|
|
2014-06-11 03:00:20 +08:00
|
|
|
if (!Subtarget->hasBFI()) {
|
|
|
|
// fcopysign can be done in a single instruction with BFI.
|
|
|
|
setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
|
|
|
|
setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
|
|
|
|
}
|
|
|
|
|
2014-07-18 16:43:24 +08:00
|
|
|
setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
|
2016-11-02 00:31:48 +08:00
|
|
|
setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom);
|
2017-03-16 03:04:26 +08:00
|
|
|
setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom);
|
2014-07-18 16:43:24 +08:00
|
|
|
|
2014-06-14 01:20:53 +08:00
|
|
|
const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
|
|
|
|
for (MVT VT : ScalarIntVTs) {
|
2016-05-21 10:27:49 +08:00
|
|
|
// These should use [SU]DIVREM, so set them to expand
|
2014-08-13 01:31:17 +08:00
|
|
|
setOperationAction(ISD::SDIV, VT, Expand);
|
2016-05-21 10:27:49 +08:00
|
|
|
setOperationAction(ISD::UDIV, VT, Expand);
|
|
|
|
setOperationAction(ISD::SREM, VT, Expand);
|
|
|
|
setOperationAction(ISD::UREM, VT, Expand);
|
2014-06-16 05:08:58 +08:00
|
|
|
|
2014-06-14 01:20:53 +08:00
|
|
|
// GPU does not have divrem function for signed or unsigned.
|
2014-06-23 05:43:00 +08:00
|
|
|
setOperationAction(ISD::SDIVREM, VT, Custom);
|
2014-06-16 05:08:58 +08:00
|
|
|
setOperationAction(ISD::UDIVREM, VT, Custom);
|
2014-06-14 01:20:53 +08:00
|
|
|
|
|
|
|
// GPU does not have [S|U]MUL_LOHI functions as a single instruction.
|
|
|
|
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
|
|
|
|
setOperationAction(ISD::UMUL_LOHI, VT, Expand);
|
|
|
|
|
|
|
|
setOperationAction(ISD::BSWAP, VT, Expand);
|
|
|
|
setOperationAction(ISD::CTTZ, VT, Expand);
|
|
|
|
setOperationAction(ISD::CTLZ, VT, Expand);
|
Set ADDE/ADDC/SUBE/SUBC to expand by default
Summary:
They've been deprecated in favor of UADDO/ADDCARRY or USUBO/SUBCARRY for a while.
Target that uses these opcodes are changed in order to ensure their behavior doesn't change.
Reviewers: efriedma, craig.topper, dblaikie, bkramer
Subscribers: jholewinski, arsenm, jyknight, sdardis, nemanjai, nhaehnle, kbarton, fedor.sergeev, asb, rbar, johnrusso, simoncook, jordy.potman.lists, apazos, sabuasal, niosHD, jrtc27, zzheng, edward-jones, mgrang, atanasyan, llvm-commits
Differential Revision: https://reviews.llvm.org/D47422
llvm-svn: 333748
2018-06-01 21:21:33 +08:00
|
|
|
|
|
|
|
// AMDGPU uses ADDC/SUBC/ADDE/SUBE
|
|
|
|
setOperationAction(ISD::ADDC, VT, Legal);
|
|
|
|
setOperationAction(ISD::SUBC, VT, Legal);
|
|
|
|
setOperationAction(ISD::ADDE, VT, Legal);
|
|
|
|
setOperationAction(ISD::SUBE, VT, Legal);
|
2014-06-14 01:20:53 +08:00
|
|
|
}
|
|
|
|
|
2014-06-11 03:18:28 +08:00
|
|
|
if (!Subtarget->hasBCNT(32))
|
|
|
|
setOperationAction(ISD::CTPOP, MVT::i32, Expand);
|
|
|
|
|
|
|
|
if (!Subtarget->hasBCNT(64))
|
|
|
|
setOperationAction(ISD::CTPOP, MVT::i64, Expand);
|
|
|
|
|
2014-06-16 05:08:58 +08:00
|
|
|
// The hardware supports 32-bit ROTR, but not ROTL.
|
|
|
|
setOperationAction(ISD::ROTL, MVT::i32, Expand);
|
|
|
|
setOperationAction(ISD::ROTL, MVT::i64, Expand);
|
|
|
|
setOperationAction(ISD::ROTR, MVT::i64, Expand);
|
|
|
|
|
|
|
|
setOperationAction(ISD::MUL, MVT::i64, Expand);
|
|
|
|
setOperationAction(ISD::MULHU, MVT::i64, Expand);
|
|
|
|
setOperationAction(ISD::MULHS, MVT::i64, Expand);
|
|
|
|
setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
|
2014-10-04 07:54:41 +08:00
|
|
|
setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
|
2014-10-04 07:54:56 +08:00
|
|
|
setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
|
2014-06-16 05:08:58 +08:00
|
|
|
setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
|
2014-06-11 03:18:21 +08:00
|
|
|
|
2015-06-09 08:52:37 +08:00
|
|
|
setOperationAction(ISD::SMIN, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::UMIN, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::SMAX, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::UMAX, MVT::i32, Legal);
|
|
|
|
|
2016-01-12 01:02:00 +08:00
|
|
|
if (Subtarget->hasFFBH())
|
|
|
|
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Custom);
|
2014-07-15 23:51:09 +08:00
|
|
|
|
2016-04-28 11:34:31 +08:00
|
|
|
if (Subtarget->hasFFBL())
|
2017-10-13 03:37:14 +08:00
|
|
|
setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Custom);
|
2016-01-12 00:37:46 +08:00
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
setOperationAction(ISD::CTTZ, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Custom);
|
2016-01-12 00:50:29 +08:00
|
|
|
setOperationAction(ISD::CTLZ, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
|
|
|
|
|
2016-03-01 12:58:17 +08:00
|
|
|
// We only really have 32-bit BFE instructions (and 16-bit on VI).
|
|
|
|
//
|
|
|
|
// On SI+ there are 64-bit BFEs, but they are scalar only and there isn't any
|
|
|
|
// effort to match them now. We want this to be false for i64 cases when the
|
|
|
|
// extraction isn't restricted to the upper or lower half. Ideally we would
|
|
|
|
// have some pass reduce 64-bit extracts to 32-bit if possible. Extracts that
|
|
|
|
// span the midpoint are probably relatively rare, so don't worry about them
|
|
|
|
// for now.
|
|
|
|
if (Subtarget->hasBFE())
|
|
|
|
setHasExtractBitsInsn(true);
|
|
|
|
|
2014-06-14 01:20:53 +08:00
|
|
|
static const MVT::SimpleValueType VectorIntTypes[] = {
|
2013-08-22 06:14:17 +08:00
|
|
|
MVT::v2i32, MVT::v4i32
|
2013-06-25 21:55:57 +08:00
|
|
|
};
|
|
|
|
|
2014-06-14 01:20:53 +08:00
|
|
|
for (MVT VT : VectorIntTypes) {
|
2014-06-11 03:18:21 +08:00
|
|
|
// Expand the following operations for the current type by default.
|
2013-06-25 21:55:57 +08:00
|
|
|
setOperationAction(ISD::ADD, VT, Expand);
|
|
|
|
setOperationAction(ISD::AND, VT, Expand);
|
2013-07-30 22:31:03 +08:00
|
|
|
setOperationAction(ISD::FP_TO_SINT, VT, Expand);
|
|
|
|
setOperationAction(ISD::FP_TO_UINT, VT, Expand);
|
2013-06-25 21:55:57 +08:00
|
|
|
setOperationAction(ISD::MUL, VT, Expand);
|
2016-11-01 18:26:48 +08:00
|
|
|
setOperationAction(ISD::MULHU, VT, Expand);
|
|
|
|
setOperationAction(ISD::MULHS, VT, Expand);
|
2013-06-25 21:55:57 +08:00
|
|
|
setOperationAction(ISD::OR, VT, Expand);
|
|
|
|
setOperationAction(ISD::SHL, VT, Expand);
|
|
|
|
setOperationAction(ISD::SRA, VT, Expand);
|
2014-06-13 12:00:30 +08:00
|
|
|
setOperationAction(ISD::SRL, VT, Expand);
|
|
|
|
setOperationAction(ISD::ROTL, VT, Expand);
|
|
|
|
setOperationAction(ISD::ROTR, VT, Expand);
|
2013-06-25 21:55:57 +08:00
|
|
|
setOperationAction(ISD::SUB, VT, Expand);
|
2014-06-13 12:00:30 +08:00
|
|
|
setOperationAction(ISD::SINT_TO_FP, VT, Expand);
|
2013-07-30 22:31:03 +08:00
|
|
|
setOperationAction(ISD::UINT_TO_FP, VT, Expand);
|
2014-06-23 05:43:00 +08:00
|
|
|
setOperationAction(ISD::SDIV, VT, Expand);
|
2014-06-16 05:08:58 +08:00
|
|
|
setOperationAction(ISD::UDIV, VT, Expand);
|
2014-06-14 01:20:53 +08:00
|
|
|
setOperationAction(ISD::SREM, VT, Expand);
|
2013-06-25 21:55:57 +08:00
|
|
|
setOperationAction(ISD::UREM, VT, Expand);
|
2014-06-14 01:20:53 +08:00
|
|
|
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
|
|
|
|
setOperationAction(ISD::UMUL_LOHI, VT, Expand);
|
2014-06-23 05:43:00 +08:00
|
|
|
setOperationAction(ISD::SDIVREM, VT, Custom);
|
2015-10-15 17:18:47 +08:00
|
|
|
setOperationAction(ISD::UDIVREM, VT, Expand);
|
2014-03-07 01:34:03 +08:00
|
|
|
setOperationAction(ISD::SELECT, VT, Expand);
|
2013-07-19 05:43:35 +08:00
|
|
|
setOperationAction(ISD::VSELECT, VT, Expand);
|
2014-06-24 02:00:44 +08:00
|
|
|
setOperationAction(ISD::SELECT_CC, VT, Expand);
|
2013-06-25 21:55:57 +08:00
|
|
|
setOperationAction(ISD::XOR, VT, Expand);
|
2014-06-10 00:20:25 +08:00
|
|
|
setOperationAction(ISD::BSWAP, VT, Expand);
|
2014-06-11 03:18:21 +08:00
|
|
|
setOperationAction(ISD::CTPOP, VT, Expand);
|
|
|
|
setOperationAction(ISD::CTTZ, VT, Expand);
|
|
|
|
setOperationAction(ISD::CTLZ, VT, Expand);
|
2014-06-24 02:00:44 +08:00
|
|
|
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
|
2017-10-04 05:31:24 +08:00
|
|
|
setOperationAction(ISD::SETCC, VT, Expand);
|
2013-06-25 21:55:57 +08:00
|
|
|
}
|
2013-08-17 07:51:24 +08:00
|
|
|
|
2014-06-14 01:20:53 +08:00
|
|
|
static const MVT::SimpleValueType FloatVectorTypes[] = {
|
2013-08-22 06:14:17 +08:00
|
|
|
MVT::v2f32, MVT::v4f32
|
2013-08-17 07:51:24 +08:00
|
|
|
};
|
|
|
|
|
2014-06-14 01:20:53 +08:00
|
|
|
for (MVT VT : FloatVectorTypes) {
|
2013-11-28 05:23:39 +08:00
|
|
|
setOperationAction(ISD::FABS, VT, Expand);
|
2014-10-22 07:01:01 +08:00
|
|
|
setOperationAction(ISD::FMINNUM, VT, Expand);
|
|
|
|
setOperationAction(ISD::FMAXNUM, VT, Expand);
|
2013-08-17 07:51:24 +08:00
|
|
|
setOperationAction(ISD::FADD, VT, Expand);
|
2014-06-19 01:57:29 +08:00
|
|
|
setOperationAction(ISD::FCEIL, VT, Expand);
|
2014-05-02 23:41:47 +08:00
|
|
|
setOperationAction(ISD::FCOS, VT, Expand);
|
2013-08-17 07:51:24 +08:00
|
|
|
setOperationAction(ISD::FDIV, VT, Expand);
|
2014-06-21 01:06:05 +08:00
|
|
|
setOperationAction(ISD::FEXP2, VT, Expand);
|
2014-06-21 01:06:07 +08:00
|
|
|
setOperationAction(ISD::FLOG2, VT, Expand);
|
2014-09-11 05:44:27 +08:00
|
|
|
setOperationAction(ISD::FREM, VT, Expand);
|
2017-11-27 21:26:38 +08:00
|
|
|
setOperationAction(ISD::FLOG, VT, Expand);
|
|
|
|
setOperationAction(ISD::FLOG10, VT, Expand);
|
2014-02-05 01:18:37 +08:00
|
|
|
setOperationAction(ISD::FPOW, VT, Expand);
|
2013-08-17 07:51:29 +08:00
|
|
|
setOperationAction(ISD::FFLOOR, VT, Expand);
|
2013-12-20 13:11:55 +08:00
|
|
|
setOperationAction(ISD::FTRUNC, VT, Expand);
|
2013-08-17 07:51:24 +08:00
|
|
|
setOperationAction(ISD::FMUL, VT, Expand);
|
2014-06-26 09:28:05 +08:00
|
|
|
setOperationAction(ISD::FMA, VT, Expand);
|
2013-08-17 07:51:33 +08:00
|
|
|
setOperationAction(ISD::FRINT, VT, Expand);
|
2014-06-19 06:03:45 +08:00
|
|
|
setOperationAction(ISD::FNEARBYINT, VT, Expand);
|
2013-10-30 00:37:20 +08:00
|
|
|
setOperationAction(ISD::FSQRT, VT, Expand);
|
2014-05-02 23:41:47 +08:00
|
|
|
setOperationAction(ISD::FSIN, VT, Expand);
|
2013-08-17 07:51:24 +08:00
|
|
|
setOperationAction(ISD::FSUB, VT, Expand);
|
2014-06-01 15:38:21 +08:00
|
|
|
setOperationAction(ISD::FNEG, VT, Expand);
|
|
|
|
setOperationAction(ISD::VSELECT, VT, Expand);
|
2014-06-24 02:00:44 +08:00
|
|
|
setOperationAction(ISD::SELECT_CC, VT, Expand);
|
2014-06-11 03:00:20 +08:00
|
|
|
setOperationAction(ISD::FCOPYSIGN, VT, Expand);
|
2014-06-24 02:00:44 +08:00
|
|
|
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
|
2017-10-04 05:45:01 +08:00
|
|
|
setOperationAction(ISD::SETCC, VT, Expand);
|
2013-08-17 07:51:24 +08:00
|
|
|
}
|
2014-03-18 02:58:11 +08:00
|
|
|
|
2016-05-26 01:34:58 +08:00
|
|
|
// This causes using an unrolled select operation rather than expansion with
|
|
|
|
// bit operations. This is in general better, but the alternative using BFI
|
|
|
|
// instructions may be better if the select sources are SGPRs.
|
|
|
|
setOperationAction(ISD::SELECT, MVT::v2f32, Promote);
|
|
|
|
AddPromotedToType(ISD::SELECT, MVT::v2f32, MVT::v2i32);
|
|
|
|
|
|
|
|
setOperationAction(ISD::SELECT, MVT::v4f32, Promote);
|
|
|
|
AddPromotedToType(ISD::SELECT, MVT::v4f32, MVT::v4i32);
|
|
|
|
|
2016-12-10 01:49:14 +08:00
|
|
|
// There are no libcalls of any kind.
|
|
|
|
for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I)
|
|
|
|
setLibcallName(static_cast<RTLIB::Libcall>(I), nullptr);
|
|
|
|
|
2014-11-27 05:23:15 +08:00
|
|
|
setBooleanContents(ZeroOrNegativeOneBooleanContent);
|
|
|
|
setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
|
|
|
|
|
2014-06-14 01:20:53 +08:00
|
|
|
setSchedulingPreference(Sched::RegPressure);
|
|
|
|
setJumpIsExpensive(true);
|
2017-01-11 03:08:15 +08:00
|
|
|
|
|
|
|
// FIXME: This is only partially true. If we have to do vector compares, any
|
|
|
|
// SGPR pair can be a condition register. If we have a uniform condition, we
|
|
|
|
// are better off doing SALU operations, where there is only one SCC. For now,
|
|
|
|
// we don't have a way of knowing during instruction selection if a condition
|
|
|
|
// will be uniform and we always use vector compares. Assume we are using
|
|
|
|
// vector compares until that is fixed.
|
2016-11-29 02:58:49 +08:00
|
|
|
setHasMultipleConditionRegisters(true);
|
2014-06-14 01:20:53 +08:00
|
|
|
|
2014-08-09 11:46:58 +08:00
|
|
|
// SI at least has hardware support for floating point exceptions, but no way
|
|
|
|
// of using or handling them is implemented. They are also optional in OpenCL
|
|
|
|
// (Section 7.3)
|
2016-01-29 04:53:42 +08:00
|
|
|
setHasFloatingPointExceptions(Subtarget->hasFPExceptions());
|
2014-08-09 11:46:58 +08:00
|
|
|
|
2014-06-24 02:00:52 +08:00
|
|
|
PredictableSelectIsExpensive = false;
|
|
|
|
|
2017-02-03 02:24:55 +08:00
|
|
|
// We want to find all load dependencies for long chains of stores to enable
|
|
|
|
// merging into very wide vectors. The problem is with vectors with > 4
|
|
|
|
// elements. MergeConsecutiveStores will attempt to merge these because x8/x16
|
|
|
|
// vectors are a legal type, even though we have to split the loads
|
|
|
|
// usually. When we can more precisely specify load legality per address
|
|
|
|
// space, we should be able to make FindBetterChain/MergeConsecutiveStores
|
|
|
|
// smarter so that they can figure out what to do in 2 iterations without all
|
|
|
|
// N > 4 stores on the same chain.
|
|
|
|
GatherAllAliasesMaxDepth = 16;
|
|
|
|
|
2017-02-10 06:00:42 +08:00
|
|
|
// memcpy/memmove/memset are expanded in the IR, so we shouldn't need to worry
|
|
|
|
// about these during lowering.
|
|
|
|
MaxStoresPerMemcpy = 0xffffffff;
|
|
|
|
MaxStoresPerMemmove = 0xffffffff;
|
|
|
|
MaxStoresPerMemset = 0xffffffff;
|
2016-05-21 10:27:49 +08:00
|
|
|
|
|
|
|
setTargetDAGCombine(ISD::BITCAST);
|
|
|
|
setTargetDAGCombine(ISD::SHL);
|
|
|
|
setTargetDAGCombine(ISD::SRA);
|
|
|
|
setTargetDAGCombine(ISD::SRL);
|
2018-05-10 02:37:39 +08:00
|
|
|
setTargetDAGCombine(ISD::TRUNCATE);
|
2016-05-21 10:27:49 +08:00
|
|
|
setTargetDAGCombine(ISD::MUL);
|
2016-08-27 09:32:27 +08:00
|
|
|
setTargetDAGCombine(ISD::MULHU);
|
|
|
|
setTargetDAGCombine(ISD::MULHS);
|
2016-05-21 10:27:49 +08:00
|
|
|
setTargetDAGCombine(ISD::SELECT);
|
|
|
|
setTargetDAGCombine(ISD::SELECT_CC);
|
|
|
|
setTargetDAGCombine(ISD::STORE);
|
|
|
|
setTargetDAGCombine(ISD::FADD);
|
|
|
|
setTargetDAGCombine(ISD::FSUB);
|
2017-01-12 08:09:34 +08:00
|
|
|
setTargetDAGCombine(ISD::FNEG);
|
2017-02-02 10:27:04 +08:00
|
|
|
setTargetDAGCombine(ISD::FABS);
|
2017-07-15 13:52:59 +08:00
|
|
|
setTargetDAGCombine(ISD::AssertZext);
|
|
|
|
setTargetDAGCombine(ISD::AssertSext);
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
|
|
|
|
2013-08-06 06:22:07 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Target Information
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2017-02-03 07:21:23 +08:00
|
|
|
LLVM_READNONE
|
2017-01-13 02:58:15 +08:00
|
|
|
static bool fnegFoldsIntoOp(unsigned Opc) {
|
|
|
|
switch (Opc) {
|
|
|
|
case ISD::FADD:
|
|
|
|
case ISD::FSUB:
|
|
|
|
case ISD::FMUL:
|
|
|
|
case ISD::FMA:
|
|
|
|
case ISD::FMAD:
|
2017-02-03 08:23:15 +08:00
|
|
|
case ISD::FMINNUM:
|
|
|
|
case ISD::FMAXNUM:
|
2017-01-13 02:58:15 +08:00
|
|
|
case ISD::FSIN:
|
2017-01-26 09:25:36 +08:00
|
|
|
case ISD::FTRUNC:
|
|
|
|
case ISD::FRINT:
|
|
|
|
case ISD::FNEARBYINT:
|
2017-01-13 02:58:15 +08:00
|
|
|
case AMDGPUISD::RCP:
|
|
|
|
case AMDGPUISD::RCP_LEGACY:
|
2018-06-27 23:33:33 +08:00
|
|
|
case AMDGPUISD::RCP_IFLAG:
|
2017-01-13 02:58:15 +08:00
|
|
|
case AMDGPUISD::SIN_HW:
|
|
|
|
case AMDGPUISD::FMUL_LEGACY:
|
2017-02-03 08:51:50 +08:00
|
|
|
case AMDGPUISD::FMIN_LEGACY:
|
|
|
|
case AMDGPUISD::FMAX_LEGACY:
|
2017-01-13 02:58:15 +08:00
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-02-03 07:21:23 +08:00
|
|
|
/// \p returns true if the operation will definitely need to use a 64-bit
|
|
|
|
/// encoding, and thus will use a VOP3 encoding regardless of the source
|
|
|
|
/// modifiers.
|
|
|
|
LLVM_READONLY
|
|
|
|
static bool opMustUseVOP3Encoding(const SDNode *N, MVT VT) {
|
|
|
|
return N->getNumOperands() > 2 || VT == MVT::f64;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Most FP instructions support source modifiers, but this could be refined
|
|
|
|
// slightly.
|
|
|
|
LLVM_READONLY
|
|
|
|
static bool hasSourceMods(const SDNode *N) {
|
|
|
|
if (isa<MemSDNode>(N))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
switch (N->getOpcode()) {
|
|
|
|
case ISD::CopyToReg:
|
|
|
|
case ISD::SELECT:
|
|
|
|
case ISD::FDIV:
|
|
|
|
case ISD::FREM:
|
|
|
|
case ISD::INLINEASM:
|
|
|
|
case AMDGPUISD::INTERP_P1:
|
|
|
|
case AMDGPUISD::INTERP_P2:
|
|
|
|
case AMDGPUISD::DIV_SCALE:
|
2017-05-12 01:26:25 +08:00
|
|
|
|
|
|
|
// TODO: Should really be looking at the users of the bitcast. These are
|
|
|
|
// problematic because bitcasts are used to legalize all stores to integer
|
|
|
|
// types.
|
|
|
|
case ISD::BITCAST:
|
2017-02-03 07:21:23 +08:00
|
|
|
return false;
|
|
|
|
default:
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-05-12 01:26:25 +08:00
|
|
|
bool AMDGPUTargetLowering::allUsesHaveSourceMods(const SDNode *N,
|
|
|
|
unsigned CostThreshold) {
|
2017-02-03 07:21:23 +08:00
|
|
|
// Some users (such as 3-operand FMA/MAD) must use a VOP3 encoding, and thus
|
|
|
|
// it is truly free to use a source modifier in all cases. If there are
|
|
|
|
// multiple users but for each one will necessitate using VOP3, there will be
|
|
|
|
// a code size increase. Try to avoid increasing code size unless we know it
|
|
|
|
// will save on the instruction count.
|
|
|
|
unsigned NumMayIncreaseSize = 0;
|
|
|
|
MVT VT = N->getValueType(0).getScalarType().getSimpleVT();
|
|
|
|
|
|
|
|
// XXX - Should this limit number of uses to check?
|
|
|
|
for (const SDNode *U : N->uses()) {
|
|
|
|
if (!hasSourceMods(U))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (!opMustUseVOP3Encoding(U, VT)) {
|
|
|
|
if (++NumMayIncreaseSize > CostThreshold)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2015-07-09 10:09:04 +08:00
|
|
|
MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
|
2013-08-06 06:22:07 +08:00
|
|
|
return MVT::i32;
|
|
|
|
}
|
|
|
|
|
2014-06-24 02:00:52 +08:00
|
|
|
bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2014-06-16 04:23:38 +08:00
|
|
|
// The backend supports 32 and 64 bit floating point immediates.
|
|
|
|
// FIXME: Why are we reporting vectors of FP immediates as legal?
|
|
|
|
bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
|
|
|
|
EVT ScalarVT = VT.getScalarType();
|
2016-12-22 11:05:30 +08:00
|
|
|
return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64 ||
|
|
|
|
(ScalarVT == MVT::f16 && Subtarget->has16BitInsts()));
|
2014-06-16 04:23:38 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
// We don't want to shrink f64 / f32 constants.
|
|
|
|
bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
|
|
|
|
EVT ScalarVT = VT.getScalarType();
|
|
|
|
return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
|
|
|
|
}
|
|
|
|
|
2014-12-12 08:00:24 +08:00
|
|
|
bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
|
|
|
|
ISD::LoadExtType,
|
|
|
|
EVT NewVT) const {
|
|
|
|
|
|
|
|
unsigned NewSize = NewVT.getStoreSizeInBits();
|
|
|
|
|
|
|
|
// If we are reducing to a 32-bit load, this is always better.
|
|
|
|
if (NewSize == 32)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
EVT OldVT = N->getValueType(0);
|
|
|
|
unsigned OldSize = OldVT.getStoreSizeInBits();
|
|
|
|
|
|
|
|
// Don't produce extloads from sub 32-bit types. SI doesn't have scalar
|
|
|
|
// extloads, so doing one requires using a buffer_load. In cases where we
|
|
|
|
// still couldn't use a scalar load, using the wider load shouldn't really
|
|
|
|
// hurt anything.
|
|
|
|
|
|
|
|
// If the old size already had to be an extload, there's no harm in continuing
|
|
|
|
// to reduce the width.
|
|
|
|
return (OldSize < 32);
|
|
|
|
}
|
|
|
|
|
2013-11-15 12:42:23 +08:00
|
|
|
bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
|
|
|
|
EVT CastTy) const {
|
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
assert(LoadTy.getSizeInBits() == CastTy.getSizeInBits());
|
|
|
|
|
|
|
|
if (LoadTy.getScalarType() == MVT::i32)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
unsigned LScalarSize = LoadTy.getScalarSizeInBits();
|
|
|
|
unsigned CastScalarSize = CastTy.getScalarSizeInBits();
|
2013-11-15 12:42:23 +08:00
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
return (LScalarSize < CastScalarSize) ||
|
|
|
|
(CastScalarSize >= 32);
|
2013-11-15 12:42:23 +08:00
|
|
|
}
|
2013-08-06 06:22:07 +08:00
|
|
|
|
2015-01-14 03:46:48 +08:00
|
|
|
// SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
|
|
|
|
// profitable with the expansion for 64-bit since it's generally good to
|
|
|
|
// speculate things.
|
|
|
|
// FIXME: These should really have the size as a parameter.
|
|
|
|
bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2018-03-05 23:12:21 +08:00
|
|
|
bool AMDGPUTargetLowering::isSDNodeAlwaysUniform(const SDNode * N) const {
|
|
|
|
switch (N->getOpcode()) {
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
case ISD::EntryToken:
|
|
|
|
case ISD::TokenFactor:
|
|
|
|
return true;
|
|
|
|
case ISD::INTRINSIC_WO_CHAIN:
|
|
|
|
{
|
|
|
|
unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
|
|
|
|
switch (IntrID) {
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
case Intrinsic::amdgcn_readfirstlane:
|
|
|
|
case Intrinsic::amdgcn_readlane:
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case ISD::LOAD:
|
|
|
|
{
|
|
|
|
const LoadSDNode * L = dyn_cast<LoadSDNode>(N);
|
|
|
|
if (L->getMemOperand()->getAddrSpace()
|
|
|
|
== Subtarget->getAMDGPUAS().CONSTANT_ADDRESS_32BIT)
|
|
|
|
return true;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-07-24 07:55:03 +08:00
|
|
|
//===---------------------------------------------------------------------===//
|
|
|
|
// Target Properties
|
|
|
|
//===---------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
|
|
|
|
assert(VT.isFloatingPoint());
|
2017-02-28 06:15:25 +08:00
|
|
|
|
|
|
|
// Packed operations do not have a fabs modifier.
|
|
|
|
return VT == MVT::f32 || VT == MVT::f64 ||
|
|
|
|
(Subtarget->has16BitInsts() && VT == MVT::f16);
|
2013-07-24 07:55:03 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
|
2017-02-28 06:15:25 +08:00
|
|
|
assert(VT.isFloatingPoint());
|
|
|
|
return VT == MVT::f32 || VT == MVT::f64 ||
|
|
|
|
(Subtarget->has16BitInsts() && VT == MVT::f16) ||
|
|
|
|
(Subtarget->hasVOP3PInsts() && VT == MVT::v2f16);
|
2013-07-24 07:55:03 +08:00
|
|
|
}
|
|
|
|
|
2015-05-24 08:51:27 +08:00
|
|
|
bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT,
|
|
|
|
unsigned NumElem,
|
|
|
|
unsigned AS) const {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2015-10-13 07:59:50 +08:00
|
|
|
bool AMDGPUTargetLowering::aggressivelyPreferBuildVectorSources(EVT VecVT) const {
|
|
|
|
// There are few operations which truly have vector input operands. Any vector
|
|
|
|
// operation is going to involve operations on each component, and a
|
|
|
|
// build_vector will be a copy per element, so it always makes sense to use a
|
|
|
|
// build_vector input in place of the extracted element to avoid a copy into a
|
|
|
|
// super register.
|
|
|
|
//
|
|
|
|
// We should probably only do this if all users are extracts only, but this
|
|
|
|
// should be the common case.
