1 //===-- AMDGPUSubtarget.cpp - AMDGPU Subtarget Information ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief Implements the AMDGPU specific subclass of TargetSubtarget.
13 //===----------------------------------------------------------------------===//
15 #include "AMDGPUSubtarget.h"
16 #include "R600ISelLowering.h"
17 #include "R600InstrInfo.h"
18 #include "SIFrameLowering.h"
19 #include "SIISelLowering.h"
20 #include "SIInstrInfo.h"
21 #include "SIMachineFunctionInfo.h"
22 #include "llvm/ADT/SmallString.h"
23 #include "llvm/CodeGen/MachineScheduler.h"
27 #define DEBUG_TYPE "amdgpu-subtarget"
29 #define GET_SUBTARGETINFO_ENUM
30 #define GET_SUBTARGETINFO_TARGET_DESC
31 #define GET_SUBTARGETINFO_CTOR
32 #include "AMDGPUGenSubtargetInfo.inc"
34 AMDGPUSubtarget::~AMDGPUSubtarget() {}
37 AMDGPUSubtarget::initializeSubtargetDependencies(const Triple &TT,
38 StringRef GPU, StringRef FS) {
39 // Determine default and user-specified characteristics
40 // On SI+, we want FP64 denormals to be on by default. FP32 denormals can be
41 // enabled, but some instructions do not respect them and they run at the
42 // double precision rate, so don't enable by default.
44 // We want to be able to turn these off, but making this a subtarget feature
45 // for SI has the unhelpful behavior that it unsets everything else if you
48 SmallString<256> FullFS("+promote-alloca,+fp64-denormals,+load-store-opt,");
49 if (isAmdHsaOS()) // Turn on FlatForGlobal for HSA.
50 FullFS += "+flat-for-global,+unaligned-buffer-access,";
53 ParseSubtargetFeatures(GPU, FullFS);
55 // FIXME: I don't think think Evergreen has any useful support for
56 // denormals, but should be checked. Should we issue a warning somewhere
57 // if someone tries to enable these?
58 if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
59 FP32Denormals = false;
60 FP64Denormals = false;
63 // Set defaults if needed.
64 if (MaxPrivateElementSize == 0)
65 MaxPrivateElementSize = 4;
70 AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
71 const TargetMachine &TM)
72 : AMDGPUGenSubtargetInfo(TT, GPU, FS),
74 Gen(TT.getArch() == Triple::amdgcn ? SOUTHERN_ISLANDS : R600),
75 IsaVersion(ISAVersion0_0_0),
79 MaxPrivateElementSize(0),
88 UnalignedBufferAccess(false),
91 DebuggerInsertNops(false),
92 DebuggerReserveRegs(false),
93 DebuggerEmitPrologue(false),
95 EnableVGPRSpilling(false),
96 EnablePromoteAlloca(false),
97 EnableLoadStoreOpt(false),
98 EnableUnsafeDSOffsetFolding(false),
99 EnableSIScheduler(false),
108 HasSMemRealTime(false),
109 Has16BitInsts(false),
110 FlatAddressSpace(false),
115 HasVertexCache(false),
118 FeatureDisable(false),
119 InstrItins(getInstrItineraryForCPU(GPU)) {
120 initializeSubtargetDependencies(TT, GPU, FS);
123 // FIXME: These limits are for SI. Did they change with the larger maximum LDS
125 unsigned AMDGPUSubtarget::getMaxLocalMemSizeWithWaveCount(unsigned NWaves) const {
146 return getLocalMemorySize();
150 unsigned AMDGPUSubtarget::getOccupancyWithLocalMemSize(uint32_t Bytes) const {
181 R600Subtarget::R600Subtarget(const Triple &TT, StringRef GPU, StringRef FS,
182 const TargetMachine &TM) :
183 AMDGPUSubtarget(TT, GPU, FS, TM),
185 FrameLowering(TargetFrameLowering::StackGrowsUp, getStackAlignment(), 0),
188 SISubtarget::SISubtarget(const Triple &TT, StringRef GPU, StringRef FS,
189 const TargetMachine &TM) :
190 AMDGPUSubtarget(TT, GPU, FS, TM),
192 FrameLowering(TargetFrameLowering::StackGrowsUp, getStackAlignment(), 0),
196 unsigned R600Subtarget::getStackEntrySize() const {
197 switch (getWavefrontSize()) {
201 return hasCaymanISA() ? 4 : 8;
205 llvm_unreachable("Illegal wavefront size.");
209 void SISubtarget::overrideSchedPolicy(MachineSchedPolicy &Policy,
210 unsigned NumRegionInstrs) const {
211 // Track register pressure so the scheduler can try to decrease
212 // pressure once register usage is above the threshold defined by
213 // SIRegisterInfo::getRegPressureSetLimit()
214 Policy.ShouldTrackPressure = true;
216 // Enabling both top down and bottom up scheduling seems to give us less
217 // register spills than just using one of these approaches on its own.
218 Policy.OnlyTopDown = false;
219 Policy.OnlyBottomUp = false;
221 // Enabling ShouldTrackLaneMasks crashes the SI Machine Scheduler.
222 if (!enableSIScheduler())
223 Policy.ShouldTrackLaneMasks = true;
226 bool SISubtarget::isVGPRSpillingEnabled(const Function& F) const {
227 return EnableVGPRSpilling || !AMDGPU::isShader(F.getCallingConv());
230 unsigned SISubtarget::getAmdKernelCodeChipID() const {
231 switch (getGeneration()) {
235 llvm_unreachable("ChipID unknown");
239 AMDGPU::IsaVersion SISubtarget::getIsaVersion() const {
240 return AMDGPU::getIsaVersion(getFeatureBits());