radeonsi: initial WIP SI code

This commit adds initial support for acceleration
on SI chips.  egltri is starting to work.

The SI/R600 llvm backend is currently included in mesa
but that may change in the future.

The plan is to write a single gallium driver and
use gallium to support X acceleration.

This commit contains patches from:
Tom Stellard <[email protected]>
Michel Dänzer <[email protected]>
Alex Deucher <[email protected]>
Vadim Girlin <[email protected]>

Signed-off-by: Alex Deucher <[email protected]>

The following commits were squashed in:

======================================================================

radeonsi: Remove unused winsys pointer

This was removed from r600g in commit:

commit 96d882939d612fcc8332f107befec470ed4359de
Author: Marek Olšák <[email protected]>
Date:   Fri Feb 17 01:49:49 2012 +0100

    gallium: remove unused winsys pointers in pipe_screen and pipe_context

    A winsys is already a private object of a driver.

======================================================================

radeonsi: Copy color clamping CAPs from r600

Not sure if the values of these CAPS are correct for radeonsi, but the
same changed were made to r600g in commit:

commit bc1c8369384b5e16547c5bf9728aa78f8dfd66cc
Author: Marek Olšák <[email protected]>
Date:   Mon Jan 23 03:11:17 2012 +0100

    st/mesa: do vertex and fragment color clamping in shaders

    For ARB_color_buffer_float. Most hardware can't do it and st/mesa is
    the perfect place for a fallback.
    The exceptions are:
    - r500 (vertex clamp only)
    - nv50 (both)
    - nvc0 (both)
    - softpipe (both)

    We also have to take into account that r300 can do CLAMPED vertex colors only,
    while r600 can do UNCLAMPED vertex colors only. The difference can be expressed
    with the two new CAPs.

======================================================================

radeonsi: Remove PIPE_CAP_OUTPUT_READ

This CAP was dropped in commit:

commit 04e324008759282728a95a1394bac2c4c2a1a3f9
Author: Marek Olšák <[email protected]>
Date:   Thu Feb 23 23:44:36 2012 +0100

    gallium: remove PIPE_SHADER_CAP_OUTPUT_READ

    r600g is the only driver which has made use of it. The reason the CAP was
    added was to fix some piglit tests when the GLSL pass lower_output_reads
    didn't exist.

    However, not removing output reads breaks the fallback for glClampColorARB,
    which assumes outputs are not readable. The fix would be non-trivial
    and my personal preference is to remove the CAP, considering that reading
    outputs is uncommon and that we can now use lower_output_reads to fix
    the issue that the CAP was supposed to workaround in the first place.

======================================================================

radeonsi: Add missing parameters to rws->buffer_get_tiling() call

This was changed in commit:

commit c0c979eebc076b95cc8d18a013ce2968fe6311ad
Author: Jerome Glisse <[email protected]>
Date:   Mon Jan 30 17:22:13 2012 -0500

    r600g: add support for common surface allocator for tiling v13

    Tiled surface have all kind of alignment constraint that needs to
    be met. Instead of having all this code duplicated btw ddx and
    mesa use common code in libdrm_radeon this also ensure that both
    ddx and mesa compute those alignment in the same way.

    v2 fix evergreen
    v3 fix compressed texture and workaround cube texture issue by
       disabling 2D array mode for cubemap (need to check if r7xx and
       newer are also affected by the issue)
    v4 fix texture array
    v5 fix evergreen and newer, split surface values computation from
       mipmap tree generation so that we can get them directly from the
       ddx
    v6 final fix to evergreen tile split value
    v7 fix mipmap offset to avoid to use random value, use color view
       depth view to address different layer as hardware is doing some
       magic rotation depending on the layer
    v8 fix COLOR_VIEW on r6xx for linear array mode, use COLOR_VIEW on
       evergreen, align bytes per pixel to a multiple of a dword
    v9 fix handling of stencil on evergreen, half fix for compressed
       texture
    v10 fix evergreen compressed texture proper support for stencil
        tile split. Fix stencil issue when array mode was clear by
        the kernel, always program stencil bo. On evergreen depth
        buffer bo need to be big enough to hold depth buffer + stencil
        buffer as even with stencil disabled things get written there.
    v11 rebase on top of mesa, fix pitch issue with 1d surface on evergreen,
        old ddx overestimate those. Fix linear case when pitch*height < 64.
        Fix r300g.
    v12 Fix linear case when pitch*height < 64 for old path, adapt to
        libdrm API change
    v13 add libdrm check

    Signed-off-by: Jerome Glisse <[email protected]>

======================================================================

radeonsi: Remove PIPE_TRANSFER_MAP_PERMANENTLY

This was removed in commit:

commit 62f44f670bb0162e89fd4786af877f8da9ff607c
Author: Marek Olšák <[email protected]>
Date:   Mon Mar 5 13:45:00 2012 +0100

    Revert "gallium: add flag PIPE_TRANSFER_MAP_PERMANENTLY"

    This reverts commit 0950086376b1c8b7fb89eda81ed7f2f06dee58bc.

    It was decided to refactor the transfer API instead of adding workarounds
    to address the performance issues.

======================================================================

radeonsi: Handle PIPE_VIDEO_CAP_PREFERED_FORMAT.

