R600/SI: implement range reduction for sin/cos

[oota-llvm.git] / lib / Target / R600 / AMDGPUTargetTransformInfo.cpp

diff --git a/lib/Target/R600/AMDGPUTargetTransformInfo.cpp b/lib/Target/R600/AMDGPUTargetTransformInfo.cpp
index 9cdfec59412052fbf0c41a3372f67e360c0e84aa..88934b65876e4893694c0c13d0daea2dfadfcab8 100644 (file)
--- a/lib/Target/R600/AMDGPUTargetTransformInfo.cpp
+++ b/lib/Target/R600/AMDGPUTargetTransformInfo.cpp
@@ -15,7 +15,6 @@
 //
 //===----------------------------------------------------------------------===//
 
-#define DEBUG_TYPE "AMDGPUtti"
 #include "AMDGPU.h"
 #include "AMDGPUTargetMachine.h"
 #include "llvm/Analysis/LoopInfo.h"
@@ -26,6 +25,8 @@
 #include "llvm/Target/TargetLowering.h"
 using namespace llvm;
 
+#define DEBUG_TYPE "AMDGPUtti"
+
 // Declare the pass initialization routine locally as target-specific passes
 // don't have a target-wide initialization entry point, and so we rely on the
 // pass constructor initialization.
@@ -35,7 +36,7 @@ void initializeAMDGPUTTIPass(PassRegistry &);
 
 namespace {
 
-class AMDGPUTTI LLVM_FINAL : public ImmutablePass, public TargetTransformInfo {
+class AMDGPUTTI final : public ImmutablePass, public TargetTransformInfo {
   const AMDGPUTargetMachine *TM;
   const AMDGPUSubtarget *ST;
   const AMDGPUTargetLowering *TLI;
@@ -45,7 +46,7 @@ class AMDGPUTTI LLVM_FINAL : public ImmutablePass, public TargetTransformInfo {
   unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
 
 public:
-  AMDGPUTTI() : ImmutablePass(ID), TM(0), ST(0), TLI(0) {
+  AMDGPUTTI() : ImmutablePass(ID), TM(nullptr), ST(nullptr), TLI(nullptr) {
     llvm_unreachable("This pass cannot be directly constructed");
   }
 
@@ -55,11 +56,9 @@ public:
     initializeAMDGPUTTIPass(*PassRegistry::getPassRegistry());
   }
 
-  virtual void initializePass() LLVM_OVERRIDE { pushTTIStack(this); }
+  void initializePass() override { pushTTIStack(this); }
 
-  virtual void finalizePass() { popTTIStack(); }
-
-  virtual void getAnalysisUsage(AnalysisUsage &AU) const LLVM_OVERRIDE {
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
     TargetTransformInfo::getAnalysisUsage(AU);
   }
 
@@ -67,15 +66,22 @@ public:
   static char ID;
 
   /// Provide necessary pointer adjustments for the two base classes.
-  virtual void *getAdjustedAnalysisPointer(const void *ID) LLVM_OVERRIDE {
+  void *getAdjustedAnalysisPointer(const void *ID) override {
     if (ID == &TargetTransformInfo::ID)
       return (TargetTransformInfo *)this;
     return this;
   }
 
-  virtual bool hasBranchDivergence() const LLVM_OVERRIDE;
+  bool hasBranchDivergence() const override;
+
+  void getUnrollingPreferences(Loop *L,
+                               UnrollingPreferences &UP) const override;
 
-  virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
+  PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const override;
+
+  unsigned getNumberOfRegisters(bool Vector) const override;
+  unsigned getRegisterBitWidth(bool Vector) const override;
+  unsigned getMaximumUnrollFactor() const override;
 
   /// @}
 };
@@ -95,14 +101,12 @@ bool AMDGPUTTI::hasBranchDivergence() const { return true; }
 
 void AMDGPUTTI::getUnrollingPreferences(Loop *L,
                                         UnrollingPreferences &UP) const {
-  for (Loop::block_iterator BI = L->block_begin(), BE = L->block_end();
-                                                  BI != BE; ++BI) {
-    BasicBlock *BB = *BI;
-    for (BasicBlock::const_iterator I = BB->begin(), E = BB->end();
-                                                      I != E; ++I) {
-      const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I);
-      if (!GEP)
+  for (const BasicBlock *BB : L->getBlocks()) {
+    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));
       if (Alloca) {
@@ -111,13 +115,39 @@ void AMDGPUTTI::getUnrollingPreferences(Loop *L,
         // 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.
+        // 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 = 500;
       }
     }
   }
 }
+
+AMDGPUTTI::PopcntSupportKind
+AMDGPUTTI::getPopcntSupport(unsigned TyWidth) const {
+  assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
+  return ST->hasBCNT(TyWidth) ? PSK_FastHardware : PSK_Software;
+}
+
+unsigned AMDGPUTTI::getNumberOfRegisters(bool Vec) const {
+  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 AMDGPUTTI::getRegisterBitWidth(bool) const {
+  return 32;
+}
+
+unsigned AMDGPUTTI::getMaximumUnrollFactor() const {
+  // Semi-arbitrary large amount.
+  return 64;
+}