[cleanup] Re-sort all the #include lines in LLVM using
[oota-llvm.git] / lib / Target / R600 / AMDGPUSubtarget.cpp
index e3c2a50ab828bb741d64cbff1be0e3476c3baf06..597e558e663475d5052dd5ad0c9aa61d1fdce114 100644 (file)
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPUSubtarget.h"
+#include "R600ISelLowering.h"
 #include "R600InstrInfo.h"
+#include "R600MachineScheduler.h"
+#include "SIISelLowering.h"
 #include "SIInstrInfo.h"
 #include "llvm/ADT/SmallString.h"
 
-#include "llvm/ADT/SmallString.h"
-
 using namespace llvm;
 
 #define DEBUG_TYPE "amdgpu-subtarget"
@@ -28,26 +29,23 @@ using namespace llvm;
 #define GET_SUBTARGETINFO_CTOR
 #include "AMDGPUGenSubtargetInfo.inc"
 
-AMDGPUSubtarget::AMDGPUSubtarget(StringRef TT, StringRef GPU, StringRef FS) :
-  AMDGPUGenSubtargetInfo(TT, GPU, FS),
-  DevName(GPU),
-  Is64bit(false),
-  DumpCode(false),
-  R600ALUInst(false),
-  HasVertexCache(false),
-  TexVTXClauseSize(0),
-  Gen(AMDGPUSubtarget::R600),
-  FP64(false),
-  FP64Denormals(false),
-  FP32Denormals(false),
-  CaymanISA(false),
-  EnableIRStructurizer(true),
-  EnablePromoteAlloca(false),
-  EnableIfCvt(true),
-  WavefrontSize(0),
-  CFALUBug(false),
-  LocalMemorySize(0),
-  InstrItins(getInstrItineraryForCPU(GPU)) {
+static std::string computeDataLayout(const AMDGPUSubtarget &ST) {
+  std::string Ret = "e-p:32:32";
+
+  if (ST.is64bit()) {
+    // 32-bit private, local, and region pointers. 64-bit global and constant.
+    Ret += "-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p24:64:64";
+  }
+
+  Ret += "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256"
+         "-v512:512-v1024:1024-v2048:2048-n32:64";
+
+  return Ret;
+}
+
+AMDGPUSubtarget &
+AMDGPUSubtarget::initializeSubtargetDependencies(StringRef GPU, StringRef FS) {
+  // Determine default and user-specified characteristics
   // 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.
@@ -61,16 +59,37 @@ AMDGPUSubtarget::AMDGPUSubtarget(StringRef TT, StringRef GPU, StringRef FS) :
 
   ParseSubtargetFeatures(GPU, FullFS);
 
+  // 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?
   if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
-    InstrInfo.reset(new R600InstrInfo(*this));
-
-    // 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?
     FP32Denormals = false;
     FP64Denormals = false;
+  }
+  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),
+      FP64Denormals(false), FP32Denormals(false), CaymanISA(false),
+      FlatAddressSpace(false), EnableIRStructurizer(true),
+      EnablePromoteAlloca(false), EnableIfCvt(true),
+      EnableLoadStoreOpt(false), WavefrontSize(0), CFALUBug(false), LocalMemorySize(0),
+      DL(computeDataLayout(initializeSubtargetDependencies(GPU, FS))),
+      FrameLowering(TargetFrameLowering::StackGrowsUp,
+                    64 * 16, // Maximum stack alignment (long16)
+                    0),
+      InstrItins(getInstrItineraryForCPU(GPU)),
+      TargetTriple(TT) {
+  if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
+    InstrInfo.reset(new R600InstrInfo(*this));
+    TLInfo.reset(new R600TargetLowering(TM));
   } else {
     InstrInfo.reset(new SIInstrInfo(*this));
+    TLInfo.reset(new SITargetLowering(TM));
   }
 }
 
@@ -87,3 +106,10 @@ unsigned AMDGPUSubtarget::getStackEntrySize() const {
     llvm_unreachable("Illegal wavefront size.");
   }
 }
+
+unsigned AMDGPUSubtarget::getAmdKernelCodeChipID() const {
+  switch(getGeneration()) {
+  default: llvm_unreachable("ChipID unknown");
+  case SEA_ISLANDS: return 12;
+  }
+}