Reintroduced in commit 9d9afcb5bac2931d4b8e6d1aa571e941c5110c90.

======================================================================

radeonsi: nuke the fallback for vertex and fragment color clamping

Ported from r600g commit c2b800cf38b299c1ab1c53dc0e4ea00c7acef853.

======================================================================

radeonsi: don't expose transform_feedback2 without kernel support

Ported from r600g commit 15146fd1bcbb08e44a1cbb984440ee1a5de63d48.

======================================================================

radeonsi: Handle PIPE_CAP_GLSL_FEATURE_LEVEL.

Ported from r600g part of commit 171be755223d99f8cc5cc1bdaf8bd7b4caa04b4f.

======================================================================

radeonsi: set minimum point size to 1.0 for non-sprite non-aa points.

Ported from r600g commit f183cc9ce3ad1d043bdf8b38fd519e8f437714fc.

======================================================================

radeonsi: rework and consolidate stencilref state setting.

Ported from r600g commit a2361946e782b57f0c63587841ca41c0ea707070.

======================================================================

radeonsi: cleanup setting DB_SHADER_CONTROL.

Ported from r600g commit 3d061caaed13b646ff40754f8ebe73f3d4983c5b.

======================================================================

radeonsi: Get rid of register masks.

Ported from r600g commits
3d061caaed13b646ff40754f8ebe73f3d4983c5b..9344ab382a1765c1a7c2560e771485edf4954fe2.

======================================================================

radeonsi: get rid of r600_context_reg.

Ported from r600g commits
9344ab382a1765c1a7c2560e771485edf4954fe2..bed20f02a771f43e1c5092254705701c228cfa7f.

======================================================================

radeonsi: Fix regression from 'Get rid of register masks'.

======================================================================

radeonsi: optimize r600_resource_va.

Ported from r600g commit 669d8766ff3403938794eb80d7769347b6e52174.

======================================================================

radeonsi: remove u8,u16,u32,u64 types.

Ported from r600g commit 78293b99b23268e6698f1267aaf40647c17d95a5.

======================================================================

radeonsi: merge r600_context with r600_pipe_context.

Ported from r600g commit e4340c1908a6a3b09e1a15d5195f6da7d00494d0.

======================================================================

radeonsi: Miscellaneous context cleanups.

Ported from r600g commits
e4340c1908a6a3b09e1a15d5195f6da7d00494d0..621e0db71c5ddcb379171064a4f720c9cf01e888.

======================================================================

radeonsi: add a new simple API for state emission.

Ported from r600g commits
621e0db71c5ddcb379171064a4f720c9cf01e888..f661405637bba32c2cfbeecf6e2e56e414e9521e.

======================================================================

radeonsi: Also remove sbu_flags member of struct r600_reg.

Requires using sid.h instead of r600d.h for the new CP_COHER_CNTL definitions,
so some code needs to be disabled for now.

======================================================================

radeonsi: Miscellaneous simplifications.

Ported from r600g commits 38bf2763482b4f1b6d95cd51aecec75601d8b90f and
b0337b679ad4c2feae59215104cfa60b58a619d5.

======================================================================

radeonsi: Handle PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION.

Ported from commit 8b4f7b0672d663273310fffa9490ad996f5b914a.

======================================================================

radeonsi: Use a fake reloc to sleep for fences.

Ported from r600g commit 8cd03b933cf868ff867e2db4a0937005a02fd0e4.

======================================================================

radeonsi: adapt to get_query_result interface change.

Ported from r600g commit 4445e170bee23a3607ece0e010adef7058ac6a11.