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2014-02-12 18:17:54 +08:00
|
|
|
bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
|
2014-02-11 03:57:42 +08:00
|
|
|
// Truncate is just accessing a subregister.
|
2016-11-11 00:02:37 +08:00
|
|
|
|
|
|
|
unsigned SrcSize = Source.getSizeInBits();
|
|
|
|
unsigned DestSize = Dest.getSizeInBits();
|
|
|
|
|
|
|
|
return DestSize < SrcSize && DestSize % 32 == 0 ;
|
2014-02-12 18:17:54 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
|
|
|
|
// Truncate is just accessing a subregister.
|
2016-11-11 00:02:37 +08:00
|
|
|
|
|
|
|
unsigned SrcSize = Source->getScalarSizeInBits();
|
|
|
|
unsigned DestSize = Dest->getScalarSizeInBits();
|
|
|
|
|
|
|
|
if (DestSize== 16 && Subtarget->has16BitInsts())
|
|
|
|
return SrcSize >= 32;
|
|
|
|
|
|
|
|
return DestSize < SrcSize && DestSize % 32 == 0;
|
2014-02-11 03:57:42 +08:00
|
|
|
}
|
|
|
|
|
2014-03-28 01:23:31 +08:00
|
|
|
bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
|
2015-07-09 10:09:52 +08:00
|
|
|
unsigned SrcSize = Src->getScalarSizeInBits();
|
|
|
|
unsigned DestSize = Dest->getScalarSizeInBits();
|
2014-03-28 01:23:31 +08:00
|
|
|
|
2016-11-11 00:02:37 +08:00
|
|
|
if (SrcSize == 16 && Subtarget->has16BitInsts())
|
|
|
|
return DestSize >= 32;
|
|
|
|
|
2014-03-28 01:23:31 +08:00
|
|
|
return SrcSize == 32 && DestSize == 64;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
|
|
|
|
// Any register load of a 64-bit value really requires 2 32-bit moves. For all
|
|
|
|
// practical purposes, the extra mov 0 to load a 64-bit is free. As used,
|
|
|
|
// this will enable reducing 64-bit operations the 32-bit, which is always
|
|
|
|
// good.
|
2016-11-11 00:02:37 +08:00
|
|
|
|
|
|
|
if (Src == MVT::i16)
|
|
|
|
return Dest == MVT::i32 ||Dest == MVT::i64 ;
|
|
|
|
|
2014-03-28 01:23:31 +08:00
|
|
|
return Src == MVT::i32 && Dest == MVT::i64;
|
|
|
|
}
|
|
|
|
|
2014-06-26 21:45:47 +08:00
|
|
|
bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
|
|
|
|
return isZExtFree(Val.getValueType(), VT2);
|
|
|
|
}
|
|
|
|
|
2014-03-25 03:43:31 +08:00
|
|
|
bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
|
|
|
|
// There aren't really 64-bit registers, but pairs of 32-bit ones and only a
|
|
|
|
// limited number of native 64-bit operations. Shrinking an operation to fit
|
|
|
|
// in a single 32-bit register should always be helpful. As currently used,
|
|
|
|
// this is much less general than the name suggests, and is only used in
|
|
|
|
// places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
|
|
|
|
// not profitable, and may actually be harmful.
|
|
|
|
return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
|
|
|
|
}
|
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
//===---------------------------------------------------------------------===//
|
|
|
|
// TargetLowering Callbacks
|
|
|
|
//===---------------------------------------------------------------------===//
|
|
|
|
|
Re-commit AMDGPU/GlobalISel: Add support for simple shaders
Fix build when global-isel is disabled and fix a warning.
Summary: We can select constant/global G_LOAD, global G_STORE, and G_GEP.
Reviewers: qcolombet, MatzeB, t.p.northover, ab, arsenm
Subscribers: mehdi_amini, vkalintiris, kzhuravl, wdng, nhaehnle, mgorny, yaxunl, tony-tye, modocache, llvm-commits, dberris
Differential Revision: https://reviews.llvm.org/D26730
llvm-svn: 293551
2017-01-31 05:56:46 +08:00
|
|
|
CCAssignFn *AMDGPUCallLowering::CCAssignFnForCall(CallingConv::ID CC,
|
2017-05-18 05:56:25 +08:00
|
|
|
bool IsVarArg) {
|
|
|
|
switch (CC) {
|
|
|
|
case CallingConv::AMDGPU_KERNEL:
|
|
|
|
case CallingConv::SPIR_KERNEL:
|
|
|
|
return CC_AMDGPU_Kernel;
|
|
|
|
case CallingConv::AMDGPU_VS:
|
|
|
|
case CallingConv::AMDGPU_GS:
|
|
|
|
case CallingConv::AMDGPU_PS:
|
|
|
|
case CallingConv::AMDGPU_CS:
|
|
|
|
case CallingConv::AMDGPU_HS:
|
2017-09-29 17:51:22 +08:00
|
|
|
case CallingConv::AMDGPU_ES:
|
|
|
|
case CallingConv::AMDGPU_LS:
|
2017-05-18 05:56:25 +08:00
|
|
|
return CC_AMDGPU;
|
|
|
|
case CallingConv::C:
|
|
|
|
case CallingConv::Fast:
|
2017-09-12 02:54:20 +08:00
|
|
|
case CallingConv::Cold:
|
2017-05-18 05:56:25 +08:00
|
|
|
return CC_AMDGPU_Func;
|
|
|
|
default:
|
|
|
|
report_fatal_error("Unsupported calling convention.");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
CCAssignFn *AMDGPUCallLowering::CCAssignFnForReturn(CallingConv::ID CC,
|
|
|
|
bool IsVarArg) {
|
|
|
|
switch (CC) {
|
|
|
|
case CallingConv::AMDGPU_KERNEL:
|
|
|
|
case CallingConv::SPIR_KERNEL:
|
|
|
|
return CC_AMDGPU_Kernel;
|
|
|
|
case CallingConv::AMDGPU_VS:
|
|
|
|
case CallingConv::AMDGPU_GS:
|
|
|
|
case CallingConv::AMDGPU_PS:
|
|
|
|
case CallingConv::AMDGPU_CS:
|
|
|
|
case CallingConv::AMDGPU_HS:
|
2017-09-29 17:51:22 +08:00
|
|
|
case CallingConv::AMDGPU_ES:
|
|
|
|
case CallingConv::AMDGPU_LS:
|
2017-05-18 05:56:25 +08:00
|
|
|
return RetCC_SI_Shader;
|
|
|
|
case CallingConv::C:
|
|
|
|
case CallingConv::Fast:
|
2017-09-12 02:54:20 +08:00
|
|
|
case CallingConv::Cold:
|
2017-05-18 05:56:25 +08:00
|
|
|
return RetCC_AMDGPU_Func;
|
|
|
|
default:
|
|
|
|
report_fatal_error("Unsupported calling convention.");
|
|
|
|
}
|
Re-commit AMDGPU/GlobalISel: Add support for simple shaders
Fix build when global-isel is disabled and fix a warning.
Summary: We can select constant/global G_LOAD, global G_STORE, and G_GEP.
Reviewers: qcolombet, MatzeB, t.p.northover, ab, arsenm
Subscribers: mehdi_amini, vkalintiris, kzhuravl, wdng, nhaehnle, mgorny, yaxunl, tony-tye, modocache, llvm-commits, dberris
Differential Revision: https://reviews.llvm.org/D26730
llvm-svn: 293551
2017-01-31 05:56:46 +08:00
|
|
|
}
|
|
|
|
|
2016-09-17 05:53:00 +08:00
|
|
|
/// The SelectionDAGBuilder will automatically promote function arguments
|
|
|
|
/// with illegal types. However, this does not work for the AMDGPU targets
|
|
|
|
/// since the function arguments are stored in memory as these illegal types.
|
|
|
|
/// In order to handle this properly we need to get the original types sizes
|
|
|
|
/// from the LLVM IR Function and fixup the ISD:InputArg values before
|
|
|
|
/// passing them to AnalyzeFormalArguments()
|
|
|
|
|
|
|
|
/// When the SelectionDAGBuilder computes the Ins, it takes care of splitting
|
|
|
|
/// input values across multiple registers. Each item in the Ins array
|
2017-07-16 16:11:56 +08:00
|
|
|
/// represents a single value that will be stored in registers. Ins[x].VT is
|
2016-09-17 05:53:00 +08:00
|
|
|
/// the value type of the value that will be stored in the register, so
|
|
|
|
/// whatever SDNode we lower the argument to needs to be this type.
|
|
|
|
///
|
|
|
|
/// In order to correctly lower the arguments we need to know the size of each
|
|
|
|
/// argument. Since Ins[x].VT gives us the size of the register that will
|
|
|
|
/// hold the value, we need to look at Ins[x].ArgVT to see the 'real' type
|
|
|
|
/// for the orignal function argument so that we can deduce the correct memory
|
|
|
|
/// type to use for Ins[x]. In most cases the correct memory type will be
|
|
|
|
/// Ins[x].ArgVT. However, this will not always be the case. If, for example,
|
|
|
|
/// we have a kernel argument of type v8i8, this argument will be split into
|
|
|
|
/// 8 parts and each part will be represented by its own item in the Ins array.
|
|
|
|
/// For each part the Ins[x].ArgVT will be the v8i8, which is the full type of
|
|
|
|
/// the argument before it was split. From this, we deduce that the memory type
|
|
|
|
/// for each individual part is i8. We pass the memory type as LocVT to the
|
|
|
|
/// calling convention analysis function and the register type (Ins[x].VT) as
|
|
|
|
/// the ValVT.
|
|
|
|
void AMDGPUTargetLowering::analyzeFormalArgumentsCompute(CCState &State,
|
2013-03-07 17:03:52 +08:00
|
|
|
const SmallVectorImpl<ISD::InputArg> &Ins) const {
|
2016-09-17 05:53:00 +08:00
|
|
|
for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
|
|
|
|
const ISD::InputArg &In = Ins[i];
|
|
|
|
EVT MemVT;
|
|
|
|
|
|
|
|
unsigned NumRegs = getNumRegisters(State.getContext(), In.ArgVT);
|
|
|
|
|
2016-09-17 06:20:24 +08:00
|
|
|
if (!Subtarget->isAmdHsaOS() &&
|
|
|
|
(In.ArgVT == MVT::i16 || In.ArgVT == MVT::i8 || In.ArgVT == MVT::f16)) {
|
2016-09-17 05:53:00 +08:00
|
|
|
// The ABI says the caller will extend these values to 32-bits.
|
|
|
|
MemVT = In.ArgVT.isInteger() ? MVT::i32 : MVT::f32;
|
|
|
|
} else if (NumRegs == 1) {
|
|
|
|
// This argument is not split, so the IR type is the memory type.
|
|
|
|
assert(!In.Flags.isSplit());
|
|
|
|
if (In.ArgVT.isExtended()) {
|
|
|
|
// We have an extended type, like i24, so we should just use the register type
|
|
|
|
MemVT = In.VT;
|
|
|
|
} else {
|
|
|
|
MemVT = In.ArgVT;
|
|
|
|
}
|
|
|
|
} else if (In.ArgVT.isVector() && In.VT.isVector() &&
|
|
|
|
In.ArgVT.getScalarType() == In.VT.getScalarType()) {
|
|
|
|
assert(In.ArgVT.getVectorNumElements() > In.VT.getVectorNumElements());
|
|
|
|
// We have a vector value which has been split into a vector with
|
|
|
|
// the same scalar type, but fewer elements. This should handle
|
|
|
|
// all the floating-point vector types.
|
|
|
|
MemVT = In.VT;
|
|
|
|
} else if (In.ArgVT.isVector() &&
|
|
|
|
In.ArgVT.getVectorNumElements() == NumRegs) {
|
|
|
|
// This arg has been split so that each element is stored in a separate
|
|
|
|
// register.
|
|
|
|
MemVT = In.ArgVT.getScalarType();
|
|
|
|
} else if (In.ArgVT.isExtended()) {
|
|
|
|
// We have an extended type, like i65.
|
|
|
|
MemVT = In.VT;
|
|
|
|
} else {
|
|
|
|
unsigned MemoryBits = In.ArgVT.getStoreSizeInBits() / NumRegs;
|
|
|
|
assert(In.ArgVT.getStoreSizeInBits() % NumRegs == 0);
|
|
|
|
if (In.VT.isInteger()) {
|
|
|
|
MemVT = EVT::getIntegerVT(State.getContext(), MemoryBits);
|
|
|
|
} else if (In.VT.isVector()) {
|
|
|
|
assert(!In.VT.getScalarType().isFloatingPoint());
|
|
|
|
unsigned NumElements = In.VT.getVectorNumElements();
|
|
|
|
assert(MemoryBits % NumElements == 0);
|
|
|
|
// This vector type has been split into another vector type with
|
|
|
|
// a different elements size.
|
|
|
|
EVT ScalarVT = EVT::getIntegerVT(State.getContext(),
|
|
|
|
MemoryBits / NumElements);
|
|
|
|
MemVT = EVT::getVectorVT(State.getContext(), ScalarVT, NumElements);
|
|
|
|
} else {
|
|
|
|
llvm_unreachable("cannot deduce memory type.");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Convert one element vectors to scalar.
|
|
|
|
if (MemVT.isVector() && MemVT.getVectorNumElements() == 1)
|
|
|
|
MemVT = MemVT.getScalarType();
|
|
|
|
|
|
|
|
if (MemVT.isExtended()) {
|
|
|
|
// This should really only happen if we have vec3 arguments
|
|
|
|
assert(MemVT.isVector() && MemVT.getVectorNumElements() == 3);
|
|
|
|
MemVT = MemVT.getPow2VectorType(State.getContext());
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(MemVT.isSimple());
|
|
|
|
allocateKernArg(i, In.VT, MemVT.getSimpleVT(), CCValAssign::Full, In.Flags,
|
|
|
|
State);
|
|
|
|
}
|
|
|
|
}
|
2013-03-07 17:03:52 +08:00
|
|
|
|
2017-05-18 05:56:25 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerReturn(
|
|
|
|
SDValue Chain, CallingConv::ID CallConv,
|
|
|
|
bool isVarArg,
|
|
|
|
const SmallVectorImpl<ISD::OutputArg> &Outs,
|
|
|
|
const SmallVectorImpl<SDValue> &OutVals,
|
|
|
|
const SDLoc &DL, SelectionDAG &DAG) const {
|
|
|
|
// FIXME: Fails for r600 tests
|
|
|
|
//assert(!isVarArg && Outs.empty() && OutVals.empty() &&
|
|
|
|
// "wave terminate should not have return values");
|
2016-06-23 04:15:28 +08:00
|
|
|
return DAG.getNode(AMDGPUISD::ENDPGM, DL, MVT::Other, Chain);
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
//===---------------------------------------------------------------------===//
|
|
|
|
// Target specific lowering
|
|
|
|
//===---------------------------------------------------------------------===//
|
|
|
|
|
2017-04-12 06:29:24 +08:00
|
|
|
/// Selects the correct CCAssignFn for a given CallingConvention value.
|
|
|
|
CCAssignFn *AMDGPUTargetLowering::CCAssignFnForCall(CallingConv::ID CC,
|
|
|
|
bool IsVarArg) {
|
2017-05-18 05:56:25 +08:00
|
|
|
return AMDGPUCallLowering::CCAssignFnForCall(CC, IsVarArg);
|
|
|
|
}
|
|
|
|
|
|
|
|
CCAssignFn *AMDGPUTargetLowering::CCAssignFnForReturn(CallingConv::ID CC,
|
|
|
|
bool IsVarArg) {
|
|
|
|
return AMDGPUCallLowering::CCAssignFnForReturn(CC, IsVarArg);
|
2017-04-12 06:29:24 +08:00
|
|
|
}
|
|
|
|
|
2017-08-12 04:42:08 +08:00
|
|
|
SDValue AMDGPUTargetLowering::addTokenForArgument(SDValue Chain,
|
|
|
|
SelectionDAG &DAG,
|
|
|
|
MachineFrameInfo &MFI,
|
|
|
|
int ClobberedFI) const {
|
|
|
|
SmallVector<SDValue, 8> ArgChains;
|
|
|
|
int64_t FirstByte = MFI.getObjectOffset(ClobberedFI);
|
|
|
|
int64_t LastByte = FirstByte + MFI.getObjectSize(ClobberedFI) - 1;
|
|
|
|
|
|
|
|
// Include the original chain at the beginning of the list. When this is
|
|
|
|
// used by target LowerCall hooks, this helps legalize find the
|
|
|
|
// CALLSEQ_BEGIN node.
|
|
|
|
ArgChains.push_back(Chain);
|
|
|
|
|
|
|
|
// Add a chain value for each stack argument corresponding
|
|
|
|
for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
|
|
|
|
UE = DAG.getEntryNode().getNode()->use_end();
|
|
|
|
U != UE; ++U) {
|
|
|
|
if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U)) {
|
|
|
|
if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) {
|
|
|
|
if (FI->getIndex() < 0) {
|
|
|
|
int64_t InFirstByte = MFI.getObjectOffset(FI->getIndex());
|
|
|
|
int64_t InLastByte = InFirstByte;
|
|
|
|
InLastByte += MFI.getObjectSize(FI->getIndex()) - 1;
|
|
|
|
|
|
|
|
if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
|
|
|
|
(FirstByte <= InFirstByte && InFirstByte <= LastByte))
|
|
|
|
ArgChains.push_back(SDValue(L, 1));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Build a tokenfactor for all the chains.
|
|
|
|
return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
|
|
|
|
}
|
|
|
|
|
2017-08-04 07:32:41 +08:00
|
|
|
SDValue AMDGPUTargetLowering::lowerUnhandledCall(CallLoweringInfo &CLI,
|
|
|
|
SmallVectorImpl<SDValue> &InVals,
|
|
|
|
StringRef Reason) const {
|
2014-04-23 00:42:00 +08:00
|
|
|
SDValue Callee = CLI.Callee;
|
|
|
|
SelectionDAG &DAG = CLI.DAG;
|
|
|
|
|
2017-12-16 06:22:58 +08:00
|
|
|
const Function &Fn = DAG.getMachineFunction().getFunction();
|
2014-04-23 00:42:00 +08:00
|
|
|
|
|
|
|
StringRef FuncName("<unknown>");
|
|
|
|
|
2014-04-26 06:22:01 +08:00
|
|
|
if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
|
|
|
|
FuncName = G->getSymbol();
|
|
|
|
else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
|
2014-04-23 00:42:00 +08:00
|
|
|
FuncName = G->getGlobal()->getName();
|
|
|
|
|
2016-02-02 21:52:43 +08:00
|
|
|
DiagnosticInfoUnsupported NoCalls(
|
2017-08-04 07:32:41 +08:00
|
|
|
Fn, Reason + FuncName, CLI.DL.getDebugLoc());
|
2014-04-23 00:42:00 +08:00
|
|
|
DAG.getContext()->diagnose(NoCalls);
|
2016-05-19 00:10:11 +08:00
|
|
|
|
2016-12-16 04:50:12 +08:00
|
|
|
if (!CLI.IsTailCall) {
|
|
|
|
for (unsigned I = 0, E = CLI.Ins.size(); I != E; ++I)
|
|
|
|
InVals.push_back(DAG.getUNDEF(CLI.Ins[I].VT));
|
|
|
|
}
|
2016-05-19 00:10:11 +08:00
|
|
|
|
|
|
|
return DAG.getEntryNode();
|
2014-04-23 00:42:00 +08:00
|
|
|
}
|
|
|
|
|
2017-08-04 07:32:41 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
|
|
|
|
SmallVectorImpl<SDValue> &InVals) const {
|
|
|
|
return lowerUnhandledCall(CLI, InVals, "unsupported call to function ");
|
|
|
|
}
|
|
|
|
|
2015-08-27 02:37:13 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
2017-12-16 06:22:58 +08:00
|
|
|
const Function &Fn = DAG.getMachineFunction().getFunction();
|
2015-08-27 02:37:13 +08:00
|
|
|
|
2016-02-02 21:52:43 +08:00
|
|
|
DiagnosticInfoUnsupported NoDynamicAlloca(Fn, "unsupported dynamic alloca",
|
|
|
|
SDLoc(Op).getDebugLoc());
|
2015-08-27 02:37:13 +08:00
|
|
|
DAG.getContext()->diagnose(NoDynamicAlloca);
|
2016-06-23 17:19:16 +08:00
|
|
|
auto Ops = {DAG.getConstant(0, SDLoc(), Op.getValueType()), Op.getOperand(0)};
|
|
|
|
return DAG.getMergeValues(Ops, SDLoc());
|
2015-08-27 02:37:13 +08:00
|
|
|
}
|
|
|
|
|
2014-06-16 04:23:38 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
2012-12-12 05:25:42 +08:00
|
|
|
switch (Op.getOpcode()) {
|
|
|
|
default:
|
2017-01-28 10:02:38 +08:00
|
|
|
Op->print(errs(), &DAG);
|
2013-12-11 05:37:42 +08:00
|
|
|
llvm_unreachable("Custom lowering code for this"
|
|
|
|
"instruction is not implemented yet!");
|
2012-12-12 05:25:42 +08:00
|
|
|
break;
|
|
|
|
case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
|
2013-08-15 07:25:00 +08:00
|
|
|
case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
|
|
|
|
case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
|
2012-12-12 05:25:42 +08:00
|
|
|
case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
|
2014-06-23 05:43:00 +08:00
|
|
|
case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
|
2014-09-11 05:44:27 +08:00
|
|
|
case ISD::FREM: return LowerFREM(Op, DAG);
|
2014-06-19 01:05:30 +08:00
|
|
|
case ISD::FCEIL: return LowerFCEIL(Op, DAG);
|
|
|
|
case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
|
2014-06-19 01:05:26 +08:00
|
|
|
case ISD::FRINT: return LowerFRINT(Op, DAG);
|
2014-06-19 06:03:45 +08:00
|
|
|
case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
|
2015-01-22 02:18:25 +08:00
|
|
|
case ISD::FROUND: return LowerFROUND(Op, DAG);
|
2014-06-19 01:05:30 +08:00
|
|
|
case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
|
2017-11-27 21:26:38 +08:00
|
|
|
case ISD::FLOG:
|
|
|
|
return LowerFLOG(Op, DAG, 1 / AMDGPU_LOG2E_F);
|
|
|
|
case ISD::FLOG10:
|
|
|
|
return LowerFLOG(Op, DAG, AMDGPU_LN2_F / AMDGPU_LN10_F);
|
2014-10-04 07:54:41 +08:00
|
|
|
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
|
2013-10-31 01:22:05 +08:00
|
|
|
case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
|
2016-11-02 00:31:48 +08:00
|
|
|
case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG);
|
2014-10-04 07:54:56 +08:00
|
|
|
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
|
|
|
|
case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
|
2017-10-13 03:37:14 +08:00
|
|
|
case ISD::CTTZ:
|
|
|
|
case ISD::CTTZ_ZERO_UNDEF:
|
2016-01-12 00:50:29 +08:00
|
|
|
case ISD::CTLZ:
|
|
|
|
case ISD::CTLZ_ZERO_UNDEF:
|
2017-10-13 03:37:14 +08:00
|
|
|
return LowerCTLZ_CTTZ(Op, DAG);
|
2015-08-27 02:37:13 +08:00
|
|
|
case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
|
|
|
return Op;
|
|
|
|
}
|
|
|
|
|
2014-03-28 01:23:24 +08:00
|
|
|
void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
|
|
|
|
SmallVectorImpl<SDValue> &Results,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
switch (N->getOpcode()) {
|
|
|
|
case ISD::SIGN_EXTEND_INREG:
|
|
|
|
// Different parts of legalization seem to interpret which type of
|
|
|
|
// sign_extend_inreg is the one to check for custom lowering. The extended
|
|
|
|
// from type is what really matters, but some places check for custom
|
|
|
|
// lowering of the result type. This results in trying to use
|
|
|
|
// ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
|
|
|
|
// nothing here and let the illegal result integer be handled normally.
|
|
|
|
return;
|
|
|
|
default:
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-11-14 03:56:13 +08:00
|
|
|
static bool hasDefinedInitializer(const GlobalValue *GV) {
|
|
|
|
const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
|
|
|
|
if (!GVar || !GVar->hasInitializer())
|
|
|
|
return false;
|
|
|
|
|
2016-03-03 07:00:21 +08:00
|
|
|
return !isa<UndefValue>(GVar->getInitializer());
|
2014-11-14 03:56:13 +08:00
|
|
|
}
|
|
|
|
|
2013-06-28 23:47:08 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
|
|
|
|
SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
|
2015-07-09 10:09:04 +08:00
|
|
|
const DataLayout &DL = DAG.getDataLayout();
|
2013-06-28 23:47:08 +08:00
|
|
|
GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
|
2014-01-23 03:24:21 +08:00
|
|
|
const GlobalValue *GV = G->getGlobal();
|
2013-08-26 23:05:36 +08:00
|
|
|
|
2018-06-08 16:05:54 +08:00
|
|
|
if (G->getAddressSpace() == AMDGPUASI.LOCAL_ADDRESS ||
|
|
|
|
G->getAddressSpace() == AMDGPUASI.REGION_ADDRESS) {
|
|
|
|
if (!MFI->isEntryFunction()) {
|
|
|
|
const Function &Fn = DAG.getMachineFunction().getFunction();
|
|
|
|
DiagnosticInfoUnsupported BadLDSDecl(
|
|
|
|
Fn, "local memory global used by non-kernel function", SDLoc(Op).getDebugLoc());
|
|
|
|
DAG.getContext()->diagnose(BadLDSDecl);
|
|
|
|
}
|
|
|
|
|
2014-01-23 03:24:21 +08:00
|
|
|
// XXX: What does the value of G->getOffset() mean?
|
|
|
|
assert(G->getOffset() == 0 &&
|
2013-06-28 23:47:08 +08:00
|
|
|
"Do not know what to do with an non-zero offset");
|
|
|
|
|
2014-11-14 03:56:13 +08:00
|
|
|
// TODO: We could emit code to handle the initialization somewhere.
|
2017-03-27 22:04:01 +08:00
|
|
|
if (!hasDefinedInitializer(GV)) {
|
|
|
|
unsigned Offset = MFI->allocateLDSGlobal(DL, *GV);
|
|
|
|
return DAG.getConstant(Offset, SDLoc(Op), Op.getValueType());
|
|
|
|
}
|
2013-09-06 02:37:57 +08:00
|
|
|
}
|
2014-11-14 03:56:13 +08:00
|
|
|
|
2017-12-16 06:22:58 +08:00
|
|
|
const Function &Fn = DAG.getMachineFunction().getFunction();
|
2016-02-02 21:52:43 +08:00
|
|
|
DiagnosticInfoUnsupported BadInit(
|
|
|
|
Fn, "unsupported initializer for address space", SDLoc(Op).getDebugLoc());
|
2014-11-14 03:56:13 +08:00
|
|
|
DAG.getContext()->diagnose(BadInit);
|
|
|
|
return SDValue();
|
2013-06-28 23:47:08 +08:00
|
|
|
}
|
|
|
|
|
2013-08-15 07:25:00 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
SmallVector<SDValue, 8> Args;
|
|
|
|
|
2018-06-15 23:15:46 +08:00
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
if (VT == MVT::v4i16 || VT == MVT::v4f16) {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Lo = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(0));
|
|
|
|
SDValue Hi = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(1));
|
|
|
|
|
|
|
|
SDValue BV = DAG.getBuildVector(MVT::v2i32, SL, { Lo, Hi });
|
|
|
|
return DAG.getNode(ISD::BITCAST, SL, VT, BV);
|
|
|
|
}
|
|
|
|
|
2015-04-24 06:59:24 +08:00
|
|
|
for (const SDUse &U : Op->ops())
|
|
|
|
DAG.ExtractVectorElements(U.get(), Args);
|
2013-08-15 07:25:00 +08:00
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
|
2013-08-15 07:25:00 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
|
|
|
|
SmallVector<SDValue, 8> Args;
|
|
|
|
unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
|
2014-04-12 01:47:30 +08:00
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
|
|
|
|
VT.getVectorNumElements());
|
2013-08-15 07:25:00 +08:00
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
|
2013-08-15 07:25:00 +08:00
|
|
|
}
|
|
|
|
|
2018-05-01 23:54:18 +08:00
|
|
|
/// Generate Min/Max node
|
2017-02-01 08:42:40 +08:00
|
|
|
SDValue AMDGPUTargetLowering::combineFMinMaxLegacy(const SDLoc &DL, EVT VT,
|
2016-06-12 23:39:02 +08:00
|
|
|
SDValue LHS, SDValue RHS,
|
|
|
|
SDValue True, SDValue False,
|
2014-12-12 10:30:37 +08:00
|
|
|
SDValue CC,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
2014-11-14 07:03:09 +08:00
|
|
|
if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
|
2012-12-12 05:25:42 +08:00
|
|
|
return SDValue();
|
|
|
|
|
2014-12-12 10:30:37 +08:00
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
2012-12-12 05:25:42 +08:00
|
|
|
ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
|
|
|
|
switch (CCOpcode) {
|
|
|
|
case ISD::SETOEQ:
|
|
|
|
case ISD::SETONE:
|
|
|
|
case ISD::SETUNE:
|
|
|
|
case ISD::SETNE:
|
|
|
|
case ISD::SETUEQ:
|
|
|
|
case ISD::SETEQ:
|
|
|
|
case ISD::SETFALSE:
|
|
|
|
case ISD::SETFALSE2:
|
|
|
|
case ISD::SETTRUE:
|
|
|
|
case ISD::SETTRUE2:
|
|
|
|
case ISD::SETUO:
|
|
|
|
case ISD::SETO:
|
2014-11-14 07:03:09 +08:00
|
|
|
break;
|
2012-12-12 05:25:42 +08:00
|
|
|
case ISD::SETULE:
|
2014-12-12 10:30:37 +08:00
|
|
|
case ISD::SETULT: {
|
|
|
|
if (LHS == True)
|
|
|
|
return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
|
|
|
|
return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
|
|
|
|
}
|
2012-12-12 05:25:42 +08:00
|
|
|
case ISD::SETOLE:
|
|
|
|
case ISD::SETOLT:
|
|
|
|
case ISD::SETLE:
|
|
|
|
case ISD::SETLT: {
|
2014-12-12 10:30:37 +08:00
|
|
|
// Ordered. Assume ordered for undefined.
