2012-12-11 21:25:42 +00:00
|
|
|
//===-- AMDGPUSubtarget.cpp - AMDGPU Subtarget Information ----------------===//
|
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
/// \file
|
|
|
|
/// \brief Implements the AMDGPU specific subclass of TargetSubtarget.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#include "AMDGPUSubtarget.h"
|
2014-07-25 22:22:39 +00:00
|
|
|
#include "R600ISelLowering.h"
|
2014-06-13 01:32:00 +00:00
|
|
|
#include "R600InstrInfo.h"
|
2014-07-25 22:22:39 +00:00
|
|
|
#include "R600MachineScheduler.h"
|
|
|
|
#include "SIISelLowering.h"
|
2015-01-14 11:23:27 +00:00
|
|
|
#include "SIInstrInfo.h"
|
2015-01-20 19:33:04 +00:00
|
|
|
#include "SIMachineFunctionInfo.h"
|
2014-07-13 02:08:26 +00:00
|
|
|
#include "llvm/ADT/SmallString.h"
|
2015-01-29 16:55:25 +00:00
|
|
|
#include "llvm/CodeGen/MachineScheduler.h"
|
2014-07-13 02:08:26 +00:00
|
|
|
|
2012-12-11 21:25:42 +00:00
|
|
|
using namespace llvm;
|
|
|
|
|
2014-04-21 22:55:11 +00:00
|
|
|
#define DEBUG_TYPE "amdgpu-subtarget"
|
|
|
|
|
2012-12-11 21:25:42 +00:00
|
|
|
#define GET_SUBTARGETINFO_ENUM
|
|
|
|
#define GET_SUBTARGETINFO_TARGET_DESC
|
|
|
|
#define GET_SUBTARGETINFO_CTOR
|
|
|
|
#include "AMDGPUGenSubtargetInfo.inc"
|
|
|
|
|
2014-07-25 22:22:39 +00:00
|
|
|
AMDGPUSubtarget &
|
2015-02-19 00:15:33 +00:00
|
|
|
AMDGPUSubtarget::initializeSubtargetDependencies(StringRef TT, StringRef GPU,
|
|
|
|
StringRef FS) {
|
2014-07-25 22:22:39 +00:00
|
|
|
// Determine default and user-specified characteristics
|
2014-07-14 23:40:49 +00:00
|
|
|
// On SI+, we want FP64 denormals to be on by default. FP32 denormals can be
|
|
|
|
// enabled, but some instructions do not respect them and they run at the
|
|
|
|
// double precision rate, so don't enable by default.
|
|
|
|
//
|
|
|
|
// We want to be able to turn these off, but making this a subtarget feature
|
|
|
|
// for SI has the unhelpful behavior that it unsets everything else if you
|
|
|
|
// disable it.
|
2014-07-13 02:08:26 +00:00
|
|
|
|
2014-07-14 23:40:49 +00:00
|
|
|
SmallString<256> FullFS("+promote-alloca,+fp64-denormals,");
|
2014-07-13 02:08:26 +00:00
|
|
|
FullFS += FS;
|
|
|
|
|
2015-01-28 15:38:42 +00:00
|
|
|
if (GPU == "" && Triple(TT).getArch() == Triple::amdgcn)
|
|
|
|
GPU = "SI";
|
|
|
|
|
2014-07-13 02:08:26 +00:00
|
|
|
ParseSubtargetFeatures(GPU, FullFS);
|
2014-06-13 01:32:00 +00:00
|
|
|
|
2014-07-25 22:22:39 +00:00
|
|
|
// FIXME: I don't think think Evergreen has any useful support for
|
|
|
|
// denormals, but should be checked. Should we issue a warning somewhere
|
|
|
|
// if someone tries to enable these?