1 files changed, 1353 insertions, 0 deletions

diff --git a/src/gallium/drivers/radeon/AMDILGlobalManager.cpp b/src/gallium/drivers/radeon/AMDILGlobalManager.cpp
new file mode 100644
index 00000000000..eafd36eaa4e
--- /dev/null
+++ b/src/gallium/drivers/radeon/AMDILGlobalManager.cpp
@@ -0,0 +1,1353 @@
+//===-- AMDILGlobalManager.cpp - TODO: Add brief description -------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+#include "AMDILGlobalManager.h"
+#include "AMDILDevices.h"
+#include "AMDILKernelManager.h"
+#include "AMDILSubtarget.h"
+
+#include "AMDILAlgorithms.tpp"
+#include "AMDILGlobalManager.h"
+#include "AMDILDevices.h"
+#include "AMDILKernelManager.h"
+#include "AMDILSubtarget.h"
+#include "AMDILUtilityFunctions.h"
+#include "llvm/CodeGen/MachineConstantPool.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Instructions.h"
+#include "llvm/Support/FormattedStream.h"
+
+#include <cstdio>
+
+using namespace llvm;
+
+AMDILGlobalManager::AMDILGlobalManager(bool debugMode) {
+  mOffset = 0;
+  mReservedBuffs = 0;
+  symTab = NULL;
+  mCurrentCPOffset = 0;
+  mDebugMode = debugMode;
+}
+
+AMDILGlobalManager::~AMDILGlobalManager() {
+}
+
+void AMDILGlobalManager::print(llvm::raw_ostream &O) {
+  if (!mDebugMode) {
+    return;
+  }
+  O << ";AMDIL Global Manager State Dump:\n";
+  O << ";\tSubtarget: " << mSTM << "\tSymbol Table: " << symTab
+    << "\n";
+  O << ";\tConstant Offset: " << mOffset << "\tCP Offset: "
+    << mCurrentCPOffset << "\tReserved Buffers: " << mReservedBuffs
+    << "\n";
+  if (!mImageNameMap.empty()) {
+    llvm::DenseMap<uint32_t, llvm::StringRef>::iterator imb, ime;
+    O << ";\tGlobal Image Mapping: \n";
+    for (imb = mImageNameMap.begin(), ime = mImageNameMap.end(); imb != ime;
+         ++imb) {
+      O << ";\t\tImage ID: " << imb->first << "\tName: "
+        << imb->second << "\n";
+    }
+  }
+  std::set<llvm::StringRef>::iterator sb, se;
+  if (!mByteStore.empty()) {
+    O << ";Byte Store Kernels: \n";
+    for (sb = mByteStore.begin(), se = mByteStore.end(); sb != se; ++sb) {
+      O << ";\t\t" << *sb << "\n";
+    }
+  }
+  if (!mIgnoreStr.empty()) {
+    O << ";\tIgnored Data Strings: \n";
+    for (sb = mIgnoreStr.begin(), se = mIgnoreStr.end(); sb != se; ++sb) {
+      O << ";\t\t" << *sb << "\n";
+    }
+  }
+}
+
+void AMDILGlobalManager::dump() {
+  print(errs());
+}
+
+static const constPtr *getConstPtr(const kernel &krnl, const std::string &arg) {
+  llvm::SmallVector<constPtr, DEFAULT_VEC_SLOTS>::const_iterator begin, end;
+  for (begin = krnl.constPtr.begin(), end = krnl.constPtr.end();
+       begin != end; ++begin) {
+    if (!strcmp(begin->name.data(),arg.c_str())) {
+      return &(*begin);
+    }
+  }
+  return NULL;
+}
+#if 0
+static bool structContainsSub32bitType(const StructType *ST) {
+  StructType::element_iterator eib, eie;
+  for (eib = ST->element_begin(), eie = ST->element_end(); eib != eie; ++eib) {
+    Type *ptr = *eib;
+    uint32_t size = (uint32_t)GET_SCALAR_SIZE(ptr);
+    if (!size) {
+      if (const StructType *ST = dyn_cast<StructType>(ptr)) {
+        if (structContainsSub32bitType(ST)) {
+          return true;
+        }
+      }
+    } else if (size < 32) {
+      return true;
+    }
+  }
+  return false;
+}
+#endif
+
+void AMDILGlobalManager::processModule(const Module &M,
+                                       const AMDILTargetMachine *mTM)
+{
+  Module::const_global_iterator GI;
+  Module::const_global_iterator GE;
+  symTab = "NoSymTab";
+  mSTM = mTM->getSubtargetImpl();
+  for (GI = M.global_begin(), GE = M.global_end(); GI != GE; ++GI) {
+    const GlobalValue *GV = GI;
+    if (mDebugMode) {
+      GV->dump();
+      errs() << "\n";
+    }
+    llvm::StringRef GVName = GV->getName();
+    const char *name = GVName.data();
+    if (!strncmp(name, "sgv", 3)) {
+      mKernelArgs[GVName] = parseSGV(GV);
+    } else if (!strncmp(name, "fgv", 3)) {
+      // we can ignore this since we don't care about the filename
+      // string
+    } else if (!strncmp(name, "lvgv", 4)) {
+      mLocalArgs[GVName] = parseLVGV(GV);
+    } else if (!strncmp(name, "llvm.image.annotations", 22)) {
+      if (strstr(name, "__OpenCL")
+          && strstr(name, "_kernel")) {
+        // we only want to parse the image information if the
+        // image is a kernel, we might have to parse out the
+        // information if a function is found that is not
+        // inlined.
+        parseImageAnnotate(GV);
+      }
+    } else if (!strncmp(name, "llvm.global.annotations", 23)) {