|
|
|
|
|
|
|
|
// Only do this after legalization to avoid interfering with other combines
|
|
|
|
// which might occur.
|
|
|
|
if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
|
|
|
|
!DCI.isCalledByLegalizer())
|
|
|
|
return SDValue();
|
2014-12-07 20:19:03 +08:00
|
|
|
|
2014-11-15 13:02:57 +08:00
|
|
|
// We need to permute the operands to get the correct NaN behavior. The
|
|
|
|
// selected operand is the second one based on the failing compare with NaN,
|
|
|
|
// so permute it based on the compare type the hardware uses.
|
|
|
|
if (LHS == True)
|
2014-12-12 10:30:37 +08:00
|
|
|
return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
|
|
|
|
return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
|
|
|
|
}
|
|
|
|
case ISD::SETUGE:
|
|
|
|
case ISD::SETUGT: {
|
|
|
|
if (LHS == True)
|
|
|
|
return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
|
|
|
|
return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
|
|
|
case ISD::SETGT:
|
|
|
|
case ISD::SETGE:
|
|
|
|
case ISD::SETOGE:
|
|
|
|
case ISD::SETOGT: {
|
2014-12-12 10:30:37 +08:00
|
|
|
if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
|
|
|
|
!DCI.isCalledByLegalizer())
|
|
|
|
return SDValue();
|
2014-12-07 20:19:03 +08:00
|
|
|
|
2014-11-15 13:02:57 +08:00
|
|
|
if (LHS == True)
|
2014-12-12 10:30:37 +08:00
|
|
|
return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
|
|
|
|
return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
|
|
|
case ISD::SETCC_INVALID:
|
2013-12-11 05:37:42 +08:00
|
|
|
llvm_unreachable("Invalid setcc condcode!");
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
2014-05-10 00:42:16 +08:00
|
|
|
return SDValue();
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
|
|
|
|
2016-01-19 06:01:13 +08:00
|
|
|
std::pair<SDValue, SDValue>
|
|
|
|
AMDGPUTargetLowering::split64BitValue(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
|
|
|
|
SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
|
|
|
|
|
|
|
|
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
|
|
|
|
const SDValue One = DAG.getConstant(1, SL, MVT::i32);
|
|
|
|
|
|
|
|
SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
|
|
|
|
SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
|
|
|
|
|
|
|
|
return std::make_pair(Lo, Hi);
|
|
|
|
}
|
|
|
|
|
2016-01-19 06:09:04 +08:00
|
|
|
SDValue AMDGPUTargetLowering::getLoHalf64(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
|
|
|
|
SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
|
|
|
|
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
|
|
|
|
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::getHiHalf64(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
|
|
|
|
SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
|
|
|
|
const SDValue One = DAG.getConstant(1, SL, MVT::i32);
|
|
|
|
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
|
|
|
|
}
|
|
|
|
|
2014-07-25 01:10:35 +08:00
|
|
|
SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
2016-04-15 07:31:26 +08:00
|
|
|
LoadSDNode *Load = cast<LoadSDNode>(Op);
|
2014-07-25 01:10:35 +08:00
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
|
2016-04-15 07:31:26 +08:00
|
|
|
|
2014-07-25 01:10:35 +08:00
|
|
|
// If this is a 2 element vector, we really want to scalarize and not create
|
|
|
|
// weird 1 element vectors.
|
|
|
|
if (VT.getVectorNumElements() == 2)
|
2016-04-15 07:31:26 +08:00
|
|
|
return scalarizeVectorLoad(Load, DAG);
|
2014-07-25 01:10:35 +08:00
|
|
|
|
|
|
|
SDValue BasePtr = Load->getBasePtr();
|
|
|
|
EVT MemVT = Load->getMemoryVT();
|
|
|
|
SDLoc SL(Op);
|
2015-12-15 00:59:40 +08:00
|
|
|
|
|
|
|
const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
|
2014-07-25 01:10:35 +08:00
|
|
|
|
|
|
|
EVT LoVT, HiVT;
|
|
|
|
EVT LoMemVT, HiMemVT;
|
|
|
|
SDValue Lo, Hi;
|
|
|
|
|
|
|
|
std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
|
|
|
|
std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
|
|
|
|
std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
|
2015-12-15 00:59:40 +08:00
|
|
|
|
|
|
|
unsigned Size = LoMemVT.getStoreSize();
|
|
|
|
unsigned BaseAlign = Load->getAlignment();
|
|
|
|
unsigned HiAlign = MinAlign(BaseAlign, Size);
|
|
|
|
|
[SelectionDAG] Get rid of bool parameters in SelectionDAG::getLoad, getStore, and friends.
Summary:
Instead, we take a single flags arg (a bitset).
Also add a default 0 alignment, and change the order of arguments so the
alignment comes before the flags.
This greatly simplifies many callsites, and fixes a bug in
AMDGPUISelLowering, wherein the order of the args to getLoad was
inverted. It also greatly simplifies the process of adding another flag
to getLoad.
Reviewers: chandlerc, tstellarAMD
Subscribers: jholewinski, arsenm, jyknight, dsanders, nemanjai, llvm-commits
Differential Revision: http://reviews.llvm.org/D22249
llvm-svn: 275592
2016-07-16 02:27:10 +08:00
|
|
|
SDValue LoLoad = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
|
|
|
|
Load->getChain(), BasePtr, SrcValue, LoMemVT,
|
|
|
|
BaseAlign, Load->getMemOperand()->getFlags());
|
2017-11-29 09:25:12 +08:00
|
|
|
SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, Size);
|
[SelectionDAG] Get rid of bool parameters in SelectionDAG::getLoad, getStore, and friends.
Summary:
Instead, we take a single flags arg (a bitset).
Also add a default 0 alignment, and change the order of arguments so the
alignment comes before the flags.
This greatly simplifies many callsites, and fixes a bug in
AMDGPUISelLowering, wherein the order of the args to getLoad was
inverted. It also greatly simplifies the process of adding another flag
to getLoad.
Reviewers: chandlerc, tstellarAMD
Subscribers: jholewinski, arsenm, jyknight, dsanders, nemanjai, llvm-commits
Differential Revision: http://reviews.llvm.org/D22249
llvm-svn: 275592
2016-07-16 02:27:10 +08:00
|
|
|
SDValue HiLoad =
|
|
|
|
DAG.getExtLoad(Load->getExtensionType(), SL, HiVT, Load->getChain(),
|
|
|
|
HiPtr, SrcValue.getWithOffset(LoMemVT.getStoreSize()),
|
|
|
|
HiMemVT, HiAlign, Load->getMemOperand()->getFlags());
|
2014-07-25 01:10:35 +08:00
|
|
|
|
|
|
|
SDValue Ops[] = {
|
|
|
|
DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
|
|
|
|
DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
|
|
|
|
LoLoad.getValue(1), HiLoad.getValue(1))
|
|
|
|
};
|
|
|
|
|
|
|
|
return DAG.getMergeValues(Ops, SL);
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
StoreSDNode *Store = cast<StoreSDNode>(Op);
|
|
|
|
SDValue Val = Store->getValue();
|
|
|
|
EVT VT = Val.getValueType();
|
|
|
|
|
|
|
|
// If this is a 2 element vector, we really want to scalarize and not create
|
|
|
|
// weird 1 element vectors.
|
|
|
|
if (VT.getVectorNumElements() == 2)
|
2016-04-15 07:31:26 +08:00
|
|
|
return scalarizeVectorStore(Store, DAG);
|
2014-07-25 01:10:35 +08:00
|
|
|
|
|
|
|
EVT MemVT = Store->getMemoryVT();
|
|
|
|
SDValue Chain = Store->getChain();
|
|
|
|
SDValue BasePtr = Store->getBasePtr();
|
|
|
|
SDLoc SL(Op);
|
|
|
|
|
|
|
|
EVT LoVT, HiVT;
|
|
|
|
EVT LoMemVT, HiMemVT;
|
|
|
|
SDValue Lo, Hi;
|
|
|
|
|
|
|
|
std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
|
|
|
|
std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
|
|
|
|
std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
|
|
|
|
|
2017-11-29 09:25:12 +08:00
|
|
|
SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, LoMemVT.getStoreSize());
|
2014-07-25 01:10:35 +08:00
|
|
|
|
2015-12-15 00:59:40 +08:00
|
|
|
const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo();
|
|
|
|
unsigned BaseAlign = Store->getAlignment();
|
|
|
|
unsigned Size = LoMemVT.getStoreSize();
|
|
|
|
unsigned HiAlign = MinAlign(BaseAlign, Size);
|
|
|
|
|
[SelectionDAG] Get rid of bool parameters in SelectionDAG::getLoad, getStore, and friends.
Summary:
Instead, we take a single flags arg (a bitset).
Also add a default 0 alignment, and change the order of arguments so the
alignment comes before the flags.
This greatly simplifies many callsites, and fixes a bug in
AMDGPUISelLowering, wherein the order of the args to getLoad was
inverted. It also greatly simplifies the process of adding another flag
to getLoad.
Reviewers: chandlerc, tstellarAMD
Subscribers: jholewinski, arsenm, jyknight, dsanders, nemanjai, llvm-commits
Differential Revision: http://reviews.llvm.org/D22249
llvm-svn: 275592
2016-07-16 02:27:10 +08:00
|
|
|
SDValue LoStore =
|
|
|
|
DAG.getTruncStore(Chain, SL, Lo, BasePtr, SrcValue, LoMemVT, BaseAlign,
|
|
|
|
Store->getMemOperand()->getFlags());
|
|
|
|
SDValue HiStore =
|
|
|
|
DAG.getTruncStore(Chain, SL, Hi, HiPtr, SrcValue.getWithOffset(Size),
|
|
|
|
HiMemVT, HiAlign, Store->getMemOperand()->getFlags());
|
2014-07-25 01:10:35 +08:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
|
|
|
|
}
|
|
|
|
|
2014-07-24 14:59:20 +08:00
|
|
|
// This is a shortcut for integer division because we have fast i32<->f32
|
|
|
|
// conversions, and fast f32 reciprocal instructions. The fractional part of a
|
2016-05-21 09:53:33 +08:00
|
|
|
// float is enough to accurately represent up to a 24-bit signed integer.
|
2016-05-20 05:09:58 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG,
|
|
|
|
bool Sign) const {
|
2014-06-16 04:08:02 +08:00
|
|
|
SDLoc DL(Op);
|
2014-07-24 14:59:20 +08:00
|
|
|
EVT VT = Op.getValueType();
|
2014-06-16 04:08:02 +08:00
|
|
|
SDValue LHS = Op.getOperand(0);
|
|
|
|
SDValue RHS = Op.getOperand(1);
|
2014-07-24 14:59:20 +08:00
|
|
|
MVT IntVT = MVT::i32;
|
|
|
|
MVT FltVT = MVT::f32;
|
|
|
|
|
2016-05-21 09:53:33 +08:00
|
|
|
unsigned LHSSignBits = DAG.ComputeNumSignBits(LHS);
|
|
|
|
if (LHSSignBits < 9)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
unsigned RHSSignBits = DAG.ComputeNumSignBits(RHS);
|
|
|
|
if (RHSSignBits < 9)
|
|
|
|
return SDValue();
|
2014-08-13 01:31:20 +08:00
|
|
|
|
2016-05-20 05:09:58 +08:00
|
|
|
unsigned BitSize = VT.getSizeInBits();
|
2016-05-21 09:53:33 +08:00
|
|
|
unsigned SignBits = std::min(LHSSignBits, RHSSignBits);
|
|
|
|
unsigned DivBits = BitSize - SignBits;
|
|
|
|
if (Sign)
|
|
|
|
++DivBits;
|
|
|
|
|
|
|
|
ISD::NodeType ToFp = Sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
|
|
|
|
ISD::NodeType ToInt = Sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
|
2014-07-24 14:59:20 +08:00
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue jq = DAG.getConstant(1, DL, IntVT);
|
2014-08-13 01:31:20 +08:00
|
|
|
|
2016-05-20 05:09:58 +08:00
|
|
|
if (Sign) {
|
2014-08-13 01:31:20 +08:00
|
|
|
// char|short jq = ia ^ ib;
|
|
|
|
jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
2014-08-13 01:31:20 +08:00
|
|
|
// jq = jq >> (bitsize - 2)
|
2015-04-28 22:05:47 +08:00
|
|
|
jq = DAG.getNode(ISD::SRA, DL, VT, jq,
|
|
|
|
DAG.getConstant(BitSize - 2, DL, VT));
|
2014-06-16 04:08:02 +08:00
|
|
|
|
2014-08-13 01:31:20 +08:00
|
|
|
// jq = jq | 0x1
|
2015-04-28 22:05:47 +08:00
|
|
|
jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
|
2014-08-13 01:31:20 +08:00
|
|
|
}
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// int ia = (int)LHS;
|
2016-05-20 05:09:58 +08:00
|
|
|
SDValue ia = LHS;
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// int ib, (int)RHS;
|
2016-05-20 05:09:58 +08:00
|
|
|
SDValue ib = RHS;
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// float fa = (float)ia;
|
2014-08-13 01:31:20 +08:00
|
|
|
SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// float fb = (float)ib;
|
2014-08-13 01:31:20 +08:00
|
|
|
SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
2014-07-24 14:59:20 +08:00
|
|
|
SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
|
|
|
|
fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// fq = trunc(fq);
|
2014-07-24 14:59:20 +08:00
|
|
|
fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// float fqneg = -fq;
|
2014-07-24 14:59:20 +08:00
|
|
|
SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// float fr = mad(fqneg, fb, fa);
|
2017-03-08 08:48:46 +08:00
|
|
|
unsigned OpCode = Subtarget->hasFP32Denormals() ?
|
|
|
|
(unsigned)AMDGPUISD::FMAD_FTZ :
|
2017-02-25 07:00:29 +08:00
|
|
|
(unsigned)ISD::FMAD;
|
|
|
|
SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// int iq = (int)fq;
|
2014-08-13 01:31:20 +08:00
|
|
|
SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// fr = fabs(fr);
|
2014-07-24 14:59:20 +08:00
|
|
|
fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// fb = fabs(fb);
|
2014-07-24 14:59:20 +08:00
|
|
|
fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
|
|
|
|
|
2015-07-09 10:09:04 +08:00
|
|
|
EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
|
2014-06-16 04:08:02 +08:00
|
|
|
|
|
|
|
// int cv = fr >= fb;
|
2014-07-24 14:59:20 +08:00
|
|
|
SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
|
|
|
|
|
2014-06-16 04:08:02 +08:00
|
|
|
// jq = (cv ? jq : 0);
|
2015-04-28 22:05:47 +08:00
|
|
|
jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));
|
2014-07-24 14:59:20 +08:00
|
|
|
|
2014-08-13 01:31:20 +08:00
|
|
|
// dst = iq + jq;
|
2014-08-13 01:31:17 +08:00
|
|
|
SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
|
|
|
|
|
2014-08-13 01:31:20 +08:00
|
|
|
// Rem needs compensation, it's easier to recompute it
|
2014-08-13 01:31:17 +08:00
|
|
|
SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
|
|
|
|
Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
|
|
|
|
|
2016-05-21 09:53:33 +08:00
|
|
|
// Truncate to number of bits this divide really is.
|
|
|
|
if (Sign) {
|
|
|
|
SDValue InRegSize
|
|
|
|
= DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), DivBits));
|
|
|
|
Div = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Div, InRegSize);
|
|
|
|
Rem = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Rem, InRegSize);
|
|
|
|
} else {
|
|
|
|
SDValue TruncMask = DAG.getConstant((UINT64_C(1) << DivBits) - 1, DL, VT);
|
|
|
|
Div = DAG.getNode(ISD::AND, DL, VT, Div, TruncMask);
|
|
|
|
Rem = DAG.getNode(ISD::AND, DL, VT, Rem, TruncMask);
|
|
|
|
}
|
|
|
|
|
2016-05-20 05:09:58 +08:00
|
|
|
return DAG.getMergeValues({ Div, Rem }, DL);
|
2014-06-16 04:08:02 +08:00
|
|
|
}
|
|
|
|
|
2014-11-15 09:07:53 +08:00
|
|
|
void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
|
|
|
|
SelectionDAG &DAG,
|
|
|
|
SmallVectorImpl<SDValue> &Results) const {
|
|
|
|
SDLoc DL(Op);
|
|
|
|
EVT VT = Op.getValueType();
|
2017-10-07 01:24:45 +08:00
|
|
|
|
|
|
|
assert(VT == MVT::i64 && "LowerUDIVREM64 expects an i64");
|
|
|
|
|
2014-11-15 09:07:53 +08:00
|
|
|
EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
|
|
|
|
|
2017-10-07 01:24:45 +08:00
|
|
|
SDValue One = DAG.getConstant(1, DL, HalfVT);
|
|
|
|
SDValue Zero = DAG.getConstant(0, DL, HalfVT);
|
2014-11-15 09:07:53 +08:00
|
|
|
|
|
|
|
//HiLo split
|
|
|
|
SDValue LHS = Op.getOperand(0);
|
2017-10-07 01:24:45 +08:00
|
|
|
SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
|
|
|
|
SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, One);
|
2014-11-15 09:07:53 +08:00
|
|
|
|
|
|
|
SDValue RHS = Op.getOperand(1);
|
2017-10-07 01:24:45 +08:00
|
|
|
SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
|
|
|
|
SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, One);
|
2014-11-15 09:07:53 +08:00
|
|
|
|
2017-10-07 01:24:45 +08:00
|
|
|
if (DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
|
|
|
|
DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
|
2015-01-23 07:42:43 +08:00
|
|
|
|
|
|
|
SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
|
|
|
|
LHS_Lo, RHS_Lo);
|
|
|
|
|
2017-10-07 01:24:45 +08:00
|
|
|
SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(0), Zero});
|
|
|
|
SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(1), Zero});
|
2016-03-01 13:06:05 +08:00
|
|
|
|
|
|
|
Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV));
|
|
|
|
Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM));
|
2015-01-23 07:42:43 +08:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2017-10-07 01:24:45 +08:00
|
|
|
if (isTypeLegal(MVT::i64)) {
|
|
|
|
// Compute denominator reciprocal.
|
|
|
|
unsigned FMAD = Subtarget->hasFP32Denormals() ?
|
|
|
|
(unsigned)AMDGPUISD::FMAD_FTZ :
|
|
|
|
(unsigned)ISD::FMAD;
|
|
|
|
|
|
|
|
SDValue Cvt_Lo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Lo);
|
|
|
|
SDValue Cvt_Hi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Hi);
|
|
|
|
SDValue Mad1 = DAG.getNode(FMAD, DL, MVT::f32, Cvt_Hi,
|
|
|
|
DAG.getConstantFP(APInt(32, 0x4f800000).bitsToFloat(), DL, MVT::f32),
|
|
|
|
Cvt_Lo);
|
|
|
|
SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, DL, MVT::f32, Mad1);
|
|
|
|
SDValue Mul1 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Rcp,
|
|
|
|
DAG.getConstantFP(APInt(32, 0x5f7ffffc).bitsToFloat(), DL, MVT::f32));
|
|
|
|
SDValue Mul2 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Mul1,
|
|
|
|
DAG.getConstantFP(APInt(32, 0x2f800000).bitsToFloat(), DL, MVT::f32));
|
|
|
|
SDValue Trunc = DAG.getNode(ISD::FTRUNC, DL, MVT::f32, Mul2);
|
|
|
|
SDValue Mad2 = DAG.getNode(FMAD, DL, MVT::f32, Trunc,
|
|
|
|
DAG.getConstantFP(APInt(32, 0xcf800000).bitsToFloat(), DL, MVT::f32),
|
|
|
|
Mul1);
|
|
|
|
SDValue Rcp_Lo = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Mad2);
|
|
|
|
SDValue Rcp_Hi = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Trunc);
|
|
|
|
SDValue Rcp64 = DAG.getBitcast(VT,
|
|
|
|
DAG.getBuildVector(MVT::v2i32, DL, {Rcp_Lo, Rcp_Hi}));
|
|
|
|
|
|
|
|
SDValue Zero64 = DAG.getConstant(0, DL, VT);
|
|
|
|
SDValue One64 = DAG.getConstant(1, DL, VT);
|
|
|
|
SDValue Zero1 = DAG.getConstant(0, DL, MVT::i1);
|
|
|
|
SDVTList HalfCarryVT = DAG.getVTList(HalfVT, MVT::i1);
|
|
|
|
|
|
|
|
SDValue Neg_RHS = DAG.getNode(ISD::SUB, DL, VT, Zero64, RHS);
|
|
|
|
SDValue Mullo1 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Rcp64);
|
|
|
|
SDValue Mulhi1 = DAG.getNode(ISD::MULHU, DL, VT, Rcp64, Mullo1);
|
|
|
|
SDValue Mulhi1_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1,
|
|
|
|
Zero);
|
|
|
|
SDValue Mulhi1_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1,
|
|
|
|
One);
|
|
|
|
|
|
|
|
SDValue Add1_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Lo,
|
|
|
|
Mulhi1_Lo, Zero1);
|
|
|
|
SDValue Add1_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Hi,
|
|
|
|
Mulhi1_Hi, Add1_Lo.getValue(1));
|
|
|
|
SDValue Add1_HiNc = DAG.getNode(ISD::ADD, DL, HalfVT, Rcp_Hi, Mulhi1_Hi);
|
|
|
|
SDValue Add1 = DAG.getBitcast(VT,
|
|
|
|
DAG.getBuildVector(MVT::v2i32, DL, {Add1_Lo, Add1_Hi}));
|
|
|
|
|
|
|
|
SDValue Mullo2 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Add1);
|
|
|
|
SDValue Mulhi2 = DAG.getNode(ISD::MULHU, DL, VT, Add1, Mullo2);
|
|
|
|
SDValue Mulhi2_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2,
|
|
|
|
Zero);
|
|
|
|
SDValue Mulhi2_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2,
|
|
|
|
One);
|
|
|
|
|
|
|
|
SDValue Add2_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Lo,
|
|
|
|
Mulhi2_Lo, Zero1);
|
|
|
|
SDValue Add2_HiC = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_HiNc,
|
|
|
|
Mulhi2_Hi, Add1_Lo.getValue(1));
|
|
|
|
SDValue Add2_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add2_HiC,
|
|
|
|
Zero, Add2_Lo.getValue(1));
|
|
|
|
SDValue Add2 = DAG.getBitcast(VT,
|
|
|
|
DAG.getBuildVector(MVT::v2i32, DL, {Add2_Lo, Add2_Hi}));
|
|
|
|
SDValue Mulhi3 = DAG.getNode(ISD::MULHU, DL, VT, LHS, Add2);
|
|
|
|
|
|
|
|
SDValue Mul3 = DAG.getNode(ISD::MUL, DL, VT, RHS, Mulhi3);
|
|
|
|
|
|
|
|
SDValue Mul3_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, Zero);
|
|
|
|
SDValue Mul3_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, One);
|
|
|
|
SDValue Sub1_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Lo,
|
|
|
|
Mul3_Lo, Zero1);
|
|
|
|
SDValue Sub1_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Hi,
|
|
|
|
Mul3_Hi, Sub1_Lo.getValue(1));
|
|
|
|
SDValue Sub1_Mi = DAG.getNode(ISD::SUB, DL, HalfVT, LHS_Hi, Mul3_Hi);
|
|
|
|
SDValue Sub1 = DAG.getBitcast(VT,
|
|
|
|
DAG.getBuildVector(MVT::v2i32, DL, {Sub1_Lo, Sub1_Hi}));
|
|
|
|
|
|
|
|
SDValue MinusOne = DAG.getConstant(0xffffffffu, DL, HalfVT);
|
|
|
|
SDValue C1 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, MinusOne, Zero,
|
|
|
|
ISD::SETUGE);
|
|
|
|
SDValue C2 = DAG.getSelectCC(DL, Sub1_Lo, RHS_Lo, MinusOne, Zero,
|
|
|
|
ISD::SETUGE);
|
|
|
|
SDValue C3 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, C2, C1, ISD::SETEQ);
|
|
|
|
|
|
|
|
// TODO: Here and below portions of the code can be enclosed into if/endif.
|
|
|
|
// Currently control flow is unconditional and we have 4 selects after
|
|
|
|
// potential endif to substitute PHIs.
|
|
|
|
|
|
|
|
// if C3 != 0 ...
|
|
|
|
SDValue Sub2_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Lo,
|
|
|
|
RHS_Lo, Zero1);
|
|
|
|
SDValue Sub2_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Mi,
|
|
|
|
RHS_Hi, Sub1_Lo.getValue(1));
|
|
|
|
SDValue Sub2_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
|
|
|
|
Zero, Sub2_Lo.getValue(1));
|
|
|
|
SDValue Sub2 = DAG.getBitcast(VT,
|
|
|
|
DAG.getBuildVector(MVT::v2i32, DL, {Sub2_Lo, Sub2_Hi}));
|
|
|
|
|
|
|
|
SDValue Add3 = DAG.getNode(ISD::ADD, DL, VT, Mulhi3, One64);
|
|
|
|
|
|
|
|
SDValue C4 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, MinusOne, Zero,
|
|
|
|
ISD::SETUGE);
|
|
|
|
SDValue C5 = DAG.getSelectCC(DL, Sub2_Lo, RHS_Lo, MinusOne, Zero,
|
|
|
|
ISD::SETUGE);
|
|
|
|
SDValue C6 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, C5, C4, ISD::SETEQ);
|
|
|
|
|
|
|
|
// if (C6 != 0)
|
|
|
|
SDValue Add4 = DAG.getNode(ISD::ADD, DL, VT, Add3, One64);
|
|
|
|
|
|
|
|
SDValue Sub3_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Lo,
|
|
|
|
RHS_Lo, Zero1);
|
|
|
|
SDValue Sub3_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
|
|
|
|
RHS_Hi, Sub2_Lo.getValue(1));
|
|
|
|
SDValue Sub3_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub3_Mi,
|
|
|
|
Zero, Sub3_Lo.getValue(1));
|
|
|
|
SDValue Sub3 = DAG.getBitcast(VT,
|
|
|
|
DAG.getBuildVector(MVT::v2i32, DL, {Sub3_Lo, Sub3_Hi}));
|
|
|
|
|
|
|
|
// endif C6
|
|
|
|
// endif C3
|
|
|
|
|
|
|
|
SDValue Sel1 = DAG.getSelectCC(DL, C6, Zero, Add4, Add3, ISD::SETNE);
|
|
|
|
SDValue Div = DAG.getSelectCC(DL, C3, Zero, Sel1, Mulhi3, ISD::SETNE);
|
|
|
|
|
|
|
|
SDValue Sel2 = DAG.getSelectCC(DL, C6, Zero, Sub3, Sub2, ISD::SETNE);
|
|
|
|
SDValue Rem = DAG.getSelectCC(DL, C3, Zero, Sel2, Sub1, ISD::SETNE);
|
|
|
|
|
|
|
|
Results.push_back(Div);
|
|
|
|
Results.push_back(Rem);
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// r600 expandion.
