llvm.org GIT mirror llvm / 28fe7a0 lib / Target / AMDGPU / AMDGPUTargetTransformInfo.cpp
28fe7a0

Tree @28fe7a0 (Download .tar.gz)

AMDGPUTargetTransformInfo.cpp @28fe7a0

57e6b2d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1937233
01df2fa
57e6b2d
01df2fa
a6a87b5
529919f
57e6b2d
 
974a445
57e6b2d
 
42e8630
 
26bc071
a6a87b5
ee17bf3
c719852
ee17bf3
 
 
 
21a9a66
529919f
21a9a66
 
df4663b
01df2fa
21a9a66
01df2fa
529919f
 
01df2fa
 
 
 
 
431e9f0
9121540
 
 
 
 
ee17bf3
01df2fa
 
 
 
a32c319
a6a87b5
fd22980
 
 
 
 
 
 
 
 
 
1e79dfc
 
 
fd22980
cac51be
fd22980
 
 
2320de6
b11dd50
 
 
 
 
 
 
 
 
 
 
2320de6
 
 
 
 
 
 
 
 
 
6e2ee6c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28fe7a0
6e2ee6c
 
 
 
 
 
 
 
 
 
 
 
688dd45
 
 
6e2ee6c
 
688dd45
 
6e2ee6c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
//===-- AMDGPUTargetTransformInfo.cpp - AMDGPU specific TTI pass ---------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// \file
// This file implements a TargetTransformInfo analysis pass specific to the
// AMDGPU target machine. It uses the target's detailed information to provide
// more precise answers to certain TTI queries, while letting the target
// independent and default TTI implementations handle the rest.
//
//===----------------------------------------------------------------------===//

#include "AMDGPUTargetTransformInfo.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/CodeGen/BasicTTIImpl.h"
#include "llvm/IR/Module.h"
#include "llvm/Support/Debug.h"
#include "llvm/Target/CostTable.h"
#include "llvm/Target/TargetLowering.h"
using namespace llvm;

#define DEBUG_TYPE "AMDGPUtti"

void AMDGPUTTIImpl::getUnrollingPreferences(Loop *L,
                                            TTI::UnrollingPreferences &UP) {
  UP.Threshold = 300; // Twice the default.
  UP.MaxCount = UINT_MAX;
  UP.Partial = true;

  // TODO: Do we want runtime unrolling?

  for (const BasicBlock *BB : L->getBlocks()) {
    const DataLayout &DL = BB->getModule()->getDataLayout();
    for (const Instruction &I : *BB) {
      const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I);
      if (!GEP || GEP->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
        continue;

      const Value *Ptr = GEP->getPointerOperand();
      const AllocaInst *Alloca =
          dyn_cast<AllocaInst>(GetUnderlyingObject(Ptr, DL));
      if (Alloca) {
        // We want to do whatever we can to limit the number of alloca
        // instructions that make it through to the code generator.  allocas
        // require us to use indirect addressing, which is slow and prone to
        // compiler bugs.  If this loop does an address calculation on an
        // alloca ptr, then we want to use a higher than normal loop unroll
        // threshold. This will give SROA a better chance to eliminate these
        // allocas.
        //
        // Don't use the maximum allowed value here as it will make some
        // programs way too big.
        UP.Threshold = 800;
      }
    }
  }
}

unsigned AMDGPUTTIImpl::getNumberOfRegisters(bool Vec) {
  if (Vec)
    return 0;

  // Number of VGPRs on SI.
  if (ST->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)
    return 256;

  return 4 * 128; // XXX - 4 channels. Should these count as vector instead?
}

unsigned AMDGPUTTIImpl::getRegisterBitWidth(bool Vector) {
  return Vector ? 0 : 32;
}

unsigned AMDGPUTTIImpl::getMaxInterleaveFactor(unsigned VF) {
  // Semi-arbitrary large amount.
  return 64;
}

unsigned AMDGPUTTIImpl::getCFInstrCost(unsigned Opcode) {
  // XXX - For some reason this isn't called for switch.
  switch (Opcode) {
  case Instruction::Br:
  case Instruction::Ret:
    return 10;
  default:
    return BaseT::getCFInstrCost(Opcode);
  }
}

int AMDGPUTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,
                                      unsigned Index) {
  switch (Opcode) {
  case Instruction::ExtractElement:
    // Dynamic indexing isn't free and is best avoided.
    return Index == ~0u ? 2 : 0;
  default:
    return BaseT::getVectorInstrCost(Opcode, ValTy, Index);
  }
}

static bool isIntrinsicSourceOfDivergence(const TargetIntrinsicInfo *TII,
                                          const IntrinsicInst *I) {
  switch (I->getIntrinsicID()) {
  default:
    return false;
  case Intrinsic::not_intrinsic:
    // This means we have an intrinsic that isn't defined in
    // IntrinsicsAMDGPU.td
    break;

  case Intrinsic::amdgcn_interp_p1:
  case Intrinsic::amdgcn_interp_p2:
  case Intrinsic::amdgcn_mbcnt_hi:
  case Intrinsic::amdgcn_mbcnt_lo:
  case Intrinsic::r600_read_tidig_x:
  case Intrinsic::r600_read_tidig_y:
  case Intrinsic::r600_read_tidig_z:
    return true;
  }

  StringRef Name = I->getCalledFunction()->getName();
  switch (TII->lookupName((const char *)Name.bytes_begin(), Name.size())) {
  default:
    return false;
  case AMDGPUIntrinsic::SI_tid:
  case AMDGPUIntrinsic::SI_fs_interp:
  case AMDGPUIntrinsic::SI_fs_constant:
    return true;
  }
}

static bool isArgPassedInSGPR(const Argument *A) {
  const Function *F = A->getParent();
  unsigned ShaderType = AMDGPU::getShaderType(*F);

  // Arguments to compute shaders are never a source of divergence.
  if (ShaderType == ShaderType::COMPUTE)
    return true;

  // For non-compute shaders, SGPR inputs are marked with either inreg or byval.
  if (F->getAttributes().hasAttribute(A->getArgNo() + 1, Attribute::InReg) ||
      F->getAttributes().hasAttribute(A->getArgNo() + 1, Attribute::ByVal))
    return true;

  // Everything else is in VGPRs.
  return false;
}

///
/// \returns true if the result of the value could potentially be
/// different across workitems in a wavefront.
bool AMDGPUTTIImpl::isSourceOfDivergence(const Value *V) const {

  if (const Argument *A = dyn_cast<Argument>(V))
    return !isArgPassedInSGPR(A);

  // Loads from the private address space are divergent, because threads
  // can execute the load instruction with the same inputs and get different
  // results.
  //
  // All other loads are not divergent, because if threads issue loads with the
  // same arguments, they will always get the same result.
  if (const LoadInst *Load = dyn_cast<LoadInst>(V))
    return Load->getPointerAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS;

  if (const IntrinsicInst *Intrinsic = dyn_cast<IntrinsicInst>(V)) {
    const TargetMachine &TM = getTLI()->getTargetMachine();
    return isIntrinsicSourceOfDivergence(TM.getIntrinsicInfo(), Intrinsic);
  }

  // Assume all function calls are a source of divergence.
  if (isa<CallInst>(V) || isa<InvokeInst>(V))
    return true;

  return false;
}