+      parseGlobalAnnotate(GV);
+    } else if (!strncmp(name, "llvm.constpointer.annotations", 29)) {
+      if (strstr(name, "__OpenCL")
+          && strstr(name, "_kernel")) {
+        // we only want to parse constant pointer information
+        // if it is a kernel
+        parseConstantPtrAnnotate(GV);
+      }
+    } else if (!strncmp(name, "llvm.readonlypointer.annotations", 32)) {
+      // These are skipped as we handle them later in AMDILPointerManager.cpp
+    } else if (GV->getType()->getAddressSpace() == 3) { // *** Match cl_kernel.h local AS #
+      parseAutoArray(GV, false);
+    } else if (strstr(name, "clregion")) {
+      parseAutoArray(GV, true);
+    } else if (!GV->use_empty()
+               && mIgnoreStr.find(GVName) == mIgnoreStr.end()) {
+      parseConstantPtr(GV);
+    }
+  }
+  allocateGlobalCB();
+
+  safeForEach(M.begin(), M.end(),
+      std::bind1st(
+        std::mem_fun(&AMDILGlobalManager::checkConstPtrsUseHW),
+        this));
+}
+
+void AMDILGlobalManager::allocateGlobalCB(void) {
+  uint32_t maxCBSize = mSTM->device()->getMaxCBSize();
+  uint32_t offset = 0;
+  uint32_t curCB = 0;
+  uint32_t swoffset = 0;
+  for (StringMap<constPtr>::iterator cpb = mConstMems.begin(),
+       cpe = mConstMems.end(); cpb != cpe; ++cpb) {
+    bool constHW = mSTM->device()->usesHardware(AMDILDeviceInfo::ConstantMem);
+    cpb->second.usesHardware = false;
+    if (constHW) {
+      // If we have a limit on the max CB Size, then we need to make sure that
+      // the constant sizes fall within the limits.
+      if (cpb->second.size <= maxCBSize) {
+        if (offset + cpb->second.size > maxCBSize) {
+          offset = 0;
+          curCB++;
+        }
+        if (curCB < mSTM->device()->getMaxNumCBs()) {
+          cpb->second.cbNum = curCB + CB_BASE_OFFSET;
+          cpb->second.offset = offset;
+          offset += (cpb->second.size + 15) & (~15);
+          cpb->second.usesHardware = true;
+          continue;
+        }
+      }
+    }
+    cpb->second.cbNum = 0;
+    cpb->second.offset = swoffset;
+    swoffset += (cpb->second.size + 15) & (~15);
+  }
+  if (!mConstMems.empty()) {
+    mReservedBuffs = curCB + 1;
+  }
+}
+
+bool AMDILGlobalManager::checkConstPtrsUseHW(llvm::Module::const_iterator *FCI)
+{
+  Function::const_arg_iterator AI, AE;
+  const Function *func = *FCI;
+  std::string name = func->getName();
+  if (!strstr(name.c_str(), "__OpenCL")
+      || !strstr(name.c_str(), "_kernel")) {
+    return false;
+  }
+  kernel &krnl =  mKernels[name];
+  if (mSTM->device()->usesHardware(AMDILDeviceInfo::ConstantMem)) {
+    for (AI = func->arg_begin(), AE = func->arg_end();
+         AI != AE; ++AI) {
+      const Argument *Arg = &(*AI);
+      const PointerType *P = dyn_cast<PointerType>(Arg->getType());
+      if (!P) {
+        continue;
+      }
+      if (P->getAddressSpace() != AMDILAS::CONSTANT_ADDRESS) {
+        continue;
+      }
+      const constPtr *ptr = getConstPtr(krnl, Arg->getName());
+      if (ptr) {
+        continue;
+      }
+      constPtr constAttr;
+      constAttr.name = Arg->getName();
+      constAttr.size = this->mSTM->device()->getMaxCBSize();
+      constAttr.base = Arg;
+      constAttr.isArgument = true;
+      constAttr.isArray = false;
+      constAttr.offset = 0;
+      constAttr.usesHardware =
+        mSTM->device()->usesHardware(AMDILDeviceInfo::ConstantMem);
+      if (constAttr.usesHardware) {
+        constAttr.cbNum = krnl.constPtr.size() + 2;
+      } else {
+        constAttr.cbNum = 0;
+      }
+      krnl.constPtr.push_back(constAttr);
+    }
+  }
+  // Now lets make sure that only the N largest buffers
+  // get allocated in hardware if we have too many buffers
+  uint32_t numPtrs = krnl.constPtr.size();
+  if (numPtrs > (this->mSTM->device()->getMaxNumCBs() - mReservedBuffs)) {
+    // TODO: Change this routine so it sorts
+    // constPtr instead of pulling the sizes out
+    // and then grab the N largest and disable the rest
+    llvm::SmallVector<uint32_t, 16> sizes;
+    for (uint32_t x = 0; x < numPtrs; ++x) {
+      sizes.push_back(krnl.constPtr[x].size);
+    }
+    std::sort(sizes.begin(), sizes.end());
+    uint32_t numToDisable = numPtrs - (mSTM->device()->getMaxNumCBs() -
+                                       mReservedBuffs);
+    uint32_t safeSize = sizes[numToDisable-1];
+    for (uint32_t x = 0; x < numPtrs && numToDisable; ++x) {
+      if (krnl.constPtr[x].size <= safeSize) {