|
2014-06-13 01:32:00 +00:00
|
|
|
if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
|
2014-07-14 23:40:49 +00:00
|
|
|
FP32Denormals = false;
|
|
|
|
FP64Denormals = false;
|
2014-07-25 22:22:39 +00:00
|
|
|
}
|
|
|
|
return *this;
|
|
|
|
}
|
|
|
|
|
|
|
|
AMDGPUSubtarget::AMDGPUSubtarget(StringRef TT, StringRef GPU, StringRef FS,
|
|
|
|
TargetMachine &TM)
|
|
|
|
: AMDGPUGenSubtargetInfo(TT, GPU, FS), DevName(GPU), Is64bit(false),
|
|
|
|
DumpCode(false), R600ALUInst(false), HasVertexCache(false),
|
|
|
|
TexVTXClauseSize(0), Gen(AMDGPUSubtarget::R600), FP64(false),
|
2015-02-19 00:15:33 +00:00
|
|
|
FP64Denormals(false), FP32Denormals(false), FastFMAF32(false),
|
|
|
|
CaymanISA(false), FlatAddressSpace(false), EnableIRStructurizer(true),
|
|
|
|
EnablePromoteAlloca(false), EnableIfCvt(true), EnableLoadStoreOpt(false),
|
|
|
|
WavefrontSize(0), CFALUBug(false), LocalMemorySize(0),
|
2015-03-09 15:48:09 +00:00
|
|
|
EnableVGPRSpilling(false), SGPRInitBug(false),
|
2015-04-23 19:33:54 +00:00
|
|
|
IsGCN(false), GCN1Encoding(false), GCN3Encoding(false), CIInsts(false),
|
2014-07-25 22:22:39 +00:00
|
|
|
FrameLowering(TargetFrameLowering::StackGrowsUp,
|
|
|
|
64 * 16, // Maximum stack alignment (long16)
|
|
|
|
0),
|
2015-02-19 00:15:33 +00:00
|
|
|
InstrItins(getInstrItineraryForCPU(GPU)), TargetTriple(TT) {
|
2015-01-28 16:04:26 +00:00
|
|
|
|
|
|
|
initializeSubtargetDependencies(TT, GPU, FS);
|
|
|
|
|
2014-07-25 22:22:39 +00:00
|
|
|
if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
|
|
|
|
InstrInfo.reset(new R600InstrInfo(*this));
|
2015-01-30 23:24:40 +00:00
|
|
|
TLInfo.reset(new R600TargetLowering(TM, *this));
|
2014-06-13 01:32:00 +00:00
|
|
|
} else {
|
|
|
|
InstrInfo.reset(new SIInstrInfo(*this));
|
2015-01-30 23:24:40 +00:00
|
|
|
TLInfo.reset(new SITargetLowering(TM, *this));
|
2014-06-13 01:32:00 +00:00
|
|
|
}
|
2012-12-11 21:25:42 +00:00
|
|
|
}
|
|
|
|
|
2014-06-27 17:57:00 +00:00
|
|
|
unsigned AMDGPUSubtarget::getStackEntrySize() const {
|
2014-01-22 21:55:43 +00:00
|
|
|
assert(getGeneration() <= NORTHERN_ISLANDS);
|
|
|
|
switch(getWavefrontSize()) {
|
|
|
|
case 16:
|
|
|
|
return 8;
|
|
|
|
case 32:
|
2014-06-27 17:57:00 +00:00
|
|
|
return hasCaymanISA() ? 4 : 8;
|
2014-01-22 21:55:43 +00:00
|
|
|
case 64:
|
|
|
|
return 4;
|
|
|
|
default:
|
|
|
|
llvm_unreachable("Illegal wavefront size.");
|
|
|
|
}
|
|
|
|
}
|
2014-12-02 22:00:07 +00:00
|
|
|
|
|
|
|
unsigned AMDGPUSubtarget::getAmdKernelCodeChipID() const {
|
|
|
|
switch(getGeneration()) {
|
|
|
|
default: llvm_unreachable("ChipID unknown");
|
|
|
|
case SEA_ISLANDS: return 12;
|
|
|
|
}
|
|
|
|
}
|
2015-01-20 19:33:04 +00:00
|
|
|
|
|
|
|
bool AMDGPUSubtarget::isVGPRSpillingEnabled(
|
|
|
|
const SIMachineFunctionInfo *MFI) const {
|
|
|
|
return MFI->getShaderType() == ShaderType::COMPUTE || EnableVGPRSpilling;
|
|
|
|
}
|
2015-01-29 16:55:25 +00:00
|
|
|
|
|
|
|
void AMDGPUSubtarget::overrideSchedPolicy(MachineSchedPolicy &Policy,
|
|
|
|
MachineInstr *begin,
|
|
|
|
MachineInstr *end,
|
|
|
|
unsigned NumRegionInstrs) const {
|
|
|
|
if (getGeneration() >= SOUTHERN_ISLANDS) {
|
|
|
|
|
|
|
|
// Track register pressure so the scheduler can try to decrease
|
|
|
|
// pressure once register usage is above the threshold defined by
|
|
|
|
// SIRegisterInfo::getRegPressureSetLimit()
|
|
|
|
Policy.ShouldTrackPressure = true;
|
|
|
|
|
|
|
|
// Enabling both top down and bottom up scheduling seems to give us less
|
|
|
|
// register spills than just using one of these approaches on its own.
|
|
|
|
Policy.OnlyTopDown = false;
|
|
|
|
Policy.OnlyBottomUp = false;
|
|
|
|
}
|
|
|
|
}
|