|
2014-11-15 09:07:53 +08:00
|
|
|
// Get Speculative values
|
|
|
|
SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
|
|
|
|
SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
|
|
|
|
|
2017-10-07 01:24:45 +08:00
|
|
|
SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, Zero, REM_Part, LHS_Hi, ISD::SETEQ);
|
|
|
|
SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {REM_Lo, Zero});
|
2016-03-01 13:06:05 +08:00
|
|
|
REM = DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM);
|
2014-11-15 09:07:53 +08:00
|
|
|
|
2017-10-07 01:24:45 +08:00
|
|
|
SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, Zero, DIV_Part, Zero, ISD::SETEQ);
|
|
|
|
SDValue DIV_Lo = Zero;
|
2014-11-15 09:07:53 +08:00
|
|
|
|
|
|
|
const unsigned halfBitWidth = HalfVT.getSizeInBits();
|
|
|
|
|
|
|
|
for (unsigned i = 0; i < halfBitWidth; ++i) {
|
2015-01-23 07:42:39 +08:00
|
|
|
const unsigned bitPos = halfBitWidth - i - 1;
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
|
2015-01-23 07:42:39 +08:00
|
|
|
// Get value of high bit
|
2015-04-13 07:45:01 +08:00
|
|
|
SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
|
2017-10-07 01:24:45 +08:00
|
|
|
HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, One);
|
2015-01-23 07:42:39 +08:00
|
|
|
HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
|
2014-11-15 09:07:53 +08:00
|
|
|
|
2015-01-23 07:42:39 +08:00
|
|
|
// Shift
|
2015-04-28 22:05:47 +08:00
|
|
|
REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
|
2015-01-23 07:42:39 +08:00
|
|
|
// Add LHS high bit
|
|
|
|
REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
|
2014-11-15 09:07:53 +08:00
|
|
|
|
2016-03-31 05:30:00 +08:00
|
|
|
SDValue BIT = DAG.getConstant(1ULL << bitPos, DL, HalfVT);
|
2017-10-07 01:24:45 +08:00
|
|
|
SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, Zero, ISD::SETUGE);
|
2014-11-15 09:07:53 +08:00
|
|
|
|
|
|
|
DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
|
|
|
|
|
|
|
|
// Update REM
|
|
|
|
SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
|
2014-11-15 09:07:57 +08:00
|
|
|
REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
|
2014-11-15 09:07:53 +08:00
|
|
|
}
|
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {DIV_Lo, DIV_Hi});
|
2016-03-01 13:06:05 +08:00
|
|
|
DIV = DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV);
|
2014-11-15 09:07:53 +08:00
|
|
|
Results.push_back(DIV);
|
|
|
|
Results.push_back(REM);
|
|
|
|
}
|
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
|
2014-05-12 05:24:41 +08:00
|
|
|
SelectionDAG &DAG) const {
|
2013-05-25 10:42:55 +08:00
|
|
|
SDLoc DL(Op);
|
2012-12-12 05:25:42 +08:00
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
|
2014-11-15 09:07:53 +08:00
|
|
|
if (VT == MVT::i64) {
|
|
|
|
SmallVector<SDValue, 2> Results;
|
|
|
|
LowerUDIVREM64(Op, DAG, Results);
|
|
|
|
return DAG.getMergeValues(Results, DL);
|
|
|
|
}
|
|
|
|
|
2014-08-13 01:31:20 +08:00
|
|
|
if (VT == MVT::i32) {
|
2016-05-21 09:53:33 +08:00
|
|
|
if (SDValue Res = LowerDIVREM24(Op, DAG, false))
|
|
|
|
return Res;
|
2014-08-13 01:31:20 +08:00
|
|
|
}
|
|
|
|
|
2016-05-21 09:53:33 +08:00
|
|
|
SDValue Num = Op.getOperand(0);
|
|
|
|
SDValue Den = Op.getOperand(1);
|
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
// RCP = URECIP(Den) = 2^32 / Den + e
|
|
|
|
// e is rounding error.
|
|
|
|
SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
|
|
|
|
|
2014-09-22 23:35:30 +08:00
|
|
|
// RCP_LO = mul(RCP, Den) */
|
|
|
|
SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
|
2012-12-12 05:25:42 +08:00
|
|
|
|
|
|
|
// RCP_HI = mulhu (RCP, Den) */
|
|
|
|
SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
|
|
|
|
|
|
|
|
// NEG_RCP_LO = -RCP_LO
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
|
2012-12-12 05:25:42 +08:00
|
|
|
RCP_LO);
|
|
|
|
|
|
|
|
// ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
|
2012-12-12 05:25:42 +08:00
|
|
|
NEG_RCP_LO, RCP_LO,
|
|
|
|
ISD::SETEQ);
|
|
|
|
// Calculate the rounding error from the URECIP instruction
|
|
|
|
// E = mulhu(ABS_RCP_LO, RCP)
|
|
|
|
SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
|
|
|
|
|
|
|
|
// RCP_A_E = RCP + E
|
|
|
|
SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
|
|
|
|
|
|
|
|
// RCP_S_E = RCP - E
|
|
|
|
SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
|
|
|
|
|
|
|
|
// Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
|
2012-12-12 05:25:42 +08:00
|
|
|
RCP_A_E, RCP_S_E,
|
|
|
|
ISD::SETEQ);
|
|
|
|
// Quotient = mulhu(Tmp0, Num)
|
|
|
|
SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
|
|
|
|
|
|
|
|
// Num_S_Remainder = Quotient * Den
|
2014-09-22 23:35:30 +08:00
|
|
|
SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
|
2012-12-12 05:25:42 +08:00
|
|
|
|
|
|
|
// Remainder = Num - Num_S_Remainder
|
|
|
|
SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
|
|
|
|
|
|
|
|
// Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
|
|
|
|
SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(-1, DL, VT),
|
|
|
|
DAG.getConstant(0, DL, VT),
|
2013-11-07 01:36:04 +08:00
|
|
|
ISD::SETUGE);
|
|
|
|
// Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
|
|
|
|
SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
|
|
|
|
Num_S_Remainder,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(-1, DL, VT),
|
|
|
|
DAG.getConstant(0, DL, VT),
|
2013-11-07 01:36:04 +08:00
|
|
|
ISD::SETUGE);
|
2012-12-12 05:25:42 +08:00
|
|
|
// Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
|
|
|
|
SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
|
|
|
|
Remainder_GE_Zero);
|
|
|
|
|
|
|
|
// Calculate Division result:
|
|
|
|
|
|
|
|
// Quotient_A_One = Quotient + 1
|
|
|
|
SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(1, DL, VT));
|
2012-12-12 05:25:42 +08:00
|
|
|
|
|
|
|
// Quotient_S_One = Quotient - 1
|
|
|
|
SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(1, DL, VT));
|
2012-12-12 05:25:42 +08:00
|
|
|
|
|
|
|
// Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
|
2012-12-12 05:25:42 +08:00
|
|
|
Quotient, Quotient_A_One, ISD::SETEQ);
|
|
|
|
|
|
|
|
// Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
|
2015-04-28 22:05:47 +08:00
|
|
|
Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
|
2012-12-12 05:25:42 +08:00
|
|
|
Quotient_S_One, Div, ISD::SETEQ);
|
|
|
|
|
|
|
|
// Calculate Rem result:
|
|
|
|
|
|
|
|
// Remainder_S_Den = Remainder - Den
|
|
|
|
SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
|
|
|
|
|
|
|
|
// Remainder_A_Den = Remainder + Den
|
|
|
|
SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
|
|
|
|
|
|
|
|
// Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
|
2012-12-12 05:25:42 +08:00
|
|
|
Remainder, Remainder_S_Den, ISD::SETEQ);
|
|
|
|
|
|
|
|
// Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
|
2015-04-28 22:05:47 +08:00
|
|
|
Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
|
2012-12-12 05:25:42 +08:00
|
|
|
Remainder_A_Den, Rem, ISD::SETEQ);
|
2014-04-08 00:44:24 +08:00
|
|
|
SDValue Ops[2] = {
|
|
|
|
Div,
|
|
|
|
Rem
|
|
|
|
};
|
2014-04-28 03:20:57 +08:00
|
|
|
return DAG.getMergeValues(Ops, DL);
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
|
|
|
|
2014-06-23 05:43:00 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
SDLoc DL(Op);
|
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
|
|
|
|
SDValue LHS = Op.getOperand(0);
|
|
|
|
SDValue RHS = Op.getOperand(1);
|
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue Zero = DAG.getConstant(0, DL, VT);
|
|
|
|
SDValue NegOne = DAG.getConstant(-1, DL, VT);
|
2014-08-13 01:31:17 +08:00
|
|
|
|
2016-05-21 09:53:33 +08:00
|
|
|
if (VT == MVT::i32) {
|
|
|
|
if (SDValue Res = LowerDIVREM24(Op, DAG, true))
|
|
|
|
return Res;
|
2015-01-23 07:42:43 +08:00
|
|
|
}
|
2016-05-21 09:53:33 +08:00
|
|
|
|
2015-01-23 07:42:43 +08:00
|
|
|
if (VT == MVT::i64 &&
|
|
|
|
DAG.ComputeNumSignBits(LHS) > 32 &&
|
|
|
|
DAG.ComputeNumSignBits(RHS) > 32) {
|
|
|
|
EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
|
|
|
|
|
|
|
|
//HiLo split
|
|
|
|
SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
|
|
|
|
SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
|
|
|
|
SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
|
|
|
|
LHS_Lo, RHS_Lo);
|
|
|
|
SDValue Res[2] = {
|
|
|
|
DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
|
|
|
|
DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
|
|
|
|
};
|
|
|
|
return DAG.getMergeValues(Res, DL);
|
|
|
|
}
|
|
|
|
|
2014-06-23 05:43:00 +08:00
|
|
|
SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
|
|
|
|
SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
|
|
|
|
SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
|
|
|
|
SDValue RSign = LHSign; // Remainder sign is the same as LHS
|
|
|
|
|
|
|
|
LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
|
|
|
|
RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
|
|
|
|
|
|
|
|
LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
|
|
|
|
RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
|
|
|
|
|
|
|
|
SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
|
|
|
|
SDValue Rem = Div.getValue(1);
|
|
|
|
|
|
|
|
Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
|
|
|
|
Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
|
|
|
|
|
|
|
|
Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
|
|
|
|
Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
|
|
|
|
|
|
|
|
SDValue Res[2] = {
|
|
|
|
Div,
|
|
|
|
Rem
|
|
|
|
};
|
|
|
|
return DAG.getMergeValues(Res, DL);
|
|
|
|
}
|
|
|
|
|
2014-09-11 05:44:27 +08:00
|
|
|
// (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
|
|
|
|
SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
SDValue X = Op.getOperand(0);
|
|
|
|
SDValue Y = Op.getOperand(1);
|
|
|
|
|
2015-09-17 00:31:21 +08:00
|
|
|
// TODO: Should this propagate fast-math-flags?
|
|
|
|
|
2014-09-11 05:44:27 +08:00
|
|
|
SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
|
|
|
|
SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
|
|
|
|
SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
|
|
|
|
|
|
|
|
return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
|
|
|
|
}
|
|
|
|
|
2014-06-19 01:05:30 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
|
|
|
// result = trunc(src)
|
|
|
|
// if (src > 0.0 && src != result)
|
|
|
|
// result += 1.0
|
|
|
|
|
|
|
|
SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
|
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
|
|
|
|
const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
2015-07-09 10:09:04 +08:00
|
|
|
EVT SetCCVT =
|
|
|
|
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
|
|
|
SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
|
|
|
|
SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
|
|
|
|
SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
|
|
|
|
|
|
|
|
SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
|
2015-09-17 00:31:21 +08:00
|
|
|
// TODO: Should this propagate fast-math-flags?
|
2014-06-19 01:05:30 +08:00
|
|
|
return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
|
|
|
|
}
|
|
|
|
|
2016-06-12 23:39:02 +08:00
|
|
|
static SDValue extractF64Exponent(SDValue Hi, const SDLoc &SL,
|
|
|
|
SelectionDAG &DAG) {
|
2015-01-22 02:18:25 +08:00
|
|
|
const unsigned FractBits = 52;
|
|
|
|
const unsigned ExpBits = 11;
|
|
|
|
|
|
|
|
SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
|
|
|
|
Hi,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(FractBits - 32, SL, MVT::i32),
|
|
|
|
DAG.getConstant(ExpBits, SL, MVT::i32));
|
2015-01-22 02:18:25 +08:00
|
|
|
SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(1023, SL, MVT::i32));
|
2015-01-22 02:18:25 +08:00
|
|
|
|
|
|
|
return Exp;
|
|
|
|
}
|
|
|
|
|
2014-06-19 01:05:30 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
|
|
|
assert(Op.getValueType() == MVT::f64);
|
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
|
|
|
|
const SDValue One = DAG.getConstant(1, SL, MVT::i32);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
|
|
|
SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
|
|
|
|
|
|
|
|
// Extract the upper half, since this is where we will find the sign and
|
|
|
|
// exponent.
|
|
|
|
SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
|
|
|
|
|
2015-01-22 02:18:25 +08:00
|
|
|
SDValue Exp = extractF64Exponent(Hi, SL, DAG);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
2015-01-22 02:18:25 +08:00
|
|
|
const unsigned FractBits = 52;
|
2014-06-19 01:05:30 +08:00
|
|
|
|
|
|
|
// Extract the sign bit.
|
2015-04-28 22:05:47 +08:00
|
|
|
const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, SL, MVT::i32);
|
2014-06-19 01:05:30 +08:00
|
|
|
SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
|
|
|
|
|
2018-01-29 13:17:03 +08:00
|
|
|
// Extend back to 64-bits.
|
2016-04-27 05:15:30 +08:00
|
|
|
SDValue SignBit64 = DAG.getBuildVector(MVT::v2i32, SL, {Zero, SignBit});
|
2014-06-19 01:05:30 +08:00
|
|
|
SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
|
|
|
|
|
|
|
|
SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
|
2014-06-19 06:11:03 +08:00
|
|
|
const SDValue FractMask
|
2015-04-28 22:05:47 +08:00
|
|
|
= DAG.getConstant((UINT64_C(1) << FractBits) - 1, SL, MVT::i64);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
|
|
|
SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
|
|
|
|
SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
|
|
|
|
SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
|
|
|
|
|
2015-07-09 10:09:04 +08:00
|
|
|
EVT SetCCVT =
|
|
|
|
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
|
|
|
SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
|
|
|
|
SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
|
|
|
|
|
|
|
|
SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
|
|
|
|
SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
|
|
|
|
|
|
|
|
return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
|
|
|
|
}
|
|
|
|
|
2014-06-19 01:05:26 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
|
|
|
assert(Op.getValueType() == MVT::f64);
|
|
|
|
|
2016-12-14 19:57:17 +08:00
|
|
|
APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64);
|
2014-06-19 01:05:26 +08:00
|
|
|
SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
|
|
|
|
|
2015-09-17 00:31:21 +08:00
|
|
|
// TODO: Should this propagate fast-math-flags?
|
|
|
|
|
2014-06-19 01:05:26 +08:00
|
|
|
SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
|
|
|
|
SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
|
|
|
|
|
|
|
|
SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
|
2014-06-19 01:45:58 +08:00
|
|
|
|
2016-12-14 19:57:17 +08:00
|
|
|
APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64);
|
2014-06-19 01:05:26 +08:00
|
|
|
|
2015-07-09 10:09:04 +08:00
|
|
|
EVT SetCCVT =
|
|
|
|
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
|
2014-06-19 01:05:26 +08:00
|
|
|
SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
|
|
|
|
|
|
|
|
return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
|
|
|
|
}
|
|
|
|
|
2014-06-19 06:03:45 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
// FNEARBYINT and FRINT are the same, except in their handling of FP
|
|
|
|
// exceptions. Those aren't really meaningful for us, and OpenCL only has
|
|
|
|
// rint, so just treat them as equivalent.
|
|
|
|
return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
|
|
|
|
}
|
|
|
|
|
2015-01-22 02:18:25 +08:00
|
|
|
// XXX - May require not supporting f32 denormals?
|
2017-03-25 04:04:18 +08:00
|
|
|
|
|
|
|
// Don't handle v2f16. The extra instructions to scalarize and repack around the
|
|
|
|
// compare and vselect end up producing worse code than scalarizing the whole
|
|
|
|
// operation.
|
|
|
|
SDValue AMDGPUTargetLowering::LowerFROUND32_16(SDValue Op, SelectionDAG &DAG) const {
|
2015-01-22 02:18:25 +08:00
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue X = Op.getOperand(0);
|
2017-03-25 04:04:18 +08:00
|
|
|
EVT VT = Op.getValueType();
|
2015-01-22 02:18:25 +08:00
|
|
|
|
2017-03-25 04:04:18 +08:00
|
|
|
SDValue T = DAG.getNode(ISD::FTRUNC, SL, VT, X);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
2015-09-17 00:31:21 +08:00
|
|
|
// TODO: Should this propagate fast-math-flags?
|
|
|
|
|
2017-03-25 04:04:18 +08:00
|
|
|
SDValue Diff = DAG.getNode(ISD::FSUB, SL, VT, X, T);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
2017-03-25 04:04:18 +08:00
|
|
|
SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, VT, Diff);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
2017-03-25 04:04:18 +08:00
|
|
|
const SDValue Zero = DAG.getConstantFP(0.0, SL, VT);
|
|
|
|
const SDValue One = DAG.getConstantFP(1.0, SL, VT);
|
|
|
|
const SDValue Half = DAG.getConstantFP(0.5, SL, VT);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
2017-03-25 04:04:18 +08:00
|
|
|
SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, VT, One, X);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
2015-07-09 10:09:04 +08:00
|
|
|
EVT SetCCVT =
|
2017-03-25 04:04:18 +08:00
|
|
|
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
|
|
|
SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
|
|
|
|
|
2017-03-25 04:04:18 +08:00
|
|
|
SDValue Sel = DAG.getNode(ISD::SELECT, SL, VT, Cmp, SignOne, Zero);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
2017-03-25 04:04:18 +08:00
|
|
|
return DAG.getNode(ISD::FADD, SL, VT, T, Sel);
|
2015-01-22 02:18:25 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::LowerFROUND64(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue X = Op.getOperand(0);
|
|
|
|
|
|
|
|
SDValue L = DAG.getNode(ISD::BITCAST, SL, MVT::i64, X);
|
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
|
|
|
|
const SDValue One = DAG.getConstant(1, SL, MVT::i32);
|
|
|
|
const SDValue NegOne = DAG.getConstant(-1, SL, MVT::i32);
|
|
|
|
const SDValue FiftyOne = DAG.getConstant(51, SL, MVT::i32);
|
2015-07-09 10:09:04 +08:00
|
|
|
EVT SetCCVT =
|
|
|
|
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
|
|
|
SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
|
|
|
|
|
|
|
|
SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC, One);
|
|
|
|
|
|
|
|
SDValue Exp = extractF64Exponent(Hi, SL, DAG);
|
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
const SDValue Mask = DAG.getConstant(INT64_C(0x000fffffffffffff), SL,
|
|
|
|
MVT::i64);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
|
|
|
SDValue M = DAG.getNode(ISD::SRA, SL, MVT::i64, Mask, Exp);
|
|
|
|
SDValue D = DAG.getNode(ISD::SRA, SL, MVT::i64,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(INT64_C(0x0008000000000000), SL,
|
|
|
|
MVT::i64),
|
2015-01-22 02:18:25 +08:00
|
|
|
Exp);
|
|
|
|
|
|
|
|
SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, L, M);
|
|
|
|
SDValue Tmp1 = DAG.getSetCC(SL, SetCCVT,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(0, SL, MVT::i64), Tmp0,
|
2015-01-22 02:18:25 +08:00
|
|
|
ISD::SETNE);
|
|
|
|
|
|
|
|
SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, Tmp1,
|
2015-04-28 22:05:47 +08:00
|
|
|
D, DAG.getConstant(0, SL, MVT::i64));
|
2015-01-22 02:18:25 +08:00
|
|
|
SDValue K = DAG.getNode(ISD::ADD, SL, MVT::i64, L, Tmp2);
|
|
|
|
|
|
|
|
K = DAG.getNode(ISD::AND, SL, MVT::i64, K, DAG.getNOT(SL, M, MVT::i64));
|
|
|
|
K = DAG.getNode(ISD::BITCAST, SL, MVT::f64, K);
|
|
|
|
|
|
|
|
SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
|
|
|
|
SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
|
|
|
|
SDValue ExpEqNegOne = DAG.getSetCC(SL, SetCCVT, NegOne, Exp, ISD::SETEQ);
|
|
|
|
|
|
|
|
SDValue Mag = DAG.getNode(ISD::SELECT, SL, MVT::f64,
|
|
|
|
ExpEqNegOne,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstantFP(1.0, SL, MVT::f64),
|
|
|
|
DAG.getConstantFP(0.0, SL, MVT::f64));
|
2015-01-22 02:18:25 +08:00
|
|
|
|
|
|
|
SDValue S = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, Mag, X);
|
|
|
|
|
|
|
|
K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpLt0, S, K);
|
|
|
|
K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpGt51, X, K);
|
|
|
|
|
|
|
|
return K;
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
|
2017-03-25 04:04:18 +08:00
|
|
|
if (VT == MVT::f32 || VT == MVT::f16)
|
|
|
|
return LowerFROUND32_16(Op, DAG);
|
2015-01-22 02:18:25 +08:00
|
|
|
|
|
|
|
if (VT == MVT::f64)
|
|
|
|
return LowerFROUND64(Op, DAG);
|
|
|
|
|
|
|
|
llvm_unreachable("unhandled type");
|
|
|
|
}
|
|
|
|
|
2014-06-19 01:05:30 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
|
|
|
// result = trunc(src);
|
|
|
|
// if (src < 0.0 && src != result)
|
|
|
|
// result += -1.0.
|
|
|
|
|
|
|
|
SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
|
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
|
|
|
|
const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
2015-07-09 10:09:04 +08:00
|
|
|
EVT SetCCVT =
|
|
|
|
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
|
2014-06-19 01:05:30 +08:00
|
|
|
|
|
|
|
SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
|
|
|
|
SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
|
|
|
|
SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
|
|
|
|
|
|
|
|
SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
|
2015-09-17 00:31:21 +08:00
|
|
|
// TODO: Should this propagate fast-math-flags?
|
2014-06-19 01:05:30 +08:00
|
|
|
return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
|
|
|
|
}
|
|
|
|
|
2017-11-27 21:26:38 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFLOG(SDValue Op, SelectionDAG &DAG,
|
|
|
|
double Log2BaseInverted) const {
|
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Operand = Op.getOperand(0);
|
|
|
|
SDValue Log2Operand = DAG.getNode(ISD::FLOG2, SL, VT, Operand);
|
|
|
|
SDValue Log2BaseInvertedOperand = DAG.getConstantFP(Log2BaseInverted, SL, VT);
|
|
|
|
|
|
|
|
return DAG.getNode(ISD::FMUL, SL, VT, Log2Operand, Log2BaseInvertedOperand);
|
|
|
|
}
|
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
static bool isCtlzOpc(unsigned Opc) {
|
|
|
|
return Opc == ISD::CTLZ || Opc == ISD::CTLZ_ZERO_UNDEF;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool isCttzOpc(unsigned Opc) {
|
|
|
|
return Opc == ISD::CTTZ || Opc == ISD::CTTZ_ZERO_UNDEF;
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const {
|
2016-01-12 00:50:29 +08:00
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
2017-10-13 03:37:14 +08:00
|
|
|
bool ZeroUndef = Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF ||
|
|
|
|
Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF;
|
|
|
|
|
|
|
|
unsigned ISDOpc, NewOpc;
|
|
|
|
if (isCtlzOpc(Op.getOpcode())) {
|
|
|
|
ISDOpc = ISD::CTLZ_ZERO_UNDEF;
|
|
|
|
NewOpc = AMDGPUISD::FFBH_U32;
|
|
|
|
} else if (isCttzOpc(Op.getOpcode())) {
|
|
|
|
ISDOpc = ISD::CTTZ_ZERO_UNDEF;
|
|
|
|
NewOpc = AMDGPUISD::FFBL_B32;
|
|
|
|
} else
|
|
|
|
llvm_unreachable("Unexpected OPCode!!!");
|
|
|
|
|
2016-01-12 01:02:00 +08:00
|
|
|
|
|
|
|
if (ZeroUndef && Src.getValueType() == MVT::i32)
|
2017-10-13 03:37:14 +08:00
|
|
|
return DAG.getNode(NewOpc, SL, MVT::i32, Src);
|
2016-01-12 01:02:00 +08:00
|
|
|
|
2016-01-12 00:50:29 +08:00
|
|
|
SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
|
|
|
|
|
|
|
|
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
|
|
|
|
const SDValue One = DAG.getConstant(1, SL, MVT::i32);
|
|
|
|
|
|
|
|
SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
|
|
|
|
SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
|
|
|
|
|
|
|
|
EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(),
|
|
|
|
*DAG.getContext(), MVT::i32);
|
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
SDValue HiOrLo = isCtlzOpc(Op.getOpcode()) ? Hi : Lo;
|
2017-10-18 05:49:52 +08:00
|
|
|
SDValue Hi0orLo0 = DAG.getSetCC(SL, SetCCVT, HiOrLo, Zero, ISD::SETEQ);
|
2016-01-12 00:50:29 +08:00
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
SDValue OprLo = DAG.getNode(ISDOpc, SL, MVT::i32, Lo);
|
|
|
|
SDValue OprHi = DAG.getNode(ISDOpc, SL, MVT::i32, Hi);
|
2016-01-12 00:50:29 +08:00
|
|
|
|
|
|
|
const SDValue Bits32 = DAG.getConstant(32, SL, MVT::i32);
|
2017-10-13 03:37:14 +08:00
|
|
|
SDValue Add, NewOpr;
|
|
|
|
if (isCtlzOpc(Op.getOpcode())) {
|
|
|
|
Add = DAG.getNode(ISD::ADD, SL, MVT::i32, OprLo, Bits32);
|
|
|
|
// ctlz(x) = hi_32(x) == 0 ? ctlz(lo_32(x)) + 32 : ctlz(hi_32(x))
|
|
|
|
NewOpr = DAG.getNode(ISD::SELECT, SL, MVT::i32, Hi0orLo0, Add, OprHi);
|
|
|
|
} else {
|
|
|
|
Add = DAG.getNode(ISD::ADD, SL, MVT::i32, OprHi, Bits32);
|
|
|
|
// cttz(x) = lo_32(x) == 0 ? cttz(hi_32(x)) + 32 : cttz(lo_32(x))
|
|
|
|
NewOpr = DAG.getNode(ISD::SELECT, SL, MVT::i32, Hi0orLo0, Add, OprLo);
|
|
|
|
}
|
2016-01-12 00:50:29 +08:00
|
|
|
|
|
|
|
if (!ZeroUndef) {
|
|
|
|
// Test if the full 64-bit input is zero.
|
|
|
|
|
|
|
|
// FIXME: DAG combines turn what should be an s_and_b64 into a v_or_b32,
|
|
|
|
// which we probably don't want.
|
2017-10-13 03:37:14 +08:00
|
|
|
SDValue LoOrHi = isCtlzOpc(Op.getOpcode()) ? Lo : Hi;
|
2017-10-18 05:49:52 +08:00
|
|
|
SDValue Lo0OrHi0 = DAG.getSetCC(SL, SetCCVT, LoOrHi, Zero, ISD::SETEQ);
|
2017-10-13 03:37:14 +08:00
|
|
|
SDValue SrcIsZero = DAG.getNode(ISD::AND, SL, SetCCVT, Lo0OrHi0, Hi0orLo0);
|
2016-01-12 00:50:29 +08:00
|
|
|
|
|
|
|
// TODO: If i64 setcc is half rate, it can result in 1 fewer instruction
|
|
|
|
// with the same cycles, otherwise it is slower.
|
|
|
|
// SDValue SrcIsZero = DAG.getSetCC(SL, SetCCVT, Src,
|
|
|
|
// DAG.getConstant(0, SL, MVT::i64), ISD::SETEQ);
|
|
|
|
|
|
|
|
const SDValue Bits32 = DAG.getConstant(64, SL, MVT::i32);
|
|
|
|
|
|
|
|
// The instruction returns -1 for 0 input, but the defined intrinsic
|
|
|
|
// behavior is to return the number of bits.