+        krnl.constPtr[x].usesHardware = false;
+        --numToDisable;
+      }
+    }
+  }
+  // Renumber all of the valid CB's so that
+  // they are linear increase
+  uint32_t CBid = 2 + mReservedBuffs;
+  for (uint32_t x = 0; x < numPtrs; ++x) {
+    if (krnl.constPtr[x].usesHardware) {
+      krnl.constPtr[x].cbNum = CBid++;
+    }
+  }
+  for (StringMap<constPtr>::iterator cpb = mConstMems.begin(),
+       cpe = mConstMems.end(); cpb != cpe; ++cpb) {
+    if (cpb->second.usesHardware) {
+      krnl.constPtr.push_back(cpb->second);
+    }
+  }
+  for (uint32_t x = 0; x < krnl.constPtr.size(); ++x) {
+    constPtr &c = krnl.constPtr[x];
+    uint32_t cbNum = c.cbNum - CB_BASE_OFFSET;
+    if (cbNum < HW_MAX_NUM_CB && c.cbNum >= CB_BASE_OFFSET) {
+      if ((c.size + c.offset) > krnl.constSizes[cbNum]) {
+        krnl.constSizes[cbNum] =
+          ((c.size + c.offset) + 15) & ~15;
+      }
+    } else {
+      krnl.constPtr[x].usesHardware = false;
+    }
+  }
+  return false;
+}
+
+int32_t AMDILGlobalManager::getArrayOffset(const llvm::StringRef &a) const {
+  StringMap<arraymem>::const_iterator iter = mArrayMems.find(a);
+  if (iter != mArrayMems.end()) {
+    return iter->second.offset;
+  } else {
+    return -1;
+  }
+}
+
+int32_t AMDILGlobalManager::getConstOffset(const llvm::StringRef &a) const {
+  StringMap<constPtr>::const_iterator iter = mConstMems.find(a);
+  if (iter != mConstMems.end()) {
+    return iter->second.offset;
+  } else {
+    return -1;
+  }
+}
+
+bool AMDILGlobalManager::getConstHWBit(const llvm::StringRef &name) const {
+  StringMap<constPtr>::const_iterator iter = mConstMems.find(name);
+  if (iter != mConstMems.end()) {
+    return iter->second.usesHardware;
+  } else {
+    return false;
+  }
+}
+
+// As of right now we only care about the required group size
+// so we can skip the variable encoding
+kernelArg AMDILGlobalManager::parseSGV(const GlobalValue *G) {
+  kernelArg nArg;
+  const GlobalVariable *GV = dyn_cast<GlobalVariable>(G);
+  memset(&nArg, 0, sizeof(nArg));
+  for (int x = 0; x < 3; ++x) {
+    nArg.reqGroupSize[x] = mSTM->getDefaultSize(x);
+    nArg.reqRegionSize[x] = mSTM->getDefaultSize(x);
+  }
+  if (!GV || !GV->hasInitializer()) {
+    return nArg;
+  }
+  const Constant *CV = GV->getInitializer();
+  const ConstantDataArray *CA =dyn_cast_or_null<ConstantDataArray>(CV);
+
+  if (!CA || !CA->isString()) {
+    return nArg;
+  }
+  std::string init = CA->getAsString();
+  size_t pos = init.find("RWG");
+  if (pos != llvm::StringRef::npos) {
+    pos += 3;
+    std::string LWS = init.substr(pos, init.length() - pos);
+    const char *lws = LWS.c_str();
+    sscanf(lws, "%u,%u,%u", &(nArg.reqGroupSize[0]),
+           &(nArg.reqGroupSize[1]),
+           &(nArg.reqGroupSize[2]));
+    nArg.mHasRWG = true;
+  }
+  pos = init.find("RWR");
+  if (pos != llvm::StringRef::npos) {
+    pos += 3;
+    std::string LWS = init.substr(pos, init.length() - pos);
+    const char *lws = LWS.c_str();
+    sscanf(lws, "%u,%u,%u", &(nArg.reqRegionSize[0]),
+           &(nArg.reqRegionSize[1]),
+           &(nArg.reqRegionSize[2]));
+    nArg.mHasRWR = true;
+  }
+  return nArg;
+}
+
+localArg AMDILGlobalManager::parseLVGV(const GlobalValue *G) {
+  localArg nArg;
+  const GlobalVariable *GV = dyn_cast<GlobalVariable>(G);
+  nArg.name = "";
+  if (!GV || !GV->hasInitializer()) {
+    return nArg;
+  }
+  const ConstantArray *CA =
+    dyn_cast_or_null<ConstantArray>(GV->getInitializer());
+  if (!CA) {
+    return nArg;
+  }
+  for (size_t x = 0, y = CA->getNumOperands(); x < y; ++x) {
+    const Value *local = CA->getOperand(x);
+    const ConstantExpr *CE = dyn_cast_or_null<ConstantExpr>(local);
+    if (!CE || !CE->getNumOperands()) {
+      continue;
+    }
+    nArg.name = (*(CE->op_begin()))->getName();
+    if (mArrayMems.find(nArg.name) != mArrayMems.end()) {
+      nArg.local.push_back(&(mArrayMems[nArg.name]));
+    }
+  }
+  return nArg;
+}
+
+void AMDILGlobalManager::parseConstantPtrAnnotate(const GlobalValue *G) {
+  const GlobalVariable *GV = dyn_cast_or_null<GlobalVariable>(G);
+  const ConstantArray *CA =
+    dyn_cast_or_null<ConstantArray>(GV->getInitializer());
+  if (!CA) {
+    return;
+  }
+  uint32_t numOps = CA->getNumOperands();
+  for (uint32_t x = 0; x < numOps; ++x) {
+    const Value *V = CA->getOperand(x);
+    const ConstantStruct *CS = dyn_cast_or_null<ConstantStruct>(V);
+    if (!CS) {
+      continue;
+    }
+    assert(CS->getNumOperands() == 2 && "There can only be 2"
