Home | History | Annotate | Download | only in AMDGPU
      1 //===-- AMDGPUSubtarget.cpp - AMDGPU Subtarget Information ----------------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 /// \file
     11 /// \brief Implements the AMDGPU specific subclass of TargetSubtarget.
     12 //
     13 //===----------------------------------------------------------------------===//
     14 
     15 #include "AMDGPUSubtarget.h"
     16 #include "R600ISelLowering.h"
     17 #include "R600InstrInfo.h"
     18 #include "R600MachineScheduler.h"
     19 #include "SIFrameLowering.h"
     20 #include "SIISelLowering.h"
     21 #include "SIInstrInfo.h"
     22 #include "SIMachineFunctionInfo.h"
     23 #include "llvm/ADT/SmallString.h"
     24 #include "llvm/CodeGen/MachineScheduler.h"
     25 
     26 using namespace llvm;
     27 
     28 #define DEBUG_TYPE "amdgpu-subtarget"
     29 
     30 #define GET_SUBTARGETINFO_ENUM
     31 #define GET_SUBTARGETINFO_TARGET_DESC
     32 #define GET_SUBTARGETINFO_CTOR
     33 #include "AMDGPUGenSubtargetInfo.inc"
     34 
     35 AMDGPUSubtarget &
     36 AMDGPUSubtarget::initializeSubtargetDependencies(const Triple &TT,
     37                                                  StringRef GPU, StringRef FS) {
     38   // Determine default and user-specified characteristics
     39   // On SI+, we want FP64 denormals to be on by default. FP32 denormals can be
     40   // enabled, but some instructions do not respect them and they run at the
     41   // double precision rate, so don't enable by default.
     42   //
     43   // We want to be able to turn these off, but making this a subtarget feature
     44   // for SI has the unhelpful behavior that it unsets everything else if you
     45   // disable it.
     46 
     47   SmallString<256> FullFS("+promote-alloca,+fp64-denormals,");
     48   FullFS += FS;
     49 
     50   if (GPU == "" && TT.getArch() == Triple::amdgcn)
     51     GPU = "SI";
     52 
     53   ParseSubtargetFeatures(GPU, FullFS);
     54 
     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;
     61   }
     62   return *this;
     63 }
     64 
     65 AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     66                                  TargetMachine &TM)
     67     : AMDGPUGenSubtargetInfo(TT, GPU, FS), DevName(GPU), Is64bit(false),
     68       DumpCode(false), R600ALUInst(false), HasVertexCache(false),
     69       TexVTXClauseSize(0), Gen(AMDGPUSubtarget::R600), FP64(false),
     70       FP64Denormals(false), FP32Denormals(false), FastFMAF32(false),
     71       CaymanISA(false), FlatAddressSpace(false), EnableIRStructurizer(true),
     72       EnablePromoteAlloca(false), EnableIfCvt(true), EnableLoadStoreOpt(false),
     73       EnableUnsafeDSOffsetFolding(false),
     74       WavefrontSize(0), CFALUBug(false), LocalMemorySize(0),
     75       EnableVGPRSpilling(false), SGPRInitBug(false), IsGCN(false),
     76       GCN1Encoding(false), GCN3Encoding(false), CIInsts(false), LDSBankCount(0),
     77       IsaVersion(ISAVersion0_0_0), EnableHugeScratchBuffer(false),
     78       FrameLowering(nullptr),
     79       InstrItins(getInstrItineraryForCPU(GPU)), TargetTriple(TT) {
     80 
     81   initializeSubtargetDependencies(TT, GPU, FS);
     82 
     83   const unsigned MaxStackAlign = 64 * 16; // Maximum stack alignment (long16)
     84 
     85   if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
     86     InstrInfo.reset(new R600InstrInfo(*this));
     87     TLInfo.reset(new R600TargetLowering(TM, *this));
     88 
     89     // FIXME: Should have R600 specific FrameLowering
     90     FrameLowering.reset(new AMDGPUFrameLowering(
     91                           TargetFrameLowering::StackGrowsUp,
     92                           MaxStackAlign,
     93                           0));
     94   } else {
     95     InstrInfo.reset(new SIInstrInfo(*this));
     96     TLInfo.reset(new SITargetLowering(TM, *this));
     97     FrameLowering.reset(new SIFrameLowering(
     98                           TargetFrameLowering::StackGrowsUp,
     99                           MaxStackAlign,
    100                           0));
    101   }
    102 }
    103 
    104 unsigned AMDGPUSubtarget::getStackEntrySize() const {
    105   assert(getGeneration() <= NORTHERN_ISLANDS);
    106   switch(getWavefrontSize()) {
    107   case 16:
    108     return 8;
    109   case 32:
    110     return hasCaymanISA() ? 4 : 8;
    111   case 64:
    112     return 4;
    113   default:
    114     llvm_unreachable("Illegal wavefront size.");
    115   }
    116 }
    117 
    118 unsigned AMDGPUSubtarget::getAmdKernelCodeChipID() const {
    119   switch(getGeneration()) {
    120   default: llvm_unreachable("ChipID unknown");
    121   case SEA_ISLANDS: return 12;
    122   }
    123 }
    124 
    125 AMDGPU::IsaVersion AMDGPUSubtarget::getIsaVersion() const {
    126   return AMDGPU::getIsaVersion(getFeatureBits());
    127 }
    128 
    129 bool AMDGPUSubtarget::isVGPRSpillingEnabled(
    130                                        const SIMachineFunctionInfo *MFI) const {
    131   return MFI->getShaderType() == ShaderType::COMPUTE || EnableVGPRSpilling;
    132 }
    133 
    134 void AMDGPUSubtarget::overrideSchedPolicy(MachineSchedPolicy &Policy,
    135                                           MachineInstr *begin,
    136                                           MachineInstr *end,
    137                                           unsigned NumRegionInstrs) const {
    138   if (getGeneration() >= SOUTHERN_ISLANDS) {
    139 
    140     // Track register pressure so the scheduler can try to decrease
    141     // pressure once register usage is above the threshold defined by
    142     // SIRegisterInfo::getRegPressureSetLimit()
    143     Policy.ShouldTrackPressure = true;
    144 
    145     // Enabling both top down and bottom up scheduling seems to give us less
    146     // register spills than just using one of these approaches on its own.
    147     Policy.OnlyTopDown = false;
    148     Policy.OnlyBottomUp = false;
    149   }
    150 }
    151 
    152