|
2017-10-13 03:37:14 +08:00
|
|
|
NewOpr = DAG.getNode(ISD::SELECT, SL, MVT::i32,
|
|
|
|
SrcIsZero, Bits32, NewOpr);
|
2016-01-12 00:50:29 +08:00
|
|
|
}
|
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
return DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i64, NewOpr);
|
2016-01-12 00:50:29 +08:00
|
|
|
}
|
|
|
|
|
2016-01-12 06:01:48 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG,
|
|
|
|
bool Signed) const {
|
|
|
|
// Unsigned
|
|
|
|
// cul2f(ulong u)
|
|
|
|
//{
|
|
|
|
// uint lz = clz(u);
|
|
|
|
// uint e = (u != 0) ? 127U + 63U - lz : 0;
|
|
|
|
// u = (u << lz) & 0x7fffffffffffffffUL;
|
|
|
|
// ulong t = u & 0xffffffffffUL;
|
|
|
|
// uint v = (e << 23) | (uint)(u >> 40);
|
|
|
|
// uint r = t > 0x8000000000UL ? 1U : (t == 0x8000000000UL ? v & 1U : 0U);
|
|
|
|
// return as_float(v + r);
|
|
|
|
//}
|
|
|
|
// Signed
|
|
|
|
// cl2f(long l)
|
|
|
|
//{
|
|
|
|
// long s = l >> 63;
|
|
|
|
// float r = cul2f((l + s) ^ s);
|
|
|
|
// return s ? -r : r;
|
|
|
|
//}
|
|
|
|
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
SDValue L = Src;
|
|
|
|
|
|
|
|
SDValue S;
|
|
|
|
if (Signed) {
|
|
|
|
const SDValue SignBit = DAG.getConstant(63, SL, MVT::i64);
|
|
|
|
S = DAG.getNode(ISD::SRA, SL, MVT::i64, L, SignBit);
|
|
|
|
|
|
|
|
SDValue LPlusS = DAG.getNode(ISD::ADD, SL, MVT::i64, L, S);
|
|
|
|
L = DAG.getNode(ISD::XOR, SL, MVT::i64, LPlusS, S);
|
|
|
|
}
|
|
|
|
|
|
|
|
EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(),
|
|
|
|
*DAG.getContext(), MVT::f32);
|
|
|
|
|
|
|
|
|
|
|
|
SDValue ZeroI32 = DAG.getConstant(0, SL, MVT::i32);
|
|
|
|
SDValue ZeroI64 = DAG.getConstant(0, SL, MVT::i64);
|
|
|
|
SDValue LZ = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SL, MVT::i64, L);
|
|
|
|
LZ = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LZ);
|
|
|
|
|
|
|
|
SDValue K = DAG.getConstant(127U + 63U, SL, MVT::i32);
|
|
|
|
SDValue E = DAG.getSelect(SL, MVT::i32,
|
|
|
|
DAG.getSetCC(SL, SetCCVT, L, ZeroI64, ISD::SETNE),
|
|
|
|
DAG.getNode(ISD::SUB, SL, MVT::i32, K, LZ),
|
|
|
|
ZeroI32);
|
|
|
|
|
|
|
|
SDValue U = DAG.getNode(ISD::AND, SL, MVT::i64,
|
|
|
|
DAG.getNode(ISD::SHL, SL, MVT::i64, L, LZ),
|
|
|
|
DAG.getConstant((-1ULL) >> 1, SL, MVT::i64));
|
|
|
|
|
|
|
|
SDValue T = DAG.getNode(ISD::AND, SL, MVT::i64, U,
|
|
|
|
DAG.getConstant(0xffffffffffULL, SL, MVT::i64));
|
|
|
|
|
|
|
|
SDValue UShl = DAG.getNode(ISD::SRL, SL, MVT::i64,
|
|
|
|
U, DAG.getConstant(40, SL, MVT::i64));
|
|
|
|
|
|
|
|
SDValue V = DAG.getNode(ISD::OR, SL, MVT::i32,
|
|
|
|
DAG.getNode(ISD::SHL, SL, MVT::i32, E, DAG.getConstant(23, SL, MVT::i32)),
|
|
|
|
DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, UShl));
|
|
|
|
|
|
|
|
SDValue C = DAG.getConstant(0x8000000000ULL, SL, MVT::i64);
|
|
|
|
SDValue RCmp = DAG.getSetCC(SL, SetCCVT, T, C, ISD::SETUGT);
|
|
|
|
SDValue TCmp = DAG.getSetCC(SL, SetCCVT, T, C, ISD::SETEQ);
|
|
|
|
|
|
|
|
SDValue One = DAG.getConstant(1, SL, MVT::i32);
|
|
|
|
|
|
|
|
SDValue VTrunc1 = DAG.getNode(ISD::AND, SL, MVT::i32, V, One);
|
|
|
|
|
|
|
|
SDValue R = DAG.getSelect(SL, MVT::i32,
|
|
|
|
RCmp,
|
|
|
|
One,
|
|
|
|
DAG.getSelect(SL, MVT::i32, TCmp, VTrunc1, ZeroI32));
|
|
|
|
R = DAG.getNode(ISD::ADD, SL, MVT::i32, V, R);
|
|
|
|
R = DAG.getNode(ISD::BITCAST, SL, MVT::f32, R);
|
|
|
|
|
|
|
|
if (!Signed)
|
|
|
|
return R;
|
|
|
|
|
|
|
|
SDValue RNeg = DAG.getNode(ISD::FNEG, SL, MVT::f32, R);
|
|
|
|
return DAG.getSelect(SL, MVT::f32, DAG.getSExtOrTrunc(S, SL, SetCCVT), RNeg, R);
|
|
|
|
}
|
|
|
|
|
2014-10-04 07:54:41 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
|
|
|
|
bool Signed) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
|
|
|
SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
|
|
|
|
|
|
|
|
SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(0, SL, MVT::i32));
|
2014-10-04 07:54:41 +08:00
|
|
|
SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(1, SL, MVT::i32));
|
2014-10-04 07:54:41 +08:00
|
|
|
|
|
|
|
SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
|
|
|
|
SL, MVT::f64, Hi);
|
|
|
|
|
|
|
|
SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
|
|
|
|
|
|
|
|
SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
|
2015-04-28 22:05:47 +08:00
|
|
|
DAG.getConstant(32, SL, MVT::i32));
|
2015-09-17 00:31:21 +08:00
|
|
|
// TODO: Should this propagate fast-math-flags?
|
2014-10-04 07:54:41 +08:00
|
|
|
return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
|
|
|
|
}
|
|
|
|
|
2013-10-31 01:22:05 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
2016-01-12 06:01:48 +08:00
|
|
|
assert(Op.getOperand(0).getValueType() == MVT::i64 &&
|
|
|
|
"operation should be legal");
|
2013-10-31 01:22:05 +08:00
|
|
|
|
2016-11-13 15:01:11 +08:00
|
|
|
// TODO: Factor out code common with LowerSINT_TO_FP.
|
|
|
|
|
2014-10-04 07:54:41 +08:00
|
|
|
EVT DestVT = Op.getValueType();
|
2016-11-13 15:01:11 +08:00
|
|
|
if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
|
|
|
|
SDLoc DL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
|
|
|
SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
|
|
|
|
SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
|
|
|
|
SDValue FPRound =
|
|
|
|
DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);
|
|
|
|
|
|
|
|
return FPRound;
|
|
|
|
}
|
2014-10-04 07:54:41 +08:00
|
|
|
|
2016-01-12 06:01:48 +08:00
|
|
|
if (DestVT == MVT::f32)
|
|
|
|
return LowerINT_TO_FP32(Op, DAG, false);
|
2014-10-04 07:54:41 +08:00
|
|
|
|
2016-07-28 08:32:05 +08:00
|
|
|
assert(DestVT == MVT::f64);
|
|
|
|
return LowerINT_TO_FP64(Op, DAG, false);
|
2013-10-31 01:22:05 +08:00
|
|
|
}
|
2013-08-16 09:12:11 +08:00
|
|
|
|
2014-10-04 07:54:41 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
2016-01-12 06:01:48 +08:00
|
|
|
assert(Op.getOperand(0).getValueType() == MVT::i64 &&
|
|
|
|
"operation should be legal");
|
|
|
|
|
2016-11-13 15:01:11 +08:00
|
|
|
// TODO: Factor out code common with LowerUINT_TO_FP.
|
|
|
|
|
2016-01-12 06:01:48 +08:00
|
|
|
EVT DestVT = Op.getValueType();
|
2016-11-13 15:01:11 +08:00
|
|
|
if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
|
|
|
|
SDLoc DL(Op);
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
|
|
|
SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
|
|
|
|
SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
|
|
|
|
SDValue FPRound =
|
|
|
|
DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);
|
|
|
|
|
|
|
|
return FPRound;
|
|
|
|
}
|
|
|
|
|
2016-01-12 06:01:48 +08:00
|
|
|
if (DestVT == MVT::f32)
|
|
|
|
return LowerINT_TO_FP32(Op, DAG, true);
|
|
|
|
|
2016-07-28 08:32:05 +08:00
|
|
|
assert(DestVT == MVT::f64);
|
|
|
|
return LowerINT_TO_FP64(Op, DAG, true);
|
2014-10-04 07:54:41 +08:00
|
|
|
}
|
|
|
|
|
2014-10-04 07:54:56 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
|
|
|
|
bool Signed) const {
|
|
|
|
SDLoc SL(Op);
|
|
|
|
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
|
|
|
SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
|
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), SL,
|
|
|
|
MVT::f64);
|
|
|
|
SDValue K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), SL,
|
|
|
|
MVT::f64);
|
2015-09-17 00:31:21 +08:00
|
|
|
// TODO: Should this propagate fast-math-flags?
|
2014-10-04 07:54:56 +08:00
|
|
|
SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
|
|
|
|
|
|
|
|
SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
|
|
|
|
|
|
|
|
|
|
|
|
SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
|
|
|
|
|
|
|
|
SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
|
|
|
|
MVT::i32, FloorMul);
|
|
|
|
SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
|
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
SDValue Result = DAG.getBuildVector(MVT::v2i32, SL, {Lo, Hi});
|
2014-10-04 07:54:56 +08:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
|
|
|
|
}
|
|
|
|
|
2016-11-02 00:31:48 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const {
|
2017-03-16 03:04:26 +08:00
|
|
|
SDLoc DL(Op);
|
|
|
|
SDValue N0 = Op.getOperand(0);
|
|
|
|
|
|
|
|
// Convert to target node to get known bits
|
|
|
|
if (N0.getValueType() == MVT::f32)
|
|
|
|
return DAG.getNode(AMDGPUISD::FP_TO_FP16, DL, Op.getValueType(), N0);
|
2016-11-02 00:31:48 +08:00
|
|
|
|
|
|
|
if (getTargetMachine().Options.UnsafeFPMath) {
|
|
|
|
// There is a generic expand for FP_TO_FP16 with unsafe fast math.
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
2017-03-16 03:04:26 +08:00
|
|
|
assert(N0.getSimpleValueType() == MVT::f64);
|
2016-11-02 00:31:48 +08:00
|
|
|
|
|
|
|
// f64 -> f16 conversion using round-to-nearest-even rounding mode.
|
|
|
|
const unsigned ExpMask = 0x7ff;
|
|
|
|
const unsigned ExpBiasf64 = 1023;
|
|
|
|
const unsigned ExpBiasf16 = 15;
|
|
|
|
SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
|
|
|
|
SDValue One = DAG.getConstant(1, DL, MVT::i32);
|
|
|
|
SDValue U = DAG.getNode(ISD::BITCAST, DL, MVT::i64, N0);
|
|
|
|
SDValue UH = DAG.getNode(ISD::SRL, DL, MVT::i64, U,
|
|
|
|
DAG.getConstant(32, DL, MVT::i64));
|
|
|
|
UH = DAG.getZExtOrTrunc(UH, DL, MVT::i32);
|
|
|
|
U = DAG.getZExtOrTrunc(U, DL, MVT::i32);
|
|
|
|
SDValue E = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
|
|
|
|
DAG.getConstant(20, DL, MVT::i64));
|
|
|
|
E = DAG.getNode(ISD::AND, DL, MVT::i32, E,
|
|
|
|
DAG.getConstant(ExpMask, DL, MVT::i32));
|
|
|
|
// Subtract the fp64 exponent bias (1023) to get the real exponent and
|
|
|
|
// add the f16 bias (15) to get the biased exponent for the f16 format.
|
|
|
|
E = DAG.getNode(ISD::ADD, DL, MVT::i32, E,
|
|
|
|
DAG.getConstant(-ExpBiasf64 + ExpBiasf16, DL, MVT::i32));
|
|
|
|
|
|
|
|
SDValue M = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
|
|
|
|
DAG.getConstant(8, DL, MVT::i32));
|
|
|
|
M = DAG.getNode(ISD::AND, DL, MVT::i32, M,
|
|
|
|
DAG.getConstant(0xffe, DL, MVT::i32));
|
|
|
|
|
|
|
|
SDValue MaskedSig = DAG.getNode(ISD::AND, DL, MVT::i32, UH,
|
|
|
|
DAG.getConstant(0x1ff, DL, MVT::i32));
|
|
|
|
MaskedSig = DAG.getNode(ISD::OR, DL, MVT::i32, MaskedSig, U);
|
|
|
|
|
|
|
|
SDValue Lo40Set = DAG.getSelectCC(DL, MaskedSig, Zero, Zero, One, ISD::SETEQ);
|
|
|
|
M = DAG.getNode(ISD::OR, DL, MVT::i32, M, Lo40Set);
|
|
|
|
|
|
|
|
// (M != 0 ? 0x0200 : 0) | 0x7c00;
|
|
|
|
SDValue I = DAG.getNode(ISD::OR, DL, MVT::i32,
|
|
|
|
DAG.getSelectCC(DL, M, Zero, DAG.getConstant(0x0200, DL, MVT::i32),
|
|
|
|
Zero, ISD::SETNE), DAG.getConstant(0x7c00, DL, MVT::i32));
|
|
|
|
|
|
|
|
// N = M | (E << 12);
|
|
|
|
SDValue N = DAG.getNode(ISD::OR, DL, MVT::i32, M,
|
|
|
|
DAG.getNode(ISD::SHL, DL, MVT::i32, E,
|
|
|
|
DAG.getConstant(12, DL, MVT::i32)));
|
|
|
|
|
|
|
|
// B = clamp(1-E, 0, 13);
|
|
|
|
SDValue OneSubExp = DAG.getNode(ISD::SUB, DL, MVT::i32,
|
|
|
|
One, E);
|
|
|
|
SDValue B = DAG.getNode(ISD::SMAX, DL, MVT::i32, OneSubExp, Zero);
|
|
|
|
B = DAG.getNode(ISD::SMIN, DL, MVT::i32, B,
|
|
|
|
DAG.getConstant(13, DL, MVT::i32));
|
|
|
|
|
|
|
|
SDValue SigSetHigh = DAG.getNode(ISD::OR, DL, MVT::i32, M,
|
|
|
|
DAG.getConstant(0x1000, DL, MVT::i32));
|
|
|
|
|
|
|
|
SDValue D = DAG.getNode(ISD::SRL, DL, MVT::i32, SigSetHigh, B);
|
|
|
|
SDValue D0 = DAG.getNode(ISD::SHL, DL, MVT::i32, D, B);
|
|
|
|
SDValue D1 = DAG.getSelectCC(DL, D0, SigSetHigh, One, Zero, ISD::SETNE);
|
|
|
|
D = DAG.getNode(ISD::OR, DL, MVT::i32, D, D1);
|
|
|
|
|
|
|
|
SDValue V = DAG.getSelectCC(DL, E, One, D, N, ISD::SETLT);
|
|
|
|
SDValue VLow3 = DAG.getNode(ISD::AND, DL, MVT::i32, V,
|
|
|
|
DAG.getConstant(0x7, DL, MVT::i32));
|
|
|
|
V = DAG.getNode(ISD::SRL, DL, MVT::i32, V,
|
|
|
|
DAG.getConstant(2, DL, MVT::i32));
|
|
|
|
SDValue V0 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(3, DL, MVT::i32),
|
|
|
|
One, Zero, ISD::SETEQ);
|
|
|
|
SDValue V1 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(5, DL, MVT::i32),
|
|
|
|
One, Zero, ISD::SETGT);
|
|
|
|
V1 = DAG.getNode(ISD::OR, DL, MVT::i32, V0, V1);
|
|
|
|
V = DAG.getNode(ISD::ADD, DL, MVT::i32, V, V1);
|
|
|
|
|
|
|
|
V = DAG.getSelectCC(DL, E, DAG.getConstant(30, DL, MVT::i32),
|
|
|
|
DAG.getConstant(0x7c00, DL, MVT::i32), V, ISD::SETGT);
|
|
|
|
V = DAG.getSelectCC(DL, E, DAG.getConstant(1039, DL, MVT::i32),
|
|
|
|
I, V, ISD::SETEQ);
|
|
|
|
|
|
|
|
// Extract the sign bit.
|
|
|
|
SDValue Sign = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
|
|
|
|
DAG.getConstant(16, DL, MVT::i32));
|
|
|
|
Sign = DAG.getNode(ISD::AND, DL, MVT::i32, Sign,
|
|
|
|
DAG.getConstant(0x8000, DL, MVT::i32));
|
|
|
|
|
|
|
|
V = DAG.getNode(ISD::OR, DL, MVT::i32, Sign, V);
|
|
|
|
return DAG.getZExtOrTrunc(V, DL, Op.getValueType());
|
|
|
|
}
|
|
|
|
|
2014-10-04 07:54:56 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
2016-11-13 15:01:11 +08:00
|
|
|
// TODO: Factor out code common with LowerFP_TO_UINT.
|
|
|
|
|
|
|
|
EVT SrcVT = Src.getValueType();
|
|
|
|
if (Subtarget->has16BitInsts() && SrcVT == MVT::f16) {
|
|
|
|
SDLoc DL(Op);
|
|
|
|
|
|
|
|
SDValue FPExtend = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
|
|
|
|
SDValue FpToInt32 =
|
|
|
|
DAG.getNode(Op.getOpcode(), DL, MVT::i64, FPExtend);
|
|
|
|
|
|
|
|
return FpToInt32;
|
|
|
|
}
|
|
|
|
|
2014-10-04 07:54:56 +08:00
|
|
|
if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
|
|
|
|
return LowerFP64_TO_INT(Op, DAG, true);
|
|
|
|
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
SDValue Src = Op.getOperand(0);
|
|
|
|
|
2016-11-13 15:01:11 +08:00
|
|
|
// TODO: Factor out code common with LowerFP_TO_SINT.
|
|
|
|
|
|
|
|
EVT SrcVT = Src.getValueType();
|
|
|
|
if (Subtarget->has16BitInsts() && SrcVT == MVT::f16) {
|
|
|
|
SDLoc DL(Op);
|
|
|
|
|
|
|
|
SDValue FPExtend = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
|
|
|
|
SDValue FpToInt32 =
|
|
|
|
DAG.getNode(Op.getOpcode(), DL, MVT::i64, FPExtend);
|
|
|
|
|
|
|
|
return FpToInt32;
|
|
|
|
}
|
|
|
|
|
2014-10-04 07:54:56 +08:00
|
|
|
if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
|
|
|
|
return LowerFP64_TO_INT(Op, DAG, false);
|
|
|
|
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
2014-03-18 02:58:11 +08:00
|
|
|
SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
|
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
|
|
|
|
MVT VT = Op.getSimpleValueType();
|
|
|
|
MVT ScalarVT = VT.getScalarType();
|
|
|
|
|
2016-07-28 08:32:05 +08:00
|
|
|
assert(VT.isVector());
|
2014-03-18 02:58:11 +08:00
|
|
|
|
|
|
|
SDValue Src = Op.getOperand(0);
|
2014-04-22 11:49:30 +08:00
|
|
|
SDLoc DL(Op);
|
2014-03-18 02:58:11 +08:00
|
|
|
|
2014-04-22 11:49:30 +08:00
|
|
|
// TODO: Don't scalarize on Evergreen?
|
|
|
|
unsigned NElts = VT.getVectorNumElements();
|
|
|
|
SmallVector<SDValue, 8> Args;
|
|
|
|
DAG.ExtractVectorElements(Src, Args, 0, NElts);
|
2014-03-18 02:58:11 +08:00
|
|
|
|
2014-04-22 11:49:30 +08:00
|
|
|
SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
|
|
|
|
for (unsigned I = 0; I < NElts; ++I)
|
|
|
|
Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
|
2014-03-18 02:58:11 +08:00
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
return DAG.getBuildVector(VT, DL, Args);
|
2014-03-18 02:58:11 +08:00
|
|
|
}
|
|
|
|
|
2014-04-08 03:45:41 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Custom DAG optimizations
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
static bool isU24(SDValue Op, SelectionDAG &DAG) {
|
2017-11-07 01:04:37 +08:00
|
|
|
return AMDGPUTargetLowering::numBitsUnsigned(Op, DAG) <= 24;
|
2014-04-08 03:45:41 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static bool isI24(SDValue Op, SelectionDAG &DAG) {
|
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
|
|
|
|
// as unsigned 24-bit values.
|
2017-11-07 01:04:37 +08:00
|
|
|
AMDGPUTargetLowering::numBitsSigned(Op, DAG) < 24;
|
2014-04-08 03:45:41 +08:00
|
|
|
}
|
|
|
|
|
2016-10-15 03:14:29 +08:00
|
|
|
static bool simplifyI24(SDNode *Node24, unsigned OpIdx,
|
|
|
|
TargetLowering::DAGCombinerInfo &DCI) {
|
2014-04-08 03:45:41 +08:00
|
|
|
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
2016-10-15 03:14:29 +08:00
|
|
|
SDValue Op = Node24->getOperand(OpIdx);
|
2017-04-22 02:53:12 +08:00
|
|
|
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
|
2014-04-08 03:45:41 +08:00
|
|
|
EVT VT = Op.getValueType();
|
|
|
|
|
|
|
|
APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
|
|
|
|
APInt KnownZero, KnownOne;
|
|
|
|
TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
|
2017-04-22 02:53:12 +08:00
|
|
|
if (TLI.SimplifyDemandedBits(Node24, OpIdx, Demanded, DCI, TLO))
|
2016-08-27 09:32:27 +08:00
|
|
|
return true;
|
|
|
|
|
|
|
|
return false;
|
2014-04-08 03:45:41 +08:00
|
|
|
}
|
|
|
|
|
2014-05-23 02:09:07 +08:00
|
|
|
template <typename IntTy>
|
2016-06-12 23:39:02 +08:00
|
|
|
static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0, uint32_t Offset,
|
|
|
|
uint32_t Width, const SDLoc &DL) {
|
2014-05-23 02:09:07 +08:00
|
|
|
if (Width + Offset < 32) {
|
2014-09-19 08:42:06 +08:00
|
|
|
uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
|
|
|
|
IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
|
2015-04-28 22:05:47 +08:00
|
|
|
return DAG.getConstant(Result, DL, MVT::i32);
|
2014-05-23 02:09:07 +08:00
|
|
|
}
|
|
|
|
|
2015-04-28 22:05:47 +08:00
|
|
|
return DAG.getConstant(Src0 >> Offset, DL, MVT::i32);
|
2014-05-23 02:09:07 +08:00
|
|
|
}
|
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
static bool hasVolatileUser(SDNode *Val) {
|
|
|
|
for (SDNode *U : Val->uses()) {
|
|
|
|
if (MemSDNode *M = dyn_cast<MemSDNode>(U)) {
|
|
|
|
if (M->isVolatile())
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2016-07-02 06:55:55 +08:00
|
|
|
bool AMDGPUTargetLowering::shouldCombineMemoryType(EVT VT) const {
|
2016-07-02 06:47:50 +08:00
|
|
|
// i32 vectors are the canonical memory type.
|
|
|
|
if (VT.getScalarType() == MVT::i32 || isTypeLegal(VT))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (!VT.isByteSized())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
unsigned Size = VT.getStoreSize();
|
|
|
|
|
|
|
|
if ((Size == 1 || Size == 2 || Size == 4) && !VT.isVector())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (Size == 3 || (Size > 4 && (Size % 4 != 0)))
|
|
|
|
return false;
|
|
|
|
|
2014-07-15 10:06:31 +08:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
// Replace load of an illegal type with a store of a bitcast to a friendlier
|
|
|
|
// type.
|
|
|
|
SDValue AMDGPUTargetLowering::performLoadCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
if (!DCI.isBeforeLegalize())
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
LoadSDNode *LN = cast<LoadSDNode>(N);
|
|
|
|
if (LN->isVolatile() || !ISD::isNormalLoad(LN) || hasVolatileUser(LN))
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
SDLoc SL(N);
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
EVT VT = LN->getMemoryVT();
|
2016-07-02 06:55:55 +08:00
|
|
|
|
|
|
|
unsigned Size = VT.getStoreSize();
|
|
|
|
unsigned Align = LN->getAlignment();
|
|
|
|
if (Align < Size && isTypeLegal(VT)) {
|
|
|
|
bool IsFast;
|
|
|
|
unsigned AS = LN->getAddressSpace();
|
|
|
|
|
|
|
|
// Expand unaligned loads earlier than legalization. Due to visitation order
|
|
|
|
// problems during legalization, the emitted instructions to pack and unpack
|
|
|
|
// the bytes again are not eliminated in the case of an unaligned copy.
|
|
|
|
if (!allowsMisalignedMemoryAccesses(VT, AS, Align, &IsFast)) {
|
2016-09-01 05:52:27 +08:00
|
|
|
if (VT.isVector())
|
|
|
|
return scalarizeVectorLoad(LN, DAG);
|
|
|
|
|
2016-07-02 06:55:55 +08:00
|
|
|
SDValue Ops[2];
|
|
|
|
std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(LN, DAG);
|
|
|
|
return DAG.getMergeValues(Ops, SDLoc(N));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!IsFast)
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!shouldCombineMemoryType(VT))
|
|
|
|
return SDValue();
|
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
|
|
|
|
|
|
|
|
SDValue NewLoad
|
|
|
|
= DAG.getLoad(NewVT, SL, LN->getChain(),
|
|
|
|
LN->getBasePtr(), LN->getMemOperand());
|
|
|
|
|
|
|
|
SDValue BC = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad);
|
|
|
|
DCI.CombineTo(N, BC, NewLoad.getValue(1));
|
|
|
|
return SDValue(N, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Replace store of an illegal type with a store of a bitcast to a friendlier
|
|
|
|
// type.
|
2014-07-15 10:06:31 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
if (!DCI.isBeforeLegalize())
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
StoreSDNode *SN = cast<StoreSDNode>(N);
|
2016-07-02 06:47:50 +08:00
|
|
|
if (SN->isVolatile() || !ISD::isNormalStore(SN))
|
2014-07-15 10:06:31 +08:00
|
|
|
return SDValue();
|
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
EVT VT = SN->getMemoryVT();
|
2016-07-02 06:55:55 +08:00
|
|
|
unsigned Size = VT.getStoreSize();
|
2014-07-15 10:06:31 +08:00
|
|
|
|
|
|
|
SDLoc SL(N);
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
2016-07-02 06:55:55 +08:00
|
|
|
unsigned Align = SN->getAlignment();
|
|
|
|
if (Align < Size && isTypeLegal(VT)) {
|
|
|
|
bool IsFast;
|
|
|
|
unsigned AS = SN->getAddressSpace();
|
|
|
|
|
|
|
|
// Expand unaligned stores earlier than legalization. Due to visitation
|
|
|
|
// order problems during legalization, the emitted instructions to pack and
|
|
|
|
// unpack the bytes again are not eliminated in the case of an unaligned
|
|
|
|
// copy.
|
2016-09-01 05:52:27 +08:00
|
|
|
if (!allowsMisalignedMemoryAccesses(VT, AS, Align, &IsFast)) {
|
|
|
|
if (VT.isVector())
|
|
|
|
return scalarizeVectorStore(SN, DAG);
|
|
|
|
|
2016-07-02 06:55:55 +08:00
|
|
|
return expandUnalignedStore(SN, DAG);
|
2016-09-01 05:52:27 +08:00
|
|
|
}
|
2016-07-02 06:55:55 +08:00
|
|
|
|
|
|
|
if (!IsFast)
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!shouldCombineMemoryType(VT))
|
|
|
|
return SDValue();
|
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
|
2016-07-02 06:55:55 +08:00
|
|
|
SDValue Val = SN->getValue();
|
|
|
|
|
|
|
|
//DCI.AddToWorklist(Val.getNode());
|
2014-07-15 10:06:31 +08:00
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
bool OtherUses = !Val.hasOneUse();
|
|
|
|
SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, NewVT, Val);
|
|
|
|
if (OtherUses) {
|
|
|
|
SDValue CastBack = DAG.getNode(ISD::BITCAST, SL, VT, CastVal);
|
|
|
|
DAG.ReplaceAllUsesOfValueWith(Val, CastBack);
|
|
|
|
}
|
2014-07-15 10:06:31 +08:00
|
|
|
|
2016-07-02 06:47:50 +08:00
|
|
|
return DAG.getStore(SN->getChain(), SL, CastVal,
|
2014-07-15 10:06:31 +08:00
|
|
|
SN->getBasePtr(), SN->getMemOperand());
|
|
|
|
}
|
|
|
|
|
2017-07-15 13:52:59 +08:00
|
|
|
// FIXME: This should go in generic DAG combiner with an isTruncateFree check,
|
|
|
|
// but isTruncateFree is inaccurate for i16 now because of SALU vs. VALU
|
|
|
|
// issues.
|
|
|
|
SDValue AMDGPUTargetLowering::performAssertSZExtCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDValue N0 = N->getOperand(0);
|
|
|
|
|
|
|
|
// (vt2 (assertzext (truncate vt0:x), vt1)) ->
|
|
|
|
// (vt2 (truncate (assertzext vt0:x, vt1)))
|
|
|
|
if (N0.getOpcode() == ISD::TRUNCATE) {
|
|
|
|
SDValue N1 = N->getOperand(1);
|
|
|
|
EVT ExtVT = cast<VTSDNode>(N1)->getVT();
|
|
|
|
SDLoc SL(N);
|
|
|
|
|
|
|
|
SDValue Src = N0.getOperand(0);
|
|
|
|
EVT SrcVT = Src.getValueType();
|
|
|
|
if (SrcVT.bitsGE(ExtVT)) {
|
|
|
|
SDValue NewInReg = DAG.getNode(N->getOpcode(), SL, SrcVT, Src, N1);
|
|
|
|
return DAG.getNode(ISD::TRUNCATE, SL, N->getValueType(0), NewInReg);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return SDValue();
|
|
|
|
}
|
2016-09-14 23:19:03 +08:00
|
|
|
/// Split the 64-bit value \p LHS into two 32-bit components, and perform the
|
|
|
|
/// binary operation \p Opc to it with the corresponding constant operands.