+           " fields, a name and size");
+    const ConstantExpr *nameField = dyn_cast<ConstantExpr>(CS->getOperand(0));
+    const ConstantInt *sizeField = dyn_cast<ConstantInt>(CS->getOperand(1));
+    assert(nameField && "There must be a constant name field");
+    assert(sizeField && "There must be a constant size field");
+    const GlobalVariable *nameGV =
+      dyn_cast<GlobalVariable>(nameField->getOperand(0));
+    const ConstantDataArray *nameArray =
+      dyn_cast<ConstantDataArray>(nameGV->getInitializer());
+    // Lets add this string to the set of strings we should ignore processing
+    mIgnoreStr.insert(nameGV->getName());
+    if (mConstMems.find(nameGV->getName())
+        != mConstMems.end()) {
+      // If we already processesd this string as a constant, lets remove it from
+      // the list of known constants.  This way we don't process unneeded data
+      // and don't generate code/metadata for strings that are never used.
+      mConstMems.erase(mConstMems.find(nameGV->getName()));
+    } else {
+      mIgnoreStr.insert(CS->getOperand(0)->getName());
+    }
+    constPtr constAttr;
+    constAttr.name = nameArray->getAsString();
+    constAttr.size = (sizeField->getZExtValue() + 15) & ~15;
+    constAttr.base = CS;
+    constAttr.isArgument = true;
+    constAttr.isArray = false;
+    constAttr.cbNum = 0;
+    constAttr.offset = 0;
+    constAttr.usesHardware = (constAttr.size <= mSTM->device()->getMaxCBSize());
+    // Now that we have all our constant information,
+    // lets update the kernel
+    llvm::StringRef  kernelName = G->getName().data() + 30;
+    kernel k;
+    if (mKernels.find(kernelName) != mKernels.end()) {
+      k = mKernels[kernelName];
+    } else {
+      k.curSize = 0;
+      k.curRSize = 0;
+      k.curHWSize = 0;
+      k.curHWRSize = 0;
+      k.constSize = 0;
+      k.lvgv = NULL;
+      k.sgv = NULL;
+      memset(k.constSizes, 0, sizeof(uint32_t) * HW_MAX_NUM_CB);
+    }
+    constAttr.cbNum = k.constPtr.size() + 2;
+    k.constPtr.push_back(constAttr);
+    mKernels[kernelName] = k;
+  }
+}
+
+void AMDILGlobalManager::parseImageAnnotate(const GlobalValue *G) {
+  const GlobalVariable *GV = dyn_cast<GlobalVariable>(G);
+  const ConstantArray *CA = dyn_cast<ConstantArray>(GV->getInitializer());
+  if (!CA) {
+    return;
+  }
+  if (isa<GlobalValue>(CA)) {
+    return;
+  }
+  uint32_t e = CA->getNumOperands();
+  if (!e) {
+    return;
+  }
+  kernel k;
+  llvm::StringRef name = G->getName().data() + 23;
+  if (mKernels.find(name) != mKernels.end()) {
+    k = mKernels[name];
+  } else {
+    k.curSize = 0;
+    k.curRSize = 0;
+    k.curHWSize = 0;
+    k.curHWRSize = 0;
+    k.constSize = 0;
+    k.lvgv = NULL;
+    k.sgv = NULL;
+    memset(k.constSizes, 0, sizeof(uint32_t) * HW_MAX_NUM_CB);
+  }
+  for (uint32_t i = 0; i != e; ++i) {
+    const Value *V = CA->getOperand(i);
+    const Constant *C = dyn_cast<Constant>(V);
+    const ConstantStruct *CS = dyn_cast<ConstantStruct>(C);
+    if (CS && CS->getNumOperands() == 2) {
+      if (mConstMems.find(CS->getOperand(0)->getOperand(0)->getName()) !=
+          mConstMems.end()) {
+        // If we already processesd this string as a constant, lets remove it
+        // from the list of known constants.  This way we don't process unneeded
+        // data and don't generate code/metadata for strings that are never
+        // used.
+        mConstMems.erase(
+            mConstMems.find(CS->getOperand(0)->getOperand(0)->getName()));
+      } else {
+        mIgnoreStr.insert(CS->getOperand(0)->getOperand(0)->getName());
+      }
+      const ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(1));
+      uint32_t val = (uint32_t)CI->getZExtValue();
+      if (val == 1) {
+        k.readOnly.insert(i);
+      } else if (val == 2) {
+        k.writeOnly.insert(i);
+      } else {
+        assert(!"Unknown image type value!");
+      }
+    }
+  }
+  mKernels[name] = k;
+}
+
+void AMDILGlobalManager::parseAutoArray(const GlobalValue *GV, bool isRegion) {
+  const GlobalVariable *G = dyn_cast<GlobalVariable>(GV);
+  Type *Ty = (G) ? G->getType() : NULL;
+  arraymem tmp;
+  tmp.isHW = true;
+  tmp.offset = 0;
+  tmp.vecSize = getTypeSize(Ty, true);
+  tmp.isRegion = isRegion;
+  mArrayMems[GV->getName()] = tmp;
+}
+
+void AMDILGlobalManager::parseConstantPtr(const GlobalValue *GV) {
+  const GlobalVariable *G = dyn_cast<GlobalVariable>(GV);
+  Type *Ty = (G) ? G->getType() : NULL;
+  constPtr constAttr;
