R600/SI: Fix assertion from copying a TargetGlobalAddress

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

diff --git a/lib/Target/R600/AMDGPUAsmPrinter.cpp b/lib/Target/R600/AMDGPUAsmPrinter.cpp
index 8c5fc846ed4c042fe789aef7cd3360ca9343fc45..7f0cbe575d09702c31a9493702fae34bbeeec4d4 100644 (file)
--- a/lib/Target/R600/AMDGPUAsmPrinter.cpp
+++ b/lib/Target/R600/AMDGPUAsmPrinter.cpp
@@ -25,6 +25,7 @@
 #include "SIDefines.h"
 #include "SIMachineFunctionInfo.h"
 #include "SIRegisterInfo.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCSectionELF.h"
 #include "llvm/MC/MCStreamer.h"
@@ -47,11 +48,28 @@ using namespace llvm;
 // precision, and leaves single precision to flush all and does not report
 // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports
 // CL_FP_DENORM for both.
-static uint32_t getFPMode(const MachineFunction &) {
+//
+// FIXME: It seems some instructions do not support single precision denormals
+// regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
+// and sin_f32, cos_f32 on most parts).
+
+// We want to use these instructions, and using fp32 denormals also causes
+// instructions to run at the double precision rate for the device so it's
+// probably best to just report no single precision denormals.
+static uint32_t getFPMode(const MachineFunction &F) {
+  const AMDGPUSubtarget& ST = F.getTarget().getSubtarget<AMDGPUSubtarget>();
+  // TODO: Is there any real use for the flush in only / flush out only modes?
+
+  uint32_t FP32Denormals =
+    ST.hasFP32Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
+
+  uint32_t FP64Denormals =
+    ST.hasFP64Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
+
   return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) |
          FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) |
-         FP_DENORM_MODE_SP(FP_DENORM_FLUSH_NONE) |
-         FP_DENORM_MODE_DP(FP_DENORM_FLUSH_NONE);
+         FP_DENORM_MODE_SP(FP32Denormals) |
+         FP_DENORM_MODE_DP(FP64Denormals);
 }
 
 static AsmPrinter *createAMDGPUAsmPrinterPass(TargetMachine &tm,
@@ -68,6 +86,16 @@ AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
   DisasmEnabled = TM.getSubtarget<AMDGPUSubtarget>().dumpCode();
 }
 
+void AMDGPUAsmPrinter::EmitEndOfAsmFile(Module &M) {
+
+  // This label is used to mark the end of the .text section.
+  const TargetLoweringObjectFile &TLOF = getObjFileLowering();
+  OutStreamer.SwitchSection(TLOF.getTextSection());
+  MCSymbol *EndOfTextLabel =
+      OutContext.GetOrCreateSymbol(StringRef(END_OF_TEXT_LABEL_NAME));
+  OutStreamer.EmitLabel(EndOfTextLabel);
+}
+
 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
   SetupMachineFunction(MF);
 
@@ -114,6 +142,8 @@ bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
                                  false);
       OutStreamer.emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode),
                                  false);
+      OutStreamer.emitRawComment(" ScratchSize: " + Twine(KernelInfo.ScratchSize),
+                                 false);
     } else {
       R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
       OutStreamer.emitRawComment(
@@ -147,8 +177,8 @@ bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
 void AMDGPUAsmPrinter::EmitProgramInfoR600(const MachineFunction &MF) {
   unsigned MaxGPR = 0;
   bool killPixel = false;
-  const R600RegisterInfo *RI
-    = static_cast<const R600RegisterInfo*>(TM.getRegisterInfo());
+  const R600RegisterInfo *RI = static_cast<const R600RegisterInfo *>(
+      TM.getSubtargetImpl()->getRegisterInfo());
   const R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
   const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
 
@@ -210,8 +240,8 @@ void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
   unsigned MaxSGPR = 0;
   unsigned MaxVGPR = 0;
   bool VCCUsed = false;
-  const SIRegisterInfo *RI
-    = static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
+  const SIRegisterInfo *RI = static_cast<const SIRegisterInfo *>(
+      TM.getSubtargetImpl()->getRegisterInfo());
 
   for (const MachineBasicBlock &MBB : MF) {
     for (const MachineInstr &MI : MBB) {
@@ -305,6 +335,9 @@ void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
   // Do not clamp NAN to 0.
   ProgInfo.DX10Clamp = 0;
 
+  const MachineFrameInfo *FrameInfo = MF.getFrameInfo();
+  ProgInfo.ScratchSize = FrameInfo->estimateStackSize(MF);
+
   ProgInfo.CodeLen = CodeSize;
 }
 
@@ -334,6 +367,15 @@ void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
   unsigned LDSBlocks =
     RoundUpToAlignment(MFI->LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
 
+  // Scratch is allocated in 256 dword blocks.
+  unsigned ScratchAlignShift = 10;
+  // We need to program the hardware with the amount of scratch memory that
+  // is used by the entire wave.  KernelInfo.ScratchSize is the amount of
+  // scratch memory used per thread.
+  unsigned ScratchBlocks =
+    RoundUpToAlignment(KernelInfo.ScratchSize * STM.getWavefrontSize(),
+                       1 << ScratchAlignShift) >> ScratchAlignShift;
+
   if (MFI->getShaderType() == ShaderType::COMPUTE) {
     OutStreamer.EmitIntValue(R_00B848_COMPUTE_PGM_RSRC1, 4);
 
@@ -350,7 +392,14 @@ void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
     OutStreamer.EmitIntValue(ComputePGMRSrc1, 4);
 
     OutStreamer.EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4);
-    OutStreamer.EmitIntValue(S_00B84C_LDS_SIZE(LDSBlocks), 4);
+    const uint32_t ComputePGMRSrc2 =
+      S_00B84C_LDS_SIZE(LDSBlocks) |
+      S_00B02C_SCRATCH_EN(ScratchBlocks > 0);
+
+    OutStreamer.EmitIntValue(ComputePGMRSrc2, 4);
+
+    OutStreamer.EmitIntValue(R_00B860_COMPUTE_TMPRING_SIZE, 4);
+    OutStreamer.EmitIntValue(S_00B860_WAVESIZE(ScratchBlocks), 4);
   } else {
     OutStreamer.EmitIntValue(RsrcReg, 4);
     OutStreamer.EmitIntValue(S_00B028_VGPRS(KernelInfo.NumVGPR / 4) |