|
|
|
|
SDValue AMDGPUTargetLowering::splitBinaryBitConstantOpImpl(
|
|
|
|
DAGCombinerInfo &DCI, const SDLoc &SL,
|
|
|
|
unsigned Opc, SDValue LHS,
|
|
|
|
uint32_t ValLo, uint32_t ValHi) const {
|
2016-01-19 06:01:13 +08:00
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDValue Lo, Hi;
|
2016-09-14 23:19:03 +08:00
|
|
|
std::tie(Lo, Hi) = split64BitValue(LHS, DAG);
|
2016-01-19 06:01:13 +08:00
|
|
|
|
2016-09-14 23:19:03 +08:00
|
|
|
SDValue LoRHS = DAG.getConstant(ValLo, SL, MVT::i32);
|
|
|
|
SDValue HiRHS = DAG.getConstant(ValHi, SL, MVT::i32);
|
2016-01-19 06:01:13 +08:00
|
|
|
|
2016-09-14 23:19:03 +08:00
|
|
|
SDValue LoAnd = DAG.getNode(Opc, SL, MVT::i32, Lo, LoRHS);
|
|
|
|
SDValue HiAnd = DAG.getNode(Opc, SL, MVT::i32, Hi, HiRHS);
|
2016-01-19 06:01:13 +08:00
|
|
|
|
2016-04-23 06:48:38 +08:00
|
|
|
// Re-visit the ands. It's possible we eliminated one of them and it could
|
|
|
|
// simplify the vector.
|
|
|
|
DCI.AddToWorklist(Lo.getNode());
|
|
|
|
DCI.AddToWorklist(Hi.getNode());
|
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {LoAnd, HiAnd});
|
2016-01-19 06:01:13 +08:00
|
|
|
return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
|
|
|
|
}
|
|
|
|
|
2015-07-15 02:20:33 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
2017-05-23 00:58:10 +08:00
|
|
|
EVT VT = N->getValueType(0);
|
2015-07-15 02:20:33 +08:00
|
|
|
|
2017-05-23 00:58:10 +08:00
|
|
|
ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
|
2016-01-19 05:55:14 +08:00
|
|
|
if (!RHS)
|
|
|
|
return SDValue();
|
|
|
|
|
2015-07-15 02:20:33 +08:00
|
|
|
SDValue LHS = N->getOperand(0);
|
2017-05-23 00:58:10 +08:00
|
|
|
unsigned RHSVal = RHS->getZExtValue();
|
|
|
|
if (!RHSVal)
|
|
|
|
return LHS;
|
2015-07-15 02:20:33 +08:00
|
|
|
|
|
|
|
SDLoc SL(N);
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
|
2017-05-23 00:58:10 +08:00
|
|
|
switch (LHS->getOpcode()) {
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
case ISD::ZERO_EXTEND:
|
|
|
|
case ISD::SIGN_EXTEND:
|
|
|
|
case ISD::ANY_EXTEND: {
|
2017-09-01 05:17:22 +08:00
|
|
|
SDValue X = LHS->getOperand(0);
|
|
|
|
|
|
|
|
if (VT == MVT::i32 && RHSVal == 16 && X.getValueType() == MVT::i16 &&
|
2018-05-22 14:32:10 +08:00
|
|
|
isOperationLegal(ISD::BUILD_VECTOR, MVT::v2i16)) {
|
2017-09-01 05:17:22 +08:00
|
|
|
// Prefer build_vector as the canonical form if packed types are legal.
|
|
|
|
// (shl ([asz]ext i16:x), 16 -> build_vector 0, x
|
|
|
|
SDValue Vec = DAG.getBuildVector(MVT::v2i16, SL,
|
|
|
|
{ DAG.getConstant(0, SL, MVT::i16), LHS->getOperand(0) });
|
|
|
|
return DAG.getNode(ISD::BITCAST, SL, MVT::i32, Vec);
|
|
|
|
}
|
|
|
|
|
2017-05-23 00:58:10 +08:00
|
|
|
// shl (ext x) => zext (shl x), if shift does not overflow int
|
2017-05-23 23:59:58 +08:00
|
|
|
if (VT != MVT::i64)
|
|
|
|
break;
|
2017-05-23 00:58:10 +08:00
|
|
|
KnownBits Known;
|
|
|
|
DAG.computeKnownBits(X, Known);
|
|
|
|
unsigned LZ = Known.countMinLeadingZeros();
|
|
|
|
if (LZ < RHSVal)
|
|
|
|
break;
|
|
|
|
EVT XVT = X.getValueType();
|
|
|
|
SDValue Shl = DAG.getNode(ISD::SHL, SL, XVT, X, SDValue(RHS, 0));
|
|
|
|
return DAG.getZExtOrTrunc(Shl, SL, VT);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-05-23 23:59:58 +08:00
|
|
|
if (VT != MVT::i64)
|
|
|
|
return SDValue();
|
|
|
|
|
2017-05-23 00:58:10 +08:00
|
|
|
// i64 (shl x, C) -> (build_pair 0, (shl x, C -32))
|
|
|
|
|
|
|
|
// On some subtargets, 64-bit shift is a quarter rate instruction. In the
|
|
|
|
// common case, splitting this into a move and a 32-bit shift is faster and
|
|
|
|
// the same code size.
|
|
|
|
if (RHSVal < 32)
|
|
|
|
return SDValue();
|
|
|
|
|
2016-01-19 05:55:14 +08:00
|
|
|
SDValue ShiftAmt = DAG.getConstant(RHSVal - 32, SL, MVT::i32);
|
|
|
|
|
2015-07-15 02:20:33 +08:00
|
|
|
SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
|
2016-01-19 05:55:14 +08:00
|
|
|
SDValue NewShift = DAG.getNode(ISD::SHL, SL, MVT::i32, Lo, ShiftAmt);
|
2015-07-15 02:20:33 +08:00
|
|
|
|
|
|
|
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
|
2016-01-19 05:43:36 +08:00
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {Zero, NewShift});
|
2016-01-19 05:55:14 +08:00
|
|
|
return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
|
2015-07-15 02:20:33 +08:00
|
|
|
}
|
|
|
|
|
2016-01-19 06:09:04 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performSraCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
if (N->getValueType(0) != MVT::i64)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
|
|
|
|
if (!RHS)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDLoc SL(N);
|
|
|
|
unsigned RHSVal = RHS->getZExtValue();
|
|
|
|
|
|
|
|
// (sra i64:x, 32) -> build_pair x, (sra hi_32(x), 31)
|
|
|
|
if (RHSVal == 32) {
|
|
|
|
SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
|
|
|
|
SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
|
|
|
|
DAG.getConstant(31, SL, MVT::i32));
|
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {Hi, NewShift});
|
2016-01-19 06:09:04 +08:00
|
|
|
return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
|
|
|
|
}
|
|
|
|
|
|
|
|
// (sra i64:x, 63) -> build_pair (sra hi_32(x), 31), (sra hi_32(x), 31)
|
|
|
|
if (RHSVal == 63) {
|
|
|
|
SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
|
|
|
|
SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
|
|
|
|
DAG.getConstant(31, SL, MVT::i32));
|
2016-04-27 05:15:30 +08:00
|
|
|
SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, NewShift});
|
2016-01-19 06:09:04 +08:00
|
|
|
return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
|
|
|
|
}
|
|
|
|
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
2016-01-19 05:43:36 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performSrlCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
if (N->getValueType(0) != MVT::i64)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
|
|
|
|
if (!RHS)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
unsigned ShiftAmt = RHS->getZExtValue();
|
|
|
|
if (ShiftAmt < 32)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
// srl i64:x, C for C >= 32
|
|
|
|
// =>
|
|
|
|
// build_pair (srl hi_32(x), C - 32), 0
|
|
|
|
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDLoc SL(N);
|
|
|
|
|
|
|
|
SDValue One = DAG.getConstant(1, SL, MVT::i32);
|
|
|
|
SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
|
|
|
|
|
|
|
|
SDValue VecOp = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, N->getOperand(0));
|
|
|
|
SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32,
|
|
|
|
VecOp, One);
|
|
|
|
|
|
|
|
SDValue NewConst = DAG.getConstant(ShiftAmt - 32, SL, MVT::i32);
|
|
|
|
SDValue NewShift = DAG.getNode(ISD::SRL, SL, MVT::i32, Hi, NewConst);
|
|
|
|
|
2016-04-27 05:15:30 +08:00
|
|
|
SDValue BuildPair = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, Zero});
|
2016-01-19 05:43:36 +08:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildPair);
|
|
|
|
}
|
|
|
|
|
2018-05-10 02:37:39 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performTruncateCombine(
|
|
|
|
SDNode *N, DAGCombinerInfo &DCI) const {
|
|
|
|
SDLoc SL(N);
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
EVT VT = N->getValueType(0);
|
|
|
|
SDValue Src = N->getOperand(0);
|
|
|
|
|
|
|
|
// vt1 (truncate (bitcast (build_vector vt0:x, ...))) -> vt1 (bitcast vt0:x)
|
|
|
|
if (Src.getOpcode() == ISD::BITCAST) {
|
|
|
|
SDValue Vec = Src.getOperand(0);
|
|
|
|
if (Vec.getOpcode() == ISD::BUILD_VECTOR) {
|
|
|
|
SDValue Elt0 = Vec.getOperand(0);
|
|
|
|
EVT EltVT = Elt0.getValueType();
|
|
|
|
if (VT.getSizeInBits() <= EltVT.getSizeInBits()) {
|
|
|
|
if (EltVT.isFloatingPoint()) {
|
|
|
|
Elt0 = DAG.getNode(ISD::BITCAST, SL,
|
|
|
|
EltVT.changeTypeToInteger(), Elt0);
|
|
|
|
}
|
|
|
|
|
|
|
|
return DAG.getNode(ISD::TRUNCATE, SL, VT, Elt0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-05-16 19:47:30 +08:00
|
|
|
// Equivalent of above for accessing the high element of a vector as an
|
|
|
|
// integer operation.
|
|
|
|
// trunc (srl (bitcast (build_vector x, y))), 16 -> trunc (bitcast y)
|
|
|
|
if (Src.getOpcode() == ISD::SRL) {
|
|
|
|
if (auto K = isConstOrConstSplat(Src.getOperand(1))) {
|
|
|
|
if (2 * K->getZExtValue() == Src.getValueType().getScalarSizeInBits()) {
|
|
|
|
SDValue BV = stripBitcast(Src.getOperand(0));
|
|
|
|
if (BV.getOpcode() == ISD::BUILD_VECTOR &&
|
|
|
|
BV.getValueType().getVectorNumElements() == 2) {
|
|
|
|
SDValue SrcElt = BV.getOperand(1);
|
|
|
|
EVT SrcEltVT = SrcElt.getValueType();
|
|
|
|
if (SrcEltVT.isFloatingPoint()) {
|
|
|
|
SrcElt = DAG.getNode(ISD::BITCAST, SL,
|
|
|
|
SrcEltVT.changeTypeToInteger(), SrcElt);
|
|
|
|
}
|
|
|
|
|
|
|
|
return DAG.getNode(ISD::TRUNCATE, SL, VT, SrcElt);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-05-10 04:52:43 +08:00
|
|
|
// Partially shrink 64-bit shifts to 32-bit if reduced to 16-bit.
|
|
|
|
//
|
|
|
|
// i16 (trunc (srl i64:x, K)), K <= 16 ->
|
|
|
|
// i16 (trunc (srl (i32 (trunc x), K)))
|
|
|
|
if (VT.getScalarSizeInBits() < 32) {
|
|
|
|
EVT SrcVT = Src.getValueType();
|
|
|
|
if (SrcVT.getScalarSizeInBits() > 32 &&
|
|
|
|
(Src.getOpcode() == ISD::SRL ||
|
|
|
|
Src.getOpcode() == ISD::SRA ||
|
|
|
|
Src.getOpcode() == ISD::SHL)) {
|
2018-05-10 04:52:54 +08:00
|
|
|
SDValue Amt = Src.getOperand(1);
|
|
|
|
KnownBits Known;
|
|
|
|
DAG.computeKnownBits(Amt, Known);
|
|
|
|
unsigned Size = VT.getScalarSizeInBits();
|
|
|
|
if ((Known.isConstant() && Known.getConstant().ule(Size)) ||
|
|
|
|
(Known.getBitWidth() - Known.countMinLeadingZeros() <= Log2_32(Size))) {
|
|
|
|
EVT MidVT = VT.isVector() ?
|
|
|
|
EVT::getVectorVT(*DAG.getContext(), MVT::i32,
|
|
|
|
VT.getVectorNumElements()) : MVT::i32;
|
|
|
|
|
|
|
|
EVT NewShiftVT = getShiftAmountTy(MidVT, DAG.getDataLayout());
|
|
|
|
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, MidVT,
|
|
|
|
Src.getOperand(0));
|
|
|
|
DCI.AddToWorklist(Trunc.getNode());
|
|
|
|
|
|
|
|
if (Amt.getValueType() != NewShiftVT) {
|
|
|
|
Amt = DAG.getZExtOrTrunc(Amt, SL, NewShiftVT);
|
|
|
|
DCI.AddToWorklist(Amt.getNode());
|
2018-05-10 04:52:43 +08:00
|
|
|
}
|
2018-05-10 04:52:54 +08:00
|
|
|
|
|
|
|
SDValue ShrunkShift = DAG.getNode(Src.getOpcode(), SL, MidVT,
|
|
|
|
Trunc, Amt);
|
|
|
|
return DAG.getNode(ISD::TRUNCATE, SL, VT, ShrunkShift);
|
2018-05-10 04:52:43 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-05-10 02:37:39 +08:00
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
2016-08-27 09:32:27 +08:00
|
|
|
// We need to specifically handle i64 mul here to avoid unnecessary conversion
|
|
|
|
// instructions. If we only match on the legalized i64 mul expansion,
|
|
|
|
// SimplifyDemandedBits will be unable to remove them because there will be
|
|
|
|
// multiple uses due to the separate mul + mulh[su].
|
|
|
|
static SDValue getMul24(SelectionDAG &DAG, const SDLoc &SL,
|
|
|
|
SDValue N0, SDValue N1, unsigned Size, bool Signed) {
|
|
|
|
if (Size <= 32) {
|
|
|
|
unsigned MulOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
|
|
|
|
return DAG.getNode(MulOpc, SL, MVT::i32, N0, N1);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Because we want to eliminate extension instructions before the
|
|
|
|
// operation, we need to create a single user here (i.e. not the separate
|
|
|
|
// mul_lo + mul_hi) so that SimplifyDemandedBits will deal with it.
|
|
|
|
|
|
|
|
unsigned MulOpc = Signed ? AMDGPUISD::MUL_LOHI_I24 : AMDGPUISD::MUL_LOHI_U24;
|
|
|
|
|
|
|
|
SDValue Mul = DAG.getNode(MulOpc, SL,
|
|
|
|
DAG.getVTList(MVT::i32, MVT::i32), N0, N1);
|
|
|
|
|
|
|
|
return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64,
|
|
|
|
Mul.getValue(0), Mul.getValue(1));
|
|
|
|
}
|
|
|
|
|
2014-07-01 01:55:48 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
EVT VT = N->getValueType(0);
|
|
|
|
|
2016-08-27 09:32:27 +08:00
|
|
|
unsigned Size = VT.getSizeInBits();
|
|
|
|
if (VT.isVector() || Size > 64)
|
2014-07-01 01:55:48 +08:00
|
|
|
return SDValue();
|
|
|
|
|
2016-11-11 00:02:37 +08:00
|
|
|
// There are i16 integer mul/mad.
|
|
|
|
if (Subtarget->has16BitInsts() && VT.getScalarType().bitsLE(MVT::i16))
|
|
|
|
return SDValue();
|
|
|
|
|
2014-07-01 01:55:48 +08:00
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDLoc DL(N);
|
|
|
|
|
|
|
|
SDValue N0 = N->getOperand(0);
|
|
|
|
SDValue N1 = N->getOperand(1);
|
2018-05-10 05:11:35 +08:00
|
|
|
|
|
|
|
// SimplifyDemandedBits has the annoying habit of turning useful zero_extends
|
|
|
|
// in the source into any_extends if the result of the mul is truncated. Since
|
|
|
|
// we can assume the high bits are whatever we want, use the underlying value
|
|
|
|
// to avoid the unknown high bits from interfering.
|
|
|
|
if (N0.getOpcode() == ISD::ANY_EXTEND)
|
|
|
|
N0 = N0.getOperand(0);
|
|
|
|
|
|
|
|
if (N1.getOpcode() == ISD::ANY_EXTEND)
|
|
|
|
N1 = N1.getOperand(0);
|
|
|
|
|
2014-07-01 01:55:48 +08:00
|
|
|
SDValue Mul;
|
|
|
|
|
|
|
|
if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
|
|
|
|
N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
|
|
|
|
N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
|
2016-08-27 09:32:27 +08:00
|
|
|
Mul = getMul24(DAG, DL, N0, N1, Size, false);
|
2014-07-01 01:55:48 +08:00
|
|
|
} else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
|
|
|
|
N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
|
|
|
|
N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
|
2016-08-27 09:32:27 +08:00
|
|
|
Mul = getMul24(DAG, DL, N0, N1, Size, true);
|
2014-07-01 01:55:48 +08:00
|
|
|
} else {
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
|
|
|
// We need to use sext even for MUL_U24, because MUL_U24 is used
|
|
|
|
// for signed multiply of 8 and 16-bit types.
|
|
|
|
return DAG.getSExtOrTrunc(Mul, DL, VT);
|
|
|
|
}
|
|
|
|
|
2016-08-27 09:32:27 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performMulhsCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
EVT VT = N->getValueType(0);
|
|
|
|
|
|
|
|
if (!Subtarget->hasMulI24() || VT.isVector())
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDLoc DL(N);
|
|
|
|
|
|
|
|
SDValue N0 = N->getOperand(0);
|
|
|
|
SDValue N1 = N->getOperand(1);
|
|
|
|
|
|
|
|
if (!isI24(N0, DAG) || !isI24(N1, DAG))
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
|
|
|
|
N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
|
|
|
|
|
|
|
|
SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_I24, DL, MVT::i32, N0, N1);
|
|
|
|
DCI.AddToWorklist(Mulhi.getNode());
|
|
|
|
return DAG.getSExtOrTrunc(Mulhi, DL, VT);
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::performMulhuCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
EVT VT = N->getValueType(0);
|
|
|
|
|
|
|
|
if (!Subtarget->hasMulU24() || VT.isVector() || VT.getSizeInBits() > 32)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDLoc DL(N);
|
|
|
|
|
|
|
|
SDValue N0 = N->getOperand(0);
|
|
|
|
SDValue N1 = N->getOperand(1);
|
|
|
|
|
|
|
|
if (!isU24(N0, DAG) || !isU24(N1, DAG))
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
|
|
|
|
N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
|
|
|
|
|
|
|
|
SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_U24, DL, MVT::i32, N0, N1);
|
|
|
|
DCI.AddToWorklist(Mulhi.getNode());
|
|
|
|
return DAG.getZExtOrTrunc(Mulhi, DL, VT);
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::performMulLoHi24Combine(
|
|
|
|
SDNode *N, DAGCombinerInfo &DCI) const {
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
|
|
|
|
// Simplify demanded bits before splitting into multiple users.
|
2016-10-15 03:14:29 +08:00
|
|
|
if (simplifyI24(N, 0, DCI) || simplifyI24(N, 1, DCI))
|
2016-08-27 09:32:27 +08:00
|
|
|
return SDValue();
|
|
|
|
|
2016-10-15 03:14:29 +08:00
|
|
|
SDValue N0 = N->getOperand(0);
|
|
|
|
SDValue N1 = N->getOperand(1);
|
|
|
|
|
2016-08-27 09:32:27 +08:00
|
|
|
bool Signed = (N->getOpcode() == AMDGPUISD::MUL_LOHI_I24);
|
|
|
|
|
|
|
|
unsigned MulLoOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
|
|
|
|
unsigned MulHiOpc = Signed ? AMDGPUISD::MULHI_I24 : AMDGPUISD::MULHI_U24;
|
|
|
|
|
|
|
|
SDLoc SL(N);
|
|
|
|
|
|
|
|
SDValue MulLo = DAG.getNode(MulLoOpc, SL, MVT::i32, N0, N1);
|
|
|
|
SDValue MulHi = DAG.getNode(MulHiOpc, SL, MVT::i32, N0, N1);
|
|
|
|
return DAG.getMergeValues({ MulLo, MulHi }, SL);
|
|
|
|
}
|
|
|
|
|
2016-01-12 01:02:00 +08:00
|
|
|
static bool isNegativeOne(SDValue Val) {
|
|
|
|
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val))
|
|
|
|
return C->isAllOnesValue();
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
SDValue AMDGPUTargetLowering::getFFBX_U32(SelectionDAG &DAG,
|
2016-11-02 01:49:33 +08:00
|
|
|
SDValue Op,
|
2017-10-13 03:37:14 +08:00
|
|
|
const SDLoc &DL,
|
|
|
|
unsigned Opc) const {
|
2016-01-12 01:02:06 +08:00
|
|
|
EVT VT = Op.getValueType();
|
2016-11-02 01:49:33 +08:00
|
|
|
EVT LegalVT = getTypeToTransformTo(*DAG.getContext(), VT);
|
|
|
|
if (LegalVT != MVT::i32 && (Subtarget->has16BitInsts() &&
|
|
|
|
LegalVT != MVT::i16))
|
2016-01-12 01:02:06 +08:00
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
if (VT != MVT::i32)
|
2016-11-02 01:49:33 +08:00
|
|
|
Op = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Op);
|
2016-01-12 01:02:06 +08:00
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
SDValue FFBX = DAG.getNode(Opc, DL, MVT::i32, Op);
|
2016-01-12 01:02:06 +08:00
|
|
|
if (VT != MVT::i32)
|
2017-10-13 03:37:14 +08:00
|
|
|
FFBX = DAG.getNode(ISD::TRUNCATE, DL, VT, FFBX);
|
2016-01-12 01:02:06 +08:00
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
return FFBX;
|
2016-01-12 01:02:06 +08:00
|
|
|
}
|
|
|
|
|
2016-01-12 01:02:00 +08:00
|
|
|
// The native instructions return -1 on 0 input. Optimize out a select that
|
|
|
|
// produces -1 on 0.
|
|
|
|
//
|
|
|
|
// TODO: If zero is not undef, we could also do this if the output is compared
|
|
|
|
// against the bitwidth.
|
|
|
|
//
|
|
|
|
// TODO: Should probably combine against FFBH_U32 instead of ctlz directly.
|
2017-10-13 03:37:14 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performCtlz_CttzCombine(const SDLoc &SL, SDValue Cond,
|
2016-06-12 23:39:02 +08:00
|
|
|
SDValue LHS, SDValue RHS,
|
2016-01-12 01:02:00 +08:00
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
ConstantSDNode *CmpRhs = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
|
|
|
|
if (!CmpRhs || !CmpRhs->isNullValue())
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
ISD::CondCode CCOpcode = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
|
|
|
|
SDValue CmpLHS = Cond.getOperand(0);
|
|
|
|
|
2017-10-13 03:37:14 +08:00
|
|
|
unsigned Opc = isCttzOpc(RHS.getOpcode()) ? AMDGPUISD::FFBL_B32 :
|
|
|
|
AMDGPUISD::FFBH_U32;
|
|
|
|
|
2016-01-12 01:02:00 +08:00
|
|
|
// select (setcc x, 0, eq), -1, (ctlz_zero_undef x) -> ffbh_u32 x
|
2017-10-13 03:37:14 +08:00
|
|
|
// select (setcc x, 0, eq), -1, (cttz_zero_undef x) -> ffbl_u32 x
|
2016-01-12 01:02:00 +08:00
|
|
|
if (CCOpcode == ISD::SETEQ &&
|
2017-10-13 03:37:14 +08:00
|
|
|
(isCtlzOpc(RHS.getOpcode()) || isCttzOpc(RHS.getOpcode())) &&
|
2016-01-12 01:02:00 +08:00
|
|
|
RHS.getOperand(0) == CmpLHS &&
|
|
|
|
isNegativeOne(LHS)) {
|
2017-10-13 03:37:14 +08:00
|
|
|
return getFFBX_U32(DAG, CmpLHS, SL, Opc);
|
2016-01-12 01:02:00 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
// select (setcc x, 0, ne), (ctlz_zero_undef x), -1 -> ffbh_u32 x
|
2017-10-13 03:37:14 +08:00
|
|
|
// select (setcc x, 0, ne), (cttz_zero_undef x), -1 -> ffbl_u32 x
|
2016-01-12 01:02:00 +08:00
|
|
|
if (CCOpcode == ISD::SETNE &&
|
2017-10-13 03:37:14 +08:00
|
|
|
(isCtlzOpc(LHS.getOpcode()) || isCttzOpc(RHS.getOpcode())) &&
|
2016-01-12 01:02:00 +08:00
|
|
|
LHS.getOperand(0) == CmpLHS &&
|
|
|
|
isNegativeOne(RHS)) {
|
2017-10-13 03:37:14 +08:00
|
|
|
return getFFBX_U32(DAG, CmpLHS, SL, Opc);
|
2016-01-12 01:02:00 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
2017-01-12 07:57:38 +08:00
|
|
|
static SDValue distributeOpThroughSelect(TargetLowering::DAGCombinerInfo &DCI,
|
|
|
|
unsigned Op,
|
|
|
|
const SDLoc &SL,
|
|
|
|
SDValue Cond,
|
|
|
|
SDValue N1,
|
|
|
|
SDValue N2) {
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
EVT VT = N1.getValueType();
|
|
|
|
|
|
|
|
SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT, Cond,
|
|
|
|
N1.getOperand(0), N2.getOperand(0));
|
|
|
|
DCI.AddToWorklist(NewSelect.getNode());
|
|
|
|
return DAG.getNode(Op, SL, VT, NewSelect);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Pull a free FP operation out of a select so it may fold into uses.
|
|
|
|
//
|
|
|
|
// select c, (fneg x), (fneg y) -> fneg (select c, x, y)
|
|
|
|
// select c, (fneg x), k -> fneg (select c, x, (fneg k))
|
|
|
|
//
|
|
|
|
// select c, (fabs x), (fabs y) -> fabs (select c, x, y)
|
|
|
|
// select c, (fabs x), +k -> fabs (select c, x, k)
|
|
|
|
static SDValue foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI,
|
|
|
|
SDValue N) {
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDValue Cond = N.getOperand(0);
|
|
|
|
SDValue LHS = N.getOperand(1);
|
|
|
|
SDValue RHS = N.getOperand(2);
|
|
|
|
|
|
|
|
EVT VT = N.getValueType();
|
|
|
|
if ((LHS.getOpcode() == ISD::FABS && RHS.getOpcode() == ISD::FABS) ||
|
|
|
|
(LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG)) {
|
|
|
|
return distributeOpThroughSelect(DCI, LHS.getOpcode(),
|
|
|
|
SDLoc(N), Cond, LHS, RHS);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool Inv = false;
|
|
|
|
if (RHS.getOpcode() == ISD::FABS || RHS.getOpcode() == ISD::FNEG) {
|
|
|
|
std::swap(LHS, RHS);
|
|
|
|
Inv = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// TODO: Support vector constants.
|
|
|
|
ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
|
|
|
|
if ((LHS.getOpcode() == ISD::FNEG || LHS.getOpcode() == ISD::FABS) && CRHS) {
|
|
|
|
SDLoc SL(N);
|
|
|
|
// If one side is an fneg/fabs and the other is a constant, we can push the
|
|
|
|
// fneg/fabs down. If it's an fabs, the constant needs to be non-negative.
|
|
|
|
SDValue NewLHS = LHS.getOperand(0);
|
|
|
|
SDValue NewRHS = RHS;
|
|
|
|
|
2017-01-13 02:58:15 +08:00
|
|
|
// Careful: if the neg can be folded up, don't try to pull it back down.
|
|
|
|
bool ShouldFoldNeg = true;
|
2017-01-12 07:57:38 +08:00
|
|
|
|
2017-01-13 02:58:15 +08:00
|
|
|
if (NewLHS.hasOneUse()) {
|
|
|
|
unsigned Opc = NewLHS.getOpcode();
|
|
|
|
if (LHS.getOpcode() == ISD::FNEG && fnegFoldsIntoOp(Opc))
|
|
|
|
ShouldFoldNeg = false;
|
|
|
|
if (LHS.getOpcode() == ISD::FABS && Opc == ISD::FMUL)
|
|
|
|
ShouldFoldNeg = false;
|
|
|
|
}
|
2017-01-12 07:57:38 +08:00
|
|
|
|
2017-01-13 02:58:15 +08:00
|
|
|
if (ShouldFoldNeg) {
|
|
|
|
if (LHS.getOpcode() == ISD::FNEG)
|
|
|
|
NewRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
|
|
|
|
else if (CRHS->isNegative())
|
|
|
|
return SDValue();
|
2017-01-12 07:57:38 +08:00
|
|
|
|
2017-01-13 02:58:15 +08:00
|
|
|
if (Inv)
|
|
|
|
std::swap(NewLHS, NewRHS);
|
|
|
|
|
|
|
|
SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT,
|
|
|
|
Cond, NewLHS, NewRHS);
|
|
|
|
DCI.AddToWorklist(NewSelect.getNode());
|
|
|
|
return DAG.getNode(LHS.getOpcode(), SL, VT, NewSelect);
|
|
|
|
}
|
2017-01-12 07:57:38 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2016-01-12 01:02:00 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
2017-01-12 07:57:38 +08:00
|
|
|
if (SDValue Folded = foldFreeOpFromSelect(DCI, SDValue(N, 0)))
|
|
|
|
return Folded;
|
|
|
|
|
2016-01-12 01:02:00 +08:00
|
|
|
SDValue Cond = N->getOperand(0);
|
|
|
|
if (Cond.getOpcode() != ISD::SETCC)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
EVT VT = N->getValueType(0);
|
|
|
|
SDValue LHS = Cond.getOperand(0);
|
|
|
|
SDValue RHS = Cond.getOperand(1);
|
|
|
|
SDValue CC = Cond.getOperand(2);
|
|
|
|
|
|
|
|
SDValue True = N->getOperand(1);
|
|
|
|
SDValue False = N->getOperand(2);
|
|
|
|
|
2016-12-23 05:40:08 +08:00
|
|
|
if (Cond.hasOneUse()) { // TODO: Look for multiple select uses.
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
if ((DAG.isConstantValueOfAnyType(True) ||
|
|
|
|
DAG.isConstantValueOfAnyType(True)) &&
|
|
|
|
(!DAG.isConstantValueOfAnyType(False) &&
|
|
|
|
!DAG.isConstantValueOfAnyType(False))) {
|
|
|
|
// Swap cmp + select pair to move constant to false input.
|
|
|
|
// This will allow using VOPC cndmasks more often.
|
|
|
|
// select (setcc x, y), k, x -> select (setcc y, x) x, x
|
|
|
|
|
|
|
|
SDLoc SL(N);
|
|
|
|
ISD::CondCode NewCC = getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
|
|
|
|
LHS.getValueType().isInteger());
|
|
|
|
|
|
|
|
SDValue NewCond = DAG.getSetCC(SL, Cond.getValueType(), LHS, RHS, NewCC);
|
|
|
|
return DAG.getNode(ISD::SELECT, SL, VT, NewCond, False, True);
|
|
|
|
}
|
|
|
|
|
2017-02-01 08:42:40 +08:00
|
|
|
if (VT == MVT::f32 && Subtarget->hasFminFmaxLegacy()) {
|
|
|
|
SDValue MinMax
|
|
|
|
= combineFMinMaxLegacy(SDLoc(N), VT, LHS, RHS, True, False, CC, DCI);
|
|
|
|
// Revisit this node so we can catch min3/max3/med3 patterns.
|
|
|
|
//DCI.AddToWorklist(MinMax.getNode());
|
|
|
|
return MinMax;
|
|
|
|
}
|
2016-01-29 04:53:48 +08:00
|
|
|
}
|
2016-01-12 01:02:00 +08:00
|
|
|
|
|
|
|
// There's no reason to not do this if the condition has other uses.
|
2017-10-13 03:37:14 +08:00
|
|
|
return performCtlz_CttzCombine(SDLoc(N), Cond, True, False, DCI);
|
2016-01-12 01:02:00 +08:00
|
|
|
}
|
|
|
|
|
2017-02-03 08:23:15 +08:00
|
|
|
static bool isConstantFPZero(SDValue N) {
|
|
|
|
if (const ConstantFPSDNode *C = isConstOrConstSplatFP(N))
|
|
|
|
return C->isZero() && !C->isNegative();
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2017-02-03 08:51:50 +08:00
|
|
|
static unsigned inverseMinMax(unsigned Opc) {
|
|
|
|
switch (Opc) {
|
|
|
|
case ISD::FMAXNUM:
|
|
|
|
return ISD::FMINNUM;
|
|
|
|
case ISD::FMINNUM:
|
|
|
|
return ISD::FMAXNUM;
|
|
|
|
case AMDGPUISD::FMAX_LEGACY:
|
|
|
|
return AMDGPUISD::FMIN_LEGACY;
|
|
|
|
case AMDGPUISD::FMIN_LEGACY:
|
|
|
|
return AMDGPUISD::FMAX_LEGACY;
|
|
|
|
default:
|
|
|
|
llvm_unreachable("invalid min/max opcode");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-01-12 08:09:34 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDValue N0 = N->getOperand(0);
|
|
|
|
EVT VT = N->getValueType(0);
|
|
|
|
|
|
|
|
unsigned Opc = N0.getOpcode();
|
|
|
|
|
|
|
|
// If the input has multiple uses and we can either fold the negate down, or
|
|
|
|
// the other uses cannot, give up. This both prevents unprofitable
|
|
|
|
// transformations and infinite loops: we won't repeatedly try to fold around
|
|
|
|
// a negate that has no 'good' form.
|
2017-02-03 07:21:23 +08:00
|
|
|
if (N0.hasOneUse()) {
|
|
|
|
// This may be able to fold into the source, but at a code size cost. Don't
|
|
|
|
// fold if the fold into the user is free.
|
|
|
|
if (allUsesHaveSourceMods(N, 0))
|
|
|
|
return SDValue();
|
|
|
|
} else {
|
|
|
|
if (fnegFoldsIntoOp(Opc) &&
|
|
|
|
(allUsesHaveSourceMods(N) || !allUsesHaveSourceMods(N0.getNode())))
|
|
|
|
return SDValue();
|
|
|
|
}
|
2017-01-12 08:09:34 +08:00
|
|
|
|
|
|
|
SDLoc SL(N);
|
|
|
|
switch (Opc) {
|
|
|
|
case ISD::FADD: {
|
2017-01-19 14:35:27 +08:00
|
|
|
if (!mayIgnoreSignedZero(N0))
|
|
|
|
return SDValue();
|
|
|
|
|
2017-01-12 08:09:34 +08:00
|
|
|
// (fneg (fadd x, y)) -> (fadd (fneg x), (fneg y))
|
|
|
|
SDValue LHS = N0.getOperand(0);
|
|
|
|
SDValue RHS = N0.getOperand(1);
|
|
|
|
|
|
|
|
if (LHS.getOpcode() != ISD::FNEG)
|
|
|
|
LHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
|
|
|
|
else
|
|
|
|
LHS = LHS.getOperand(0);
|
|
|
|
|
|
|
|
if (RHS.getOpcode() != ISD::FNEG)
|
|
|
|
RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
|
|
|
|
else
|
|
|
|
RHS = RHS.getOperand(0);
|
|
|
|
|
2017-01-24 03:08:34 +08:00
|
|
|
SDValue Res = DAG.getNode(ISD::FADD, SL, VT, LHS, RHS, N0->getFlags());
|
2017-01-12 08:09:34 +08:00
|
|
|
if (!N0.hasOneUse())
|
|
|
|
DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
|
|
|
|
return Res;
|
|
|
|
}
|
2017-01-13 02:26:30 +08:00
|
|
|
case ISD::FMUL:
|
|
|
|
case AMDGPUISD::FMUL_LEGACY: {
|
2017-01-12 08:23:20 +08:00
|
|
|
// (fneg (fmul x, y)) -> (fmul x, (fneg y))
|
2017-01-13 02:26:30 +08:00
|
|
|
// (fneg (fmul_legacy x, y)) -> (fmul_legacy x, (fneg y))
|
2017-01-12 08:23:20 +08:00
|
|
|
SDValue LHS = N0.getOperand(0);
|
|
|
|
SDValue RHS = N0.getOperand(1);
|
|
|
|
|
|
|
|
if (LHS.getOpcode() == ISD::FNEG)
|
|
|
|
LHS = LHS.getOperand(0);
|
|
|
|
else if (RHS.getOpcode() == ISD::FNEG)
|
|
|
|
RHS = RHS.getOperand(0);
|
|
|
|
else
|
|
|
|
RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
|
|
|
|
|
2017-01-24 03:08:34 +08:00
|
|
|
SDValue Res = DAG.getNode(Opc, SL, VT, LHS, RHS, N0->getFlags());
|
2017-01-12 08:23:20 +08:00
|
|
|
if (!N0.hasOneUse())
|
|
|
|
DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
|
|
|
|
return Res;
|
|
|
|
}
|
2017-01-12 08:32:16 +08:00
|
|
|
case ISD::FMA:
|
|
|
|
case ISD::FMAD: {
|
2017-01-19 14:35:27 +08:00
|
|
|
if (!mayIgnoreSignedZero(N0))
|
|
|
|
return SDValue();
|
|
|
|
|
2017-01-12 08:32:16 +08:00
|
|
|
// (fneg (fma x, y, z)) -> (fma x, (fneg y), (fneg z))
|
|
|
|
SDValue LHS = N0.getOperand(0);
|
|
|
|
SDValue MHS = N0.getOperand(1);
|
|
|
|
SDValue RHS = N0.getOperand(2);
|
|
|
|
|
|
|
|
if (LHS.getOpcode() == ISD::FNEG)
|
|
|
|
LHS = LHS.getOperand(0);
|
|
|
|
else if (MHS.getOpcode() == ISD::FNEG)
|
|
|
|
MHS = MHS.getOperand(0);
|
|
|
|
else
|
|
|
|
MHS = DAG.getNode(ISD::FNEG, SL, VT, MHS);
|
|
|
|
|
|
|
|
if (RHS.getOpcode() != ISD::FNEG)
|
|
|
|
RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
|
|
|
|
else
|
|
|
|
RHS = RHS.getOperand(0);
|
|
|
|
|
|
|
|
SDValue Res = DAG.getNode(Opc, SL, VT, LHS, MHS, RHS);
|
|
|
|
if (!N0.hasOneUse())
|
|
|
|
DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
|
|
|
|
return Res;
|
|
|
|
}
|
2017-02-03 08:23:15 +08:00
|
|
|
case ISD::FMAXNUM:
|
2017-02-03 08:51:50 +08:00
|
|
|
case ISD::FMINNUM:
|
|
|
|
case AMDGPUISD::FMAX_LEGACY:
|
|
|
|
case AMDGPUISD::FMIN_LEGACY: {
|
2017-02-03 08:23:15 +08:00
|
|
|
// fneg (fmaxnum x, y) -> fminnum (fneg x), (fneg y)
|
|
|
|
// fneg (fminnum x, y) -> fmaxnum (fneg x), (fneg y)
|
2017-02-03 08:51:50 +08:00
|
|
|
// fneg (fmax_legacy x, y) -> fmin_legacy (fneg x), (fneg y)
|
|
|
|
// fneg (fmin_legacy x, y) -> fmax_legacy (fneg x), (fneg y)
|
|
|
|
|
2017-02-03 08:23:15 +08:00
|
|
|
SDValue LHS = N0.getOperand(0);
|
|
|
|
SDValue RHS = N0.getOperand(1);
|
|
|
|
|
|
|
|
// 0 doesn't have a negated inline immediate.
|
|
|
|
// TODO: Shouldn't fold 1/2pi either, and should be generalized to other
|
|
|
|
// operations.
|
|
|
|
if (isConstantFPZero(RHS))
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
SDValue NegLHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
|
|
|
|
SDValue NegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
|
2017-02-03 08:51:50 +08:00
|
|
|
unsigned Opposite = inverseMinMax(Opc);
|
2017-02-03 08:23:15 +08:00
|
|
|
|
|
|
|
SDValue Res = DAG.getNode(Opposite, SL, VT, NegLHS, NegRHS, N0->getFlags());
|
|
|
|
if (!N0.hasOneUse())
|
|
|
|
DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
|
|
|
|
return Res;
|
|
|
|
}
|
2017-01-13 01:46:35 +08:00
|
|
|
case ISD::FP_EXTEND:
|
2017-01-26 09:25:36 +08:00
|
|
|
case ISD::FTRUNC:
|
|
|
|
case ISD::FRINT:
|
|
|
|
case ISD::FNEARBYINT: // XXX - Should fround be handled?
|
|
|
|
case ISD::FSIN:
|
2017-01-13 01:46:35 +08:00
|
|
|
case AMDGPUISD::RCP:
|
2017-01-13 02:48:09 +08:00
|
|
|
case AMDGPUISD::RCP_LEGACY:
|
2018-06-27 23:33:33 +08:00
|
|
|
case AMDGPUISD::RCP_IFLAG:
|
2017-01-13 02:48:09 +08:00
|
|
|
case AMDGPUISD::SIN_HW: {
|
2017-01-13 01:46:28 +08:00
|
|
|
SDValue CvtSrc = N0.getOperand(0);
|
|
|
|
if (CvtSrc.getOpcode() == ISD::FNEG) {
|
|
|
|
// (fneg (fp_extend (fneg x))) -> (fp_extend x)
|
2017-01-13 01:46:35 +08:00
|
|
|
// (fneg (rcp (fneg x))) -> (rcp x)
|
2017-01-13 01:46:33 +08:00
|
|
|
return DAG.getNode(Opc, SL, VT, CvtSrc.getOperand(0));
|
2017-01-13 01:46:28 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!N0.hasOneUse())
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
// (fneg (fp_extend x)) -> (fp_extend (fneg x))
|
2017-01-13 01:46:35 +08:00
|
|
|
// (fneg (rcp x)) -> (rcp (fneg x))
|
2017-01-13 01:46:28 +08:00
|
|
|
SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
|
2017-01-24 03:08:34 +08:00
|
|
|
return DAG.getNode(Opc, SL, VT, Neg, N0->getFlags());
|
2017-01-13 01:46:33 +08:00
|
|
|
}
|
|
|
|
case ISD::FP_ROUND: {
|
|
|
|
SDValue CvtSrc = N0.getOperand(0);
|
|
|
|
|
|
|
|
if (CvtSrc.getOpcode() == ISD::FNEG) {
|
|
|
|
// (fneg (fp_round (fneg x))) -> (fp_round x)
|
|
|
|
return DAG.getNode(ISD::FP_ROUND, SL, VT,
|
|
|
|
CvtSrc.getOperand(0), N0.getOperand(1));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!N0.hasOneUse())
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
// (fneg (fp_round x)) -> (fp_round (fneg x))
|
|
|
|
SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
|
|
|
|
return DAG.getNode(ISD::FP_ROUND, SL, VT, Neg, N0.getOperand(1));
|
2017-01-13 01:46:28 +08:00
|
|
|
}
|
2017-02-02 10:27:04 +08:00
|
|
|
case ISD::FP16_TO_FP: {
|
|
|
|
// v_cvt_f32_f16 supports source modifiers on pre-VI targets without legal
|
|
|
|
// f16, but legalization of f16 fneg ends up pulling it out of the source.
|
|
|
|
// Put the fneg back as a legal source operation that can be matched later.
|
|
|
|
SDLoc SL(N);
|
|
|
|
|
|
|
|
SDValue Src = N0.getOperand(0);
|
|
|
|
EVT SrcVT = Src.getValueType();
|
|
|
|
|
|
|
|
// fneg (fp16_to_fp x) -> fp16_to_fp (xor x, 0x8000)
|
|
|
|
SDValue IntFNeg = DAG.getNode(ISD::XOR, SL, SrcVT, Src,
|
|
|
|
DAG.getConstant(0x8000, SL, SrcVT));
|
|
|
|
return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFNeg);
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::performFAbsCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDValue N0 = N->getOperand(0);
|
|
|
|
|
|
|
|
if (!N0.hasOneUse())
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
switch (N0.getOpcode()) {
|
|
|
|
case ISD::FP16_TO_FP: {
|
|
|
|
assert(!Subtarget->has16BitInsts() && "should only see if f16 is illegal");
|
|
|
|
SDLoc SL(N);
|
|
|
|
SDValue Src = N0.getOperand(0);
|
|
|
|
EVT SrcVT = Src.getValueType();
|
|
|
|
|
|
|
|
// fabs (fp16_to_fp x) -> fp16_to_fp (and x, 0x7fff)
|
|
|
|
SDValue IntFAbs = DAG.getNode(ISD::AND, SL, SrcVT, Src,
|
|
|
|
DAG.getConstant(0x7fff, SL, SrcVT));
|
|
|
|
return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFAbs);
|
|
|
|
}
|
2017-01-12 08:09:34 +08:00
|
|
|
default:
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-06-27 23:33:33 +08:00
|
|
|
SDValue AMDGPUTargetLowering::performRcpCombine(SDNode *N,
|
|
|
|
DAGCombinerInfo &DCI) const {
|
|
|
|
const auto *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0));
|
|
|
|
if (!CFP)
|
|
|
|
return SDValue();
|
|
|
|
|
|
|
|
// XXX - Should this flush denormals?
|
|
|
|
const APFloat &Val = CFP->getValueAPF();
|
|
|
|
APFloat One(Val.getSemantics(), "1.0");
|
|
|
|
return DCI.DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
|
|
|
|
}
|
|
|
|
|
2014-04-08 03:45:41 +08:00
|
|
|
SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
|
2014-07-15 10:06:31 +08:00
|
|
|
DAGCombinerInfo &DCI) const {
|
2014-04-08 03:45:41 +08:00
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
|
|
|
SDLoc DL(N);
|
|
|
|
|
|
|
|
switch(N->getOpcode()) {
|
2015-07-04 07:33:38 +08:00
|
|
|
default:
|
|
|
|
break;
|
2016-04-15 05:58:07 +08:00
|
|
|
case ISD::BITCAST: {
|
|
|
|
EVT DestVT = N->getValueType(0);
|
2016-09-17 23:44:16 +08:00
|
|
|
|
|
|
|
// Push casts through vector builds. This helps avoid emitting a large
|
|
|
|
// number of copies when materializing floating point vector constants.
|
|
|
|
//
|
|
|
|
// vNt1 bitcast (vNt0 (build_vector t0:x, t0:y)) =>
|
|
|
|
// vnt1 = build_vector (t1 (bitcast t0:x)), (t1 (bitcast t0:y))
|
|
|
|
if (DestVT.isVector()) {
|
|
|
|
SDValue Src = N->getOperand(0);
|
|
|
|
if (Src.getOpcode() == ISD::BUILD_VECTOR) {
|
|
|
|
EVT SrcVT = Src.getValueType();
|
|
|
|
unsigned NElts = DestVT.getVectorNumElements();
|
|
|
|
|
|
|
|
if (SrcVT.getVectorNumElements() == NElts) {
|
|
|
|
EVT DestEltVT = DestVT.getVectorElementType();
|
|
|
|
|
|
|
|
SmallVector<SDValue, 8> CastedElts;
|
|
|
|
SDLoc SL(N);
|
|
|
|
for (unsigned I = 0, E = SrcVT.getVectorNumElements(); I != E; ++I) {
|
|
|
|
SDValue Elt = Src.getOperand(I);
|
|
|
|
CastedElts.push_back(DAG.getNode(ISD::BITCAST, DL, DestEltVT, Elt));
|
|
|
|
}
|
|
|
|
|
|
|
|
return DAG.getBuildVector(DestVT, SL, CastedElts);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-04-15 05:58:07 +08:00
|
|
|
if (DestVT.getSizeInBits() != 64 && !DestVT.isVector())
|
|
|
|
break;
|
|
|
|
|
|
|
|
// Fold bitcasts of constants.
|
|
|
|
//
|
|
|
|
// v2i32 (bitcast i64:k) -> build_vector lo_32(k), hi_32(k)
|
|
|
|
// TODO: Generalize and move to DAGCombiner
|
|
|
|
SDValue Src = N->getOperand(0);
|
|
|
|
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) {
|
2018-05-22 14:32:10 +08:00
|
|
|
if (Src.getValueType() == MVT::i64) {
|
|
|
|
SDLoc SL(N);
|
|
|
|
uint64_t CVal = C->getZExtValue();
|
|
|
|
return DAG.getNode(ISD::BUILD_VECTOR, SL, DestVT,
|
|
|
|
DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
|
|
|
|
DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
|
|
|
|
}
|
2016-04-15 05:58:07 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Src)) {
|
|
|
|
const APInt &Val = C->getValueAPF().bitcastToAPInt();
|
|
|
|
SDLoc SL(N);
|
|
|
|
uint64_t CVal = Val.getZExtValue();
|
|
|
|
SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
|
|
|
|
DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
|
|
|
|
DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
|
|
|
|
|
|
|
|
return DAG.getNode(ISD::BITCAST, SL, DestVT, Vec);
|
|
|
|
}
|
|
|
|
|
|
|
|
break;
|
|
|
|
}
|
2015-07-15 02:20:33 +08:00
|
|
|
case ISD::SHL: {
|
|
|
|
if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
|
|
|
|
break;
|
|
|
|
|
|
|
|
return performShlCombine(N, DCI);
|
|
|
|
}
|
2016-01-19 05:43:36 +08:00
|
|
|
case ISD::SRL: {
|
|
|
|
if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
|
|
|
|
break;
|
|
|
|
|
|
|
|
return performSrlCombine(N, DCI);
|
|
|
|
}
|
2016-01-19 06:09:04 +08:00
|
|
|
case ISD::SRA: {
|
|
|
|
if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
|
|
|
|
break;
|
|
|
|
|
|
|
|
return performSraCombine(N, DCI);
|
|
|
|
}
|
2018-05-10 02:37:39 +08:00
|
|
|
case ISD::TRUNCATE:
|
|
|
|
return performTruncateCombine(N, DCI);
|
2015-07-04 07:33:38 +08:00
|
|
|
case ISD::MUL:
|
|
|
|
return performMulCombine(N, DCI);
|
2016-08-27 09:32:27 +08:00
|
|
|
case ISD::MULHS:
|
|
|
|
return performMulhsCombine(N, DCI);
|
|
|
|
case ISD::MULHU:
|
|
|
|
return performMulhuCombine(N, DCI);
|
2015-07-04 07:33:38 +08:00
|
|
|
case AMDGPUISD::MUL_I24:
|
2016-08-27 09:32:27 +08:00
|
|
|
case AMDGPUISD::MUL_U24:
|
|
|
|
case AMDGPUISD::MULHI_I24:
|
|
|
|
case AMDGPUISD::MULHI_U24: {
|
2016-10-22 04:25:11 +08:00
|
|
|
// If the first call to simplify is successfull, then N may end up being
|
|
|
|
// deleted, so we shouldn't call simplifyI24 again.
|
|
|
|
simplifyI24(N, 0, DCI) || simplifyI24(N, 1, DCI);
|
2015-07-04 07:33:38 +08:00
|
|
|
return SDValue();
|
|
|
|
}
|
2016-08-27 09:32:27 +08:00
|
|
|
case AMDGPUISD::MUL_LOHI_I24:
|
|
|
|
case AMDGPUISD::MUL_LOHI_U24:
|
|
|
|
return performMulLoHi24Combine(N, DCI);
|
2016-01-12 01:02:00 +08:00
|
|
|
case ISD::SELECT:
|
|
|
|
return performSelectCombine(N, DCI);
|
2017-01-12 08:09:34 +08:00
|
|
|
case ISD::FNEG:
|
|
|
|
return performFNegCombine(N, DCI);
|
2017-02-02 10:27:04 +08:00
|
|
|
case ISD::FABS:
|
|
|
|
return performFAbsCombine(N, DCI);
|
2014-05-23 02:09:07 +08:00
|
|
|
case AMDGPUISD::BFE_I32:
|
|
|
|
case AMDGPUISD::BFE_U32: {
|
|
|
|
assert(!N->getValueType(0).isVector() &&
|
|
|
|
"Vector handling of BFE not implemented");
|
|
|
|
ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
|
|
|
|
if (!Width)
|
|
|
|
break;
|
|
|
|
|
|
|
|
uint32_t WidthVal = Width->getZExtValue() & 0x1f;
|
|
|
|
if (WidthVal == 0)
|
2015-04-28 22:05:47 +08:00
|
|
|
return DAG.getConstant(0, DL, MVT::i32);
|
2014-05-23 02:09:07 +08:00
|
|
|
|
|
|
|
ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
|
|
|
|
if (!Offset)
|
|
|
|
break;
|
|
|
|
|
|
|
|
SDValue BitsFrom = N->getOperand(0);
|
|
|
|
uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
|
|
|
|
|
|
|
|
bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
|
|
|
|
|
|
|
|
if (OffsetVal == 0) {
|
|
|
|
// This is already sign / zero extended, so try to fold away extra BFEs.
|
|
|
|
unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
|
|
|
|
|
|
|
|
unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
|
|
|
|
if (OpSignBits >= SignBits)
|
|
|
|
return BitsFrom;
|
2014-05-23 02:09:12 +08:00
|
|
|
|
|
|
|
EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
|
|
|
|
if (Signed) {
|
|
|
|
// This is a sign_extend_inreg. Replace it to take advantage of existing
|
|
|
|
// DAG Combines. If not eliminated, we will match back to BFE during
|
|
|
|
// selection.
|
|
|
|
|
|
|
|
// TODO: The sext_inreg of extended types ends, although we can could
|
|
|
|
// handle them in a single BFE.
|
|
|
|
return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
|
|
|
|
DAG.getValueType(SmallVT));
|
|
|
|
}
|
|
|
|
|
|
|
|
return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
|
2014-05-23 02:09:07 +08:00
|
|
|
}
|
|
|
|
|
2014-10-15 13:07:00 +08:00
|
|
|
if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
|
2014-05-23 02:09:07 +08:00
|
|
|
if (Signed) {
|
|
|
|
return constantFoldBFE<int32_t>(DAG,
|
2014-09-19 08:42:06 +08:00
|
|
|
CVal->getSExtValue(),
|
2014-05-23 02:09:07 +08:00
|
|
|
OffsetVal,
|
2015-04-28 22:05:47 +08:00
|
|
|
WidthVal,
|
|
|
|
DL);
|
2014-05-23 02:09:07 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
return constantFoldBFE<uint32_t>(DAG,
|
2014-09-18 23:52:26 +08:00
|
|
|
CVal->getZExtValue(),
|
2014-05-23 02:09:07 +08:00
|
|
|
OffsetVal,
|
2015-04-28 22:05:47 +08:00
|
|
|
WidthVal,
|
|
|
|
DL);
|
2014-05-23 02:09:07 +08:00
|
|
|
}
|
|
|
|
|
2017-05-24 03:54:48 +08:00
|
|
|
if ((OffsetVal + WidthVal) >= 32 &&
|
|
|
|
!(Subtarget->hasSDWA() && OffsetVal == 16 && WidthVal == 16)) {
|
2015-04-28 22:05:47 +08:00
|
|
|
SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32);
|
2014-05-23 02:09:12 +08:00
|
|
|
return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
|
|
|
|
BitsFrom, ShiftVal);
|
|
|
|
}
|
|
|
|
|
2014-10-16 01:58:34 +08:00
|
|
|
if (BitsFrom.hasOneUse()) {
|
2014-10-16 07:37:42 +08:00
|
|
|
APInt Demanded = APInt::getBitsSet(32,
|
|
|
|
OffsetVal,
|
|
|
|
OffsetVal + WidthVal);
|
|
|
|
|
2017-04-28 13:31:46 +08:00
|
|
|
KnownBits Known;
|
2014-10-16 01:58:34 +08:00
|
|
|
TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
|
|
|
|
!DCI.isBeforeLegalizeOps());
|
|
|
|
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
|
2017-04-22 02:53:12 +08:00
|
|
|
if (TLI.ShrinkDemandedConstant(BitsFrom, Demanded, TLO) ||
|
2017-04-28 13:31:46 +08:00
|
|
|
TLI.SimplifyDemandedBits(BitsFrom, Demanded, Known, TLO)) {
|
2014-10-16 01:58:34 +08:00
|
|
|
DCI.CommitTargetLoweringOpt(TLO);
|
|
|
|
}
|
2014-05-23 02:09:07 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
break;
|
|
|
|
}
|
2016-07-02 06:47:50 +08:00
|
|
|
case ISD::LOAD:
|
|
|
|
return performLoadCombine(N, DCI);
|
2014-07-15 10:06:31 +08:00
|
|
|
case ISD::STORE:
|
|
|
|
return performStoreCombine(N, DCI);
|
2018-06-27 23:33:33 +08:00
|
|
|
case AMDGPUISD::RCP:
|
|
|
|
case AMDGPUISD::RCP_IFLAG:
|
|
|
|
return performRcpCombine(N, DCI);
|
2017-07-15 13:52:59 +08:00
|
|
|
case ISD::AssertZext:
|
|
|
|
case ISD::AssertSext:
|
|
|
|
return performAssertSZExtCombine(N, DCI);
|
2014-04-08 03:45:41 +08:00
|
|
|
}
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Helper functions
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
|
2017-06-20 05:52:45 +08:00
|
|
|
const TargetRegisterClass *RC,
|
|
|
|
unsigned Reg, EVT VT,
|
|
|
|
const SDLoc &SL,
|
|
|
|
bool RawReg) const {
|
2012-12-12 05:25:42 +08:00
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
|
|
|
MachineRegisterInfo &MRI = MF.getRegInfo();
|
2017-06-20 05:52:45 +08:00
|
|
|
unsigned VReg;
|
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
if (!MRI.isLiveIn(Reg)) {
|
2017-06-20 05:52:45 +08:00
|
|
|
VReg = MRI.createVirtualRegister(RC);
|
|
|
|
MRI.addLiveIn(Reg, VReg);
|
2012-12-12 05:25:42 +08:00
|
|
|
} else {
|
2017-06-20 05:52:45 +08:00
|
|
|
VReg = MRI.getLiveInVirtReg(Reg);
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
2017-06-20 05:52:45 +08:00
|
|
|
|
|
|
|
if (RawReg)
|
|
|
|
return DAG.getRegister(VReg, VT);
|
|
|
|
|
|
|
|
return DAG.getCopyFromReg(DAG.getEntryNode(), SL, VReg, VT);
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
|
|
|
|
2017-08-04 07:00:29 +08:00
|
|
|
SDValue AMDGPUTargetLowering::loadStackInputValue(SelectionDAG &DAG,
|
|
|
|
EVT VT,
|
|
|
|
const SDLoc &SL,
|
|
|
|
int64_t Offset) const {
|
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
|
|
|
MachineFrameInfo &MFI = MF.getFrameInfo();
|
|
|
|
|
|
|
|
int FI = MFI.CreateFixedObject(VT.getStoreSize(), Offset, true);
|
|
|
|
auto SrcPtrInfo = MachinePointerInfo::getStack(MF, Offset);
|
|
|
|
SDValue Ptr = DAG.getFrameIndex(FI, MVT::i32);
|
|
|
|
|
|
|
|
return DAG.getLoad(VT, SL, DAG.getEntryNode(), Ptr, SrcPtrInfo, 4,
|
|
|
|
MachineMemOperand::MODereferenceable |
|
|
|
|
MachineMemOperand::MOInvariant);
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::storeStackInputValue(SelectionDAG &DAG,
|
|
|
|
const SDLoc &SL,
|
|
|
|
SDValue Chain,
|
|
|
|
SDValue StackPtr,
|
|
|
|
SDValue ArgVal,
|
|
|
|
int64_t Offset) const {
|
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
|
|
|
MachinePointerInfo DstInfo = MachinePointerInfo::getStack(MF, Offset);
|
|
|
|
|
2017-11-29 09:25:12 +08:00
|
|
|
SDValue Ptr = DAG.getObjectPtrOffset(SL, StackPtr, Offset);
|
2017-08-04 07:00:29 +08:00
|
|
|
SDValue Store = DAG.getStore(Chain, SL, ArgVal, Ptr, DstInfo, 4,
|
|
|
|
MachineMemOperand::MODereferenceable);
|
|
|
|
return Store;
|
|
|
|
}
|
|
|
|
|
|
|
|
SDValue AMDGPUTargetLowering::loadInputValue(SelectionDAG &DAG,
|
|
|
|
const TargetRegisterClass *RC,
|
|
|
|
EVT VT, const SDLoc &SL,
|
|
|
|
const ArgDescriptor &Arg) const {
|
|
|
|
assert(Arg && "Attempting to load missing argument");
|
|
|
|
|
|
|
|
if (Arg.isRegister())
|
|
|
|
return CreateLiveInRegister(DAG, RC, Arg.getRegister(), VT, SL);
|
|
|
|
return loadStackInputValue(DAG, VT, SL, Arg.getStackOffset());
|
|
|
|
}
|
|
|
|
|
2015-07-10 05:20:37 +08:00
|
|
|
uint32_t AMDGPUTargetLowering::getImplicitParameterOffset(
|
2018-06-28 18:18:55 +08:00
|
|
|
const MachineFunction &MF, const ImplicitParameter Param) const {
|
|
|
|
const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
|
|
|
|
const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>();
|
|
|
|
unsigned ExplicitArgOffset = ST.getExplicitKernelArgOffset(MF.getFunction());
|
|
|
|
unsigned Alignment = ST.getAlignmentForImplicitArgPtr();
|
|
|
|
uint64_t ArgOffset = alignTo(MFI->getExplicitKernArgSize(), Alignment) +
|
|
|
|
ExplicitArgOffset;
|
2015-07-10 05:20:37 +08:00
|
|
|
switch (Param) {
|
|
|
|
case GRID_DIM:
|
|
|
|
return ArgOffset;
|
|
|
|
case GRID_OFFSET:
|
|
|
|
return ArgOffset + 4;
|
|
|
|
}
|
|
|
|
llvm_unreachable("unexpected implicit parameter type");
|
|
|
|
}
|
|
|
|
|
2012-12-12 05:25:42 +08:00
|
|
|
#define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
|
|
|
|
|
|
|
|
const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
|
2015-05-08 05:33:59 +08:00
|
|
|
switch ((AMDGPUISD::NodeType)Opcode) {
|
|
|
|
case AMDGPUISD::FIRST_NUMBER: break;
|
2012-12-12 05:25:42 +08:00
|
|
|
// AMDIL DAG nodes
|
|
|
|
NODE_NAME_CASE(UMUL);
|
|
|
|
NODE_NAME_CASE(BRANCH_COND);
|
|
|
|
|
|
|
|
// AMDGPU DAG nodes
|
2017-03-18 04:41:45 +08:00
|
|
|
NODE_NAME_CASE(IF)
|
|
|
|
NODE_NAME_CASE(ELSE)
|
|
|
|
NODE_NAME_CASE(LOOP)
|
2017-03-22 06:18:10 +08:00
|
|
|
NODE_NAME_CASE(CALL)
|
2017-08-12 04:42:08 +08:00
|
|
|
NODE_NAME_CASE(TC_RETURN)
|
2017-04-25 01:49:13 +08:00
|
|
|
NODE_NAME_CASE(TRAP)
|
2017-03-22 06:18:10 +08:00
|
|
|
NODE_NAME_CASE(RET_FLAG)
|
|
|
|
NODE_NAME_CASE(RETURN_TO_EPILOG)
|
2016-06-23 04:15:28 +08:00
|
|
|
NODE_NAME_CASE(ENDPGM)
|
2012-12-12 05:25:42 +08:00
|
|
|
NODE_NAME_CASE(DWORDADDR)
|
|
|
|
NODE_NAME_CASE(FRACT)
|
2016-07-29 00:42:13 +08:00
|
|
|
NODE_NAME_CASE(SETCC)
|
2016-12-07 10:42:15 +08:00
|
|
|
NODE_NAME_CASE(SETREG)
|
|
|
|
NODE_NAME_CASE(FMA_W_CHAIN)
|
|
|
|
NODE_NAME_CASE(FMUL_W_CHAIN)
|
2014-06-13 05:15:44 +08:00
|
|
|
NODE_NAME_CASE(CLAMP)
|
2015-05-08 05:33:59 +08:00
|
|
|
NODE_NAME_CASE(COS_HW)
|
|
|
|
NODE_NAME_CASE(SIN_HW)
|
2014-11-14 07:03:09 +08:00
|
|
|
NODE_NAME_CASE(FMAX_LEGACY)
|
|
|
|
NODE_NAME_CASE(FMIN_LEGACY)
|
2014-11-15 04:08:52 +08:00
|
|
|
NODE_NAME_CASE(FMAX3)
|
|
|
|
NODE_NAME_CASE(SMAX3)
|
|
|
|
NODE_NAME_CASE(UMAX3)
|
|
|
|
NODE_NAME_CASE(FMIN3)
|
|
|
|
NODE_NAME_CASE(SMIN3)
|
|
|
|
NODE_NAME_CASE(UMIN3)
|
2016-01-29 04:53:42 +08:00
|
|
|
NODE_NAME_CASE(FMED3)
|
|
|
|
NODE_NAME_CASE(SMED3)
|
|
|
|
NODE_NAME_CASE(UMED3)
|
2014-06-19 09:19:19 +08:00
|
|
|
NODE_NAME_CASE(URECIP)
|
|
|
|
NODE_NAME_CASE(DIV_SCALE)
|
|
|
|
NODE_NAME_CASE(DIV_FMAS)
|
|
|
|
NODE_NAME_CASE(DIV_FIXUP)
|
2017-02-25 07:00:29 +08:00
|
|
|
NODE_NAME_CASE(FMAD_FTZ)
|
2014-06-19 09:19:19 +08:00
|
|
|
NODE_NAME_CASE(TRIG_PREOP)
|
|
|
|
NODE_NAME_CASE(RCP)
|
|
|
|
NODE_NAME_CASE(RSQ)
|
2016-07-27 00:45:45 +08:00
|
|
|
NODE_NAME_CASE(RCP_LEGACY)
|
2014-06-25 06:13:39 +08:00
|
|
|
NODE_NAME_CASE(RSQ_LEGACY)
|
2018-06-27 23:33:33 +08:00
|
|
|
NODE_NAME_CASE(RCP_IFLAG)
|
2016-07-27 00:45:45 +08:00
|
|
|
NODE_NAME_CASE(FMUL_LEGACY)
|
2016-02-13 09:03:00 +08:00
|
|
|
NODE_NAME_CASE(RSQ_CLAMP)
|
2014-08-16 01:30:25 +08:00
|
|
|
NODE_NAME_CASE(LDEXP)
|
2015-01-07 07:00:37 +08:00
|
|
|
NODE_NAME_CASE(FP_CLASS)
|
2014-06-19 09:19:19 +08:00
|
|
|
NODE_NAME_CASE(DOT4)
|
2015-05-08 05:33:59 +08:00
|
|
|
NODE_NAME_CASE(CARRY)
|
|
|
|
NODE_NAME_CASE(BORROW)
|
2014-03-18 02:58:11 +08:00
|
|
|
NODE_NAME_CASE(BFE_U32)
|
|
|
|
NODE_NAME_CASE(BFE_I32)
|
2014-04-01 02:21:13 +08:00
|
|
|
NODE_NAME_CASE(BFI)
|
|
|
|
NODE_NAME_CASE(BFM)
|
2016-01-12 01:02:00 +08:00
|
|
|
NODE_NAME_CASE(FFBH_U32)
|
2016-07-19 02:40:51 +08:00
|
|
|
NODE_NAME_CASE(FFBH_I32)
|
2017-10-13 03:37:14 +08:00
|
|
|
NODE_NAME_CASE(FFBL_B32)
|
2014-04-08 03:45:41 +08:00
|
|
|
NODE_NAME_CASE(MUL_U24)
|
|
|
|
NODE_NAME_CASE(MUL_I24)
|
2016-08-27 09:32:27 +08:00
|
|
|
NODE_NAME_CASE(MULHI_U24)
|
|
|
|
NODE_NAME_CASE(MULHI_I24)
|
|
|
|
NODE_NAME_CASE(MUL_LOHI_U24)
|
|
|
|
NODE_NAME_CASE(MUL_LOHI_I24)
|
2014-05-23 02:00:15 +08:00
|
|
|
NODE_NAME_CASE(MAD_U24)
|
|
|
|
NODE_NAME_CASE(MAD_I24)
|
2017-11-07 01:04:37 +08:00
|
|
|
NODE_NAME_CASE(MAD_I64_I32)
|
|
|
|
NODE_NAME_CASE(MAD_U64_U32)
|
2018-06-13 07:50:37 +08:00
|
|
|
NODE_NAME_CASE(PERM)
|
2015-05-08 05:33:59 +08:00
|
|
|
NODE_NAME_CASE(TEXTURE_FETCH)
|
2012-12-12 05:25:42 +08:00
|
|
|
NODE_NAME_CASE(EXPORT)
|
2016-12-06 04:23:10 +08:00
|
|
|
NODE_NAME_CASE(EXPORT_DONE)
|
|
|
|
NODE_NAME_CASE(R600_EXPORT)
|
2013-01-23 10:09:03 +08:00
|
|
|
NODE_NAME_CASE(CONST_ADDRESS)
|
2013-02-07 01:32:29 +08:00
|
|
|
NODE_NAME_CASE(REGISTER_LOAD)
|
|
|
|
NODE_NAME_CASE(REGISTER_STORE)
|
2013-08-15 07:24:45 +08:00
|
|
|
NODE_NAME_CASE(SAMPLE)
|
|
|
|
NODE_NAME_CASE(SAMPLEB)
|
|
|
|
NODE_NAME_CASE(SAMPLED)
|
|
|
|
NODE_NAME_CASE(SAMPLEL)
|
2014-06-12 01:50:44 +08:00
|
|
|
NODE_NAME_CASE(CVT_F32_UBYTE0)
|
|
|
|
NODE_NAME_CASE(CVT_F32_UBYTE1)
|
|
|
|
NODE_NAME_CASE(CVT_F32_UBYTE2)
|
|
|
|
NODE_NAME_CASE(CVT_F32_UBYTE3)
|
2017-02-22 08:27:34 +08:00
|
|
|
NODE_NAME_CASE(CVT_PKRTZ_F16_F32)
|
2018-02-01 04:18:04 +08:00
|
|
|
NODE_NAME_CASE(CVT_PKNORM_I16_F32)
|
|
|
|
NODE_NAME_CASE(CVT_PKNORM_U16_F32)
|
|
|
|
NODE_NAME_CASE(CVT_PK_I16_I32)
|
|
|
|
NODE_NAME_CASE(CVT_PK_U16_U32)
|
2017-03-16 03:04:26 +08:00
|
|
|
NODE_NAME_CASE(FP_TO_FP16)
|
2017-04-01 03:53:03 +08:00
|
|
|
NODE_NAME_CASE(FP16_ZEXT)
|
2014-06-18 00:53:14 +08:00
|
|
|
NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
|
2014-07-21 22:01:14 +08:00
|
|
|
NODE_NAME_CASE(CONST_DATA_PTR)
|
2016-06-15 04:29:59 +08:00
|
|
|
NODE_NAME_CASE(PC_ADD_REL_OFFSET)
|
2016-07-20 00:27:56 +08:00
|
|
|
NODE_NAME_CASE(KILL)
|
2017-01-21 05:24:26 +08:00
|
|
|
NODE_NAME_CASE(DUMMY_CHAIN)
|
2015-05-08 05:33:59 +08:00
|
|
|
case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break;
|
AMDGPU: Add new amdgcn.init.exec intrinsics
v2: More tests, bug fixes, cosmetic changes.
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D31762
llvm-svn: 301677
2017-04-29 04:21:58 +08:00
|
|
|
NODE_NAME_CASE(INIT_EXEC)
|
|
|
|
NODE_NAME_CASE(INIT_EXEC_FROM_INPUT)
|
2015-05-12 22:18:14 +08:00
|
|
|
NODE_NAME_CASE(SENDMSG)
|
2017-01-05 02:06:55 +08:00
|
|
|
NODE_NAME_CASE(SENDMSGHALT)
|
2015-05-12 23:00:46 +08:00
|
|
|
NODE_NAME_CASE(INTERP_MOV)
|
|
|
|
NODE_NAME_CASE(INTERP_P1)
|
|
|
|
NODE_NAME_CASE(INTERP_P2)
|
2013-08-16 09:12:06 +08:00
|
|
|
NODE_NAME_CASE(STORE_MSKOR)
|
2016-04-26 03:27:09 +08:00
|
|
|
NODE_NAME_CASE(LOAD_CONSTANT)
|
2013-09-12 10:55:14 +08:00
|
|
|
NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
|
2017-06-23 00:29:22 +08:00
|
|
|
NODE_NAME_CASE(TBUFFER_STORE_FORMAT_X3)
|
2018-01-13 05:12:19 +08:00
|
|
|
NODE_NAME_CASE(TBUFFER_STORE_FORMAT_D16)
|
2017-06-23 00:29:22 +08:00
|
|
|
NODE_NAME_CASE(TBUFFER_LOAD_FORMAT)
|
2018-01-13 05:12:19 +08:00
|
|
|
NODE_NAME_CASE(TBUFFER_LOAD_FORMAT_D16)
|
AMDGPU: Implement {BUFFER,FLAT}_ATOMIC_CMPSWAP{,_X2}
Summary:
Implement BUFFER_ATOMIC_CMPSWAP{,_X2} instructions on all GCN targets, and FLAT_ATOMIC_CMPSWAP{,_X2} on CI+.
32-bit instruction variants tested manually on Kabini and Bonaire. Tests and parts of code provided by Jan Veselý.
Patch by: Vedran Miletić
Reviewers: arsenm, tstellarAMD, nhaehnle
Subscribers: jvesely, scchan, kanarayan, arsenm
Differential Revision: http://reviews.llvm.org/D17280
llvm-svn: 265170
2016-04-02 02:27:37 +08:00
|
|
|
NODE_NAME_CASE(ATOMIC_CMP_SWAP)
|
2016-04-12 22:05:04 +08:00
|
|
|
NODE_NAME_CASE(ATOMIC_INC)
|
|
|
|
NODE_NAME_CASE(ATOMIC_DEC)
|
2018-01-17 22:05:05 +08:00
|
|
|
NODE_NAME_CASE(ATOMIC_LOAD_FADD)
|
|
|
|
NODE_NAME_CASE(ATOMIC_LOAD_FMIN)
|
|
|
|
NODE_NAME_CASE(ATOMIC_LOAD_FMAX)
|
2016-12-21 01:19:44 +08:00
|
|
|
NODE_NAME_CASE(BUFFER_LOAD)
|
|
|
|
NODE_NAME_CASE(BUFFER_LOAD_FORMAT)
|
2018-01-13 05:12:19 +08:00
|
|
|
NODE_NAME_CASE(BUFFER_LOAD_FORMAT_D16)
|
2017-11-09 09:52:48 +08:00
|
|
|
NODE_NAME_CASE(BUFFER_STORE)
|
|
|
|
NODE_NAME_CASE(BUFFER_STORE_FORMAT)
|
2018-01-13 05:12:19 +08:00
|
|
|
NODE_NAME_CASE(BUFFER_STORE_FORMAT_D16)
|
2017-11-09 09:52:48 +08:00
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_SWAP)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_ADD)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_SUB)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_SMIN)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_UMIN)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_SMAX)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_UMAX)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_AND)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_OR)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_XOR)
|
|
|
|
NODE_NAME_CASE(BUFFER_ATOMIC_CMPSWAP)
|
2018-01-19 06:08:53 +08:00
|
|
|
|
2015-05-08 05:33:59 +08:00
|
|
|
case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
2015-05-08 05:33:59 +08:00
|
|
|
return nullptr;
|
2012-12-12 05:25:42 +08:00
|
|
|
}
|
2014-03-26 02:18:27 +08:00
|
|
|
|
2016-11-11 07:31:06 +08:00
|
|
|
SDValue AMDGPUTargetLowering::getSqrtEstimate(SDValue Operand,
|
|
|
|
SelectionDAG &DAG, int Enabled,
|
|
|
|
int &RefinementSteps,
|
|
|
|
bool &UseOneConstNR,
|
|
|
|
bool Reciprocal) const {
|
2015-01-14 04:53:18 +08:00
|
|
|
EVT VT = Operand.getValueType();
|
|
|
|
|
|
|
|
if (VT == MVT::f32) {
|
|
|
|
RefinementSteps = 0;
|
|
|
|
return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
|
|
|
|
}
|
|
|
|
|
|
|
|
// TODO: There is also f64 rsq instruction, but the documentation is less
|
|
|
|
// clear on its precision.
|
|
|
|
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
2015-01-14 04:53:23 +08:00
|
|
|
SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
|
2016-10-21 00:55:45 +08:00
|
|
|
SelectionDAG &DAG, int Enabled,
|
|
|
|
int &RefinementSteps) const {
|
2015-01-14 04:53:23 +08:00
|
|
|
EVT VT = Operand.getValueType();
|
|
|
|
|
|
|
|
if (VT == MVT::f32) {
|
|
|
|
// Reciprocal, < 1 ulp error.
|
|
|
|
//
|
|
|
|
// This reciprocal approximation converges to < 0.5 ulp error with one
|
|
|
|
// newton rhapson performed with two fused multiple adds (FMAs).
|
|
|
|
|
|
|
|
RefinementSteps = 0;
|
|
|
|
return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
|
|
|
|
}
|
|
|
|
|
|
|
|
// TODO: There is also f64 rcp instruction, but the documentation is less
|
|
|
|
// clear on its precision.
|
|
|
|
|
|
|
|
return SDValue();
|
|
|
|
}
|
|
|
|
|
2014-05-15 05:14:37 +08:00
|
|
|
void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
|
2017-04-28 13:31:46 +08:00
|
|
|
const SDValue Op, KnownBits &Known,
|
2017-03-31 19:24:16 +08:00
|
|
|
const APInt &DemandedElts, const SelectionDAG &DAG, unsigned Depth) const {
|
2014-04-01 03:35:33 +08:00
|
|
|
|
2017-05-06 01:36:09 +08:00
|
|
|
Known.resetAll(); // Don't know anything.
|
2014-05-23 02:09:00 +08:00
|
|
|
|
2014-04-01 03:35:33 +08:00
|
|
|
unsigned Opc = Op.getOpcode();
|
2014-05-23 02:09:00 +08:00
|
|
|
|
2014-04-01 03:35:33 +08:00
|
|
|
switch (Opc) {
|
2014-05-23 02:09:00 +08:00
|
|
|
default:
|
|
|
|
break;
|
2015-05-01 01:15:56 +08:00
|
|
|
case AMDGPUISD::CARRY:
|
|
|
|
case AMDGPUISD::BORROW: {
|
2017-04-28 13:31:46 +08:00
|
|
|
Known.Zero = APInt::getHighBitsSet(32, 31);
|
2015-05-01 01:15:56 +08:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2014-05-23 02:09:00 +08:00
|
|
|
case AMDGPUISD::BFE_I32:
|
|
|
|
case AMDGPUISD::BFE_U32: {
|
|
|
|
ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
|
|
|
|
if (!CWidth)
|
|
|
|
return;
|
|
|
|
|
|
|
|
uint32_t Width = CWidth->getZExtValue() & 0x1f;
|
|
|
|
|
2014-10-17 04:07:40 +08:00
|
|
|
if (Opc == AMDGPUISD::BFE_U32)
|
2017-04-28 13:31:46 +08:00
|
|
|
Known.Zero = APInt::getHighBitsSet(32, 32 - Width);
|
2014-05-23 02:09:00 +08:00
|
|
|
|
2014-04-01 03:35:33 +08:00
|
|
|
break;
|
|
|
|
}
|
2017-04-01 03:53:03 +08:00
|
|
|
case AMDGPUISD::FP_TO_FP16:
|
|
|
|
case AMDGPUISD::FP16_ZEXT: {
|
2017-04-28 13:31:46 +08:00
|
|
|
unsigned BitWidth = Known.getBitWidth();
|
2017-04-01 03:53:03 +08:00
|
|
|
|
2017-03-16 03:04:26 +08:00
|
|
|
// High bits are zero.
|
2017-04-28 13:31:46 +08:00
|
|
|
Known.Zero = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
|
2017-03-16 03:04:26 +08:00
|
|
|
break;
|
|
|
|
}
|
2017-08-29 00:35:37 +08:00
|
|
|
case AMDGPUISD::MUL_U24:
|
|
|
|
case AMDGPUISD::MUL_I24: {
|
|
|
|
KnownBits LHSKnown, RHSKnown;
|
2017-09-02 04:43:20 +08:00
|
|
|
DAG.computeKnownBits(Op.getOperand(0), LHSKnown, Depth + 1);
|
|
|
|
DAG.computeKnownBits(Op.getOperand(1), RHSKnown, Depth + 1);
|
2017-08-29 00:35:37 +08:00
|
|
|
|
|
|
|
unsigned TrailZ = LHSKnown.countMinTrailingZeros() +
|
|
|
|
RHSKnown.countMinTrailingZeros();
|
|
|
|
Known.Zero.setLowBits(std::min(TrailZ, 32u));
|
|
|
|
|
|
|
|
unsigned LHSValBits = 32 - std::max(LHSKnown.countMinSignBits(), 8u);
|
|
|
|
unsigned RHSValBits = 32 - std::max(RHSKnown.countMinSignBits(), 8u);
|
|
|
|
unsigned MaxValBits = std::min(LHSValBits + RHSValBits, 32u);
|
|
|
|
if (MaxValBits >= 32)
|
|
|
|
break;
|
|
|
|
bool Negative = false;
|
|
|
|
if (Opc == AMDGPUISD::MUL_I24) {
|
|
|
|
bool LHSNegative = !!(LHSKnown.One & (1 << 23));
|
|
|
|
bool LHSPositive = !!(LHSKnown.Zero & (1 << 23));
|
|
|
|
bool RHSNegative = !!(RHSKnown.One & (1 << 23));
|
|
|
|
bool RHSPositive = !!(RHSKnown.Zero & (1 << 23));
|
|
|
|
if ((!LHSNegative && !LHSPositive) || (!RHSNegative && !RHSPositive))
|
|
|
|
break;
|
|
|
|
Negative = (LHSNegative && RHSPositive) || (LHSPositive && RHSNegative);
|
|
|
|
}
|
|
|
|
if (Negative)
|
|
|
|
Known.One.setHighBits(32 - MaxValBits);
|
|
|
|
else
|
|
|
|
Known.Zero.setHighBits(32 - MaxValBits);
|
|
|
|
break;
|
|
|
|
}
|
2018-06-13 07:50:37 +08:00
|
|
|
case AMDGPUISD::PERM: {
|
|
|
|
ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Op.getOperand(2));
|
|
|
|
if (!CMask)
|
|
|
|
return;
|
|
|
|
|
|
|
|
KnownBits LHSKnown, RHSKnown;
|
|
|
|
DAG.computeKnownBits(Op.getOperand(0), LHSKnown, Depth + 1);
|
|
|
|
DAG.computeKnownBits(Op.getOperand(1), RHSKnown, Depth + 1);
|
|
|
|
unsigned Sel = CMask->getZExtValue();
|
|
|
|
|
|
|
|
for (unsigned I = 0; I < 32; I += 8) {
|
|
|
|
unsigned SelBits = Sel & 0xff;
|
|
|
|
if (SelBits < 4) {
|
2018-06-14 02:52:54 +08:00
|
|
|
SelBits *= 8;
|
|
|
|
Known.One |= ((RHSKnown.One.getZExtValue() >> SelBits) & 0xff) << I;
|
|
|
|
Known.Zero |= ((RHSKnown.Zero.getZExtValue() >> SelBits) & 0xff) << I;
|
2018-06-13 07:50:37 +08:00
|
|
|
} else if (SelBits < 7) {
|
2018-06-14 02:52:54 +08:00
|
|
|
SelBits = (SelBits & 3) * 8;
|
|
|
|
Known.One |= ((LHSKnown.One.getZExtValue() >> SelBits) & 0xff) << I;
|
|
|
|
Known.Zero |= ((LHSKnown.Zero.getZExtValue() >> SelBits) & 0xff) << I;
|
2018-06-13 07:50:37 +08:00
|
|
|
} else if (SelBits == 0x0c) {
|
2018-06-14 02:52:54 +08:00
|
|
|
Known.Zero |= 0xff << I;
|
2018-06-13 07:50:37 +08:00
|
|
|
} else if (SelBits > 0x0c) {
|
2018-06-14 02:52:54 +08:00
|
|
|
Known.One |= 0xff << I;
|
2018-06-13 07:50:37 +08:00
|
|
|
}
|
|
|
|
Sel >>= 8;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
2017-11-14 06:55:05 +08:00
|
|
|
case ISD::INTRINSIC_WO_CHAIN: {
|
|
|
|
unsigned IID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
|
|
|
|
switch (IID) {
|
|
|
|
case Intrinsic::amdgcn_mbcnt_lo:
|
|
|
|
case Intrinsic::amdgcn_mbcnt_hi: {
|
|
|
|
// These return at most the wavefront size - 1.
|
|
|
|
unsigned Size = Op.getValueType().getSizeInBits();
|
|
|
|
Known.Zero.setHighBits(Size - Subtarget->getWavefrontSizeLog2());
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
2014-05-23 02:09:00 +08:00
|
|
|
}
|
2014-03-26 02:18:27 +08:00
|
|
|
}
|
2014-05-23 02:09:03 +08:00
|
|
|
|
|
|
|
unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
|
2017-03-31 21:54:09 +08:00
|
|
|
SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
|
|
|
|
unsigned Depth) const {
|
2014-05-23 02:09:03 +08:00
|
|
|
switch (Op.getOpcode()) {
|
|
|
|
case AMDGPUISD::BFE_I32: {
|
|
|
|
ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
|
|
|
|
if (!Width)
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
unsigned SignBits = 32 - Width->getZExtValue() + 1;
|
2015-11-26 03:41:11 +08:00
|
|
|
if (!isNullConstant(Op.getOperand(1)))
|
2014-05-23 02:09:03 +08:00
|
|
|
return SignBits;
|
|
|
|
|
|
|
|
// TODO: Could probably figure something out with non-0 offsets.
|
|
|
|
unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
|
|
|
|
return std::max(SignBits, Op0SignBits);
|
|
|
|
}
|
|
|
|
|
2014-05-23 02:09:07 +08:00
|
|
|
case AMDGPUISD::BFE_U32: {
|
|
|
|
ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
|
|
|
|
return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;
|
|
|
|
}
|
|
|
|
|
2015-05-01 01:15:56 +08:00
|
|
|
case AMDGPUISD::CARRY:
|
|
|
|
case AMDGPUISD::BORROW:
|
|
|
|
return 31;
|
2017-04-01 03:53:03 +08:00
|
|
|
case AMDGPUISD::FP_TO_FP16:
|
|
|
|
case AMDGPUISD::FP16_ZEXT:
|
|
|
|
return 16;
|
2014-05-23 02:09:03 +08:00
|
|
|
default:
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
}
|