+  constAttr.name = G->getName();
+  constAttr.size = getTypeSize(Ty, true);
+  constAttr.base = GV;
+  constAttr.isArgument = false;
+  constAttr.isArray = true;
+  constAttr.offset = 0;
+  constAttr.cbNum = 0;
+  constAttr.usesHardware = false;
+  mConstMems[GV->getName()] = constAttr;
+}
+
+void AMDILGlobalManager::parseGlobalAnnotate(const GlobalValue *G) {
+  const GlobalVariable *GV = dyn_cast<GlobalVariable>(G);
+  if (!GV->hasInitializer()) {
+    return;
+  }
+  const Constant *CT = GV->getInitializer();
+  if (!CT || isa<GlobalValue>(CT)) {
+    return;
+  }
+  const ConstantArray *CA = dyn_cast<ConstantArray>(CT);
+  if (!CA) {
+    return;
+  }
+
+  unsigned int nKernels = CA->getNumOperands();
+  for (unsigned int i = 0, e = nKernels; i != e; ++i) {
+    parseKernelInformation(CA->getOperand(i));
+  }
+}
+
+void AMDILGlobalManager::parseKernelInformation(const Value *V) {
+  if (isa<GlobalValue>(V)) {
+    return;
+  }
+  const ConstantStruct *CS = dyn_cast_or_null<ConstantStruct>(V);
+  if (!CS) {
+    return;
+  }
+  uint32_t N = CS->getNumOperands();
+  if (N != 5) {
+    return;
+  }
+  kernel tmp;
+
+  tmp.curSize = 0;
+  tmp.curRSize = 0;
+  tmp.curHWSize = 0;
+  tmp.curHWRSize = 0;
+  // The first operand is always a pointer to the kernel.
+  const Constant *CV = dyn_cast<Constant>(CS->getOperand(0));
+  llvm::StringRef kernelName = "";
+  if (CV->getNumOperands()) {
+    kernelName = (*(CV->op_begin()))->getName();
+  }
+
+  // If we have images, then we have already created the kernel and we just need
+  // to get the kernel information.
+  if (mKernels.find(kernelName) != mKernels.end()) {
+    tmp = mKernels[kernelName];
+  } else {
+    tmp.curSize = 0;
+    tmp.curRSize = 0;
+    tmp.curHWSize = 0;
+    tmp.curHWRSize = 0;
+    tmp.constSize = 0;
+    tmp.lvgv = NULL;
+    tmp.sgv = NULL;
+    memset(tmp.constSizes, 0, sizeof(uint32_t) * HW_MAX_NUM_CB);
+  }
+
+
+  // The second operand is SGV, there can only be one so we don't need to worry
+  // about parsing out multiple data points.
+  CV = dyn_cast<Constant>(CS->getOperand(1));
+
+  llvm::StringRef sgvName;
+  if (CV->getNumOperands()) {
+    sgvName = (*(CV->op_begin()))->getName();
+  }
+
+  if (mKernelArgs.find(sgvName) != mKernelArgs.end()) {
+    tmp.sgv = &mKernelArgs[sgvName];
+  }
+  // The third operand is FGV, which is skipped
+  // The fourth operand is LVGV
+  // There can be multiple local arrays, so we
+  // need to handle each one seperatly
+  CV = dyn_cast<Constant>(CS->getOperand(3));
+  llvm::StringRef lvgvName = "";
+  if (CV->getNumOperands()) {
+    lvgvName = (*(CV->op_begin()))->getName();
+  }
+  if (mLocalArgs.find(lvgvName) != mLocalArgs.end()) {
+    localArg *ptr = &mLocalArgs[lvgvName];
+    tmp.lvgv = ptr;
+    llvm::SmallVector<arraymem *, DEFAULT_VEC_SLOTS>::iterator ib, ie;
+    for (ib = ptr->local.begin(), ie = ptr->local.end(); ib != ie; ++ib) {
+      if ((*ib)->isRegion) {
+        if ((*ib)->isHW) {
+          (*ib)->offset = tmp.curHWRSize;
+          tmp.curHWRSize += ((*ib)->vecSize + 15) & ~15;
+        } else {
+          (*ib)->offset = tmp.curRSize;
+          tmp.curRSize += ((*ib)->vecSize + 15) & ~15;
+        }
+      } else {
+        if ((*ib)->isHW) {
+          (*ib)->offset = tmp.curHWSize;
+          tmp.curHWSize += ((*ib)->vecSize + 15) & ~15;
+        } else {
+          (*ib)->offset = tmp.curSize;
+          tmp.curSize += ((*ib)->vecSize + 15) & ~15;
+        }
+      }
+    }
+  }
+
+  // The fifth operand is NULL
+  mKernels[kernelName] = tmp;
+}
+
+const kernel &AMDILGlobalManager::getKernel(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  assert(isKernel(name) && "Must be a kernel to call getKernel");
+  return iter->second;
+}
+
+bool AMDILGlobalManager::isKernel(const llvm::StringRef &name) const {
+  return (mKernels.find(name) != mKernels.end());
+}
+
+bool AMDILGlobalManager::isWriteOnlyImage(const llvm::StringRef &name,
+                                          uint32_t iID) const {
+  const StringMap<kernel>::const_iterator kiter = mKernels.find(name);
+  if (kiter == mKernels.end()) {
+    return false;
+  }
+  return kiter->second.writeOnly.count(iID);
+}
+
+uint32_t
+AMDILGlobalManager::getNumWriteImages(const llvm::StringRef &name) const {
+  char *env = NULL;
+  env = getenv("GPU_DISABLE_RAW_UAV");
+  if (env && env[0] == '1') {
+    return 8;
+  }
+  const StringMap<kernel>::const_iterator kiter = mKernels.find(name);
+  if (kiter == mKernels.end()) {
+    return 0;
+  } else {
+    return kiter->second.writeOnly.size();
+  }
+}
+
+bool AMDILGlobalManager::isReadOnlyImage(const llvm::StringRef &name,
+                                         uint32_t iID) const {
+  const StringMap<kernel>::const_iterator kiter = mKernels.find(name);
+  if (kiter == mKernels.end()) {
+    return false;
+  }
+  return kiter->second.readOnly.count(iID);
+}
+
+bool AMDILGlobalManager::hasRWG(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    kernelArg *ptr = iter->second.sgv;
+    if (ptr) {
+      return ptr->mHasRWG;
+    }
+  }
+  return false;
+}
+
+bool AMDILGlobalManager::hasRWR(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    kernelArg *ptr = iter->second.sgv;
+    if (ptr) {
+      return ptr->mHasRWR;
+    }
+  }
+  return false;
+}
+
+uint32_t
+AMDILGlobalManager::getMaxGroupSize(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    kernelArg *sgv = iter->second.sgv;
+    if (sgv) {
+      return sgv->reqGroupSize[0] * sgv->reqGroupSize[1] * sgv->reqGroupSize[2];
+    }
+  }
+  return mSTM->getDefaultSize(0) *
+         mSTM->getDefaultSize(1) *
+         mSTM->getDefaultSize(2);
+}
+
+uint32_t
+AMDILGlobalManager::getMaxRegionSize(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    kernelArg *sgv = iter->second.sgv;
+    if (sgv) {
+      return sgv->reqRegionSize[0] *
+             sgv->reqRegionSize[1] *
+             sgv->reqRegionSize[2];
+    }
+  }
+  return mSTM->getDefaultSize(0) *
+         mSTM->getDefaultSize(1) *
+         mSTM->getDefaultSize(2);
+}
+
+uint32_t AMDILGlobalManager::getRegionSize(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    return iter->second.curRSize;
+  } else {
+    return 0;
+  }
+}
+
+uint32_t AMDILGlobalManager::getLocalSize(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    return iter->second.curSize;
+  } else {
+    return 0;
+  }
+}
+
+uint32_t AMDILGlobalManager::getConstSize(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    return iter->second.constSize;
+  } else {
+    return 0;
+  }
+}
+
+uint32_t
+AMDILGlobalManager::getHWRegionSize(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    return iter->second.curHWRSize;
+  } else {
+    return 0;
+  }
+}
+
+uint32_t AMDILGlobalManager::getHWLocalSize(const llvm::StringRef &name) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end()) {
+    return iter->second.curHWSize;
+  } else {
+    return 0;
+  }
+}
+
+int32_t AMDILGlobalManager::getArgID(const Argument *arg) {
+  DenseMap<const Argument *, int32_t>::iterator argiter = mArgIDMap.find(arg);
+  if (argiter != mArgIDMap.end()) {
+    return argiter->second;
+  } else {
+    return -1;
+  }
+}
+
+
+uint32_t
+AMDILGlobalManager::getLocal(const llvm::StringRef &name, uint32_t dim) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end() && iter->second.sgv) {
+    kernelArg *sgv = iter->second.sgv;
+    switch (dim) {
+    default: break;
+    case 0:
+    case 1:
+    case 2:
+      return sgv->reqGroupSize[dim];
+      break;
+    case 3:
+      return sgv->reqGroupSize[0] * sgv->reqGroupSize[1] * sgv->reqGroupSize[2];
+    };
+  }
+  switch (dim) {
+  default:
+    return 1;
+  case 3:
+    return mSTM->getDefaultSize(0) *
+           mSTM->getDefaultSize(1) *
+           mSTM->getDefaultSize(2);
+  case 2:
+  case 1:
+  case 0:
+    return mSTM->getDefaultSize(dim);
+    break;
+  };
+  return 1;
+}
+
+uint32_t
+AMDILGlobalManager::getRegion(const llvm::StringRef &name, uint32_t dim) const {
+  StringMap<kernel>::const_iterator iter = mKernels.find(name);
+  if (iter != mKernels.end() && iter->second.sgv) {
+    kernelArg *sgv = iter->second.sgv;
+    switch (dim) {
+    default: break;
+    case 0:
+    case 1:
+    case 2:
+      return sgv->reqRegionSize[dim];
+      break;
+    case 3:
+      return sgv->reqRegionSize[0] *
+             sgv->reqRegionSize[1] *
+             sgv->reqRegionSize[2];
+    };
+  }
+  switch (dim) {
+  default:
+    return 1;
+  case 3:
+    return mSTM->getDefaultSize(0) *
+           mSTM->getDefaultSize(1) *
+           mSTM->getDefaultSize(2);
+  case 2:
+  case 1:
+  case 0:
+    return mSTM->getDefaultSize(dim);
+    break;
+  };
+  return 1;
+}
+
+StringMap<constPtr>::iterator AMDILGlobalManager::consts_begin() {
+  return mConstMems.begin();
+}
+
+
+StringMap<constPtr>::iterator AMDILGlobalManager::consts_end() {
+  return mConstMems.end();
+}
+
+bool AMDILGlobalManager::byteStoreExists(StringRef S) const {
+  return mByteStore.find(S) != mByteStore.end();
+}
+
+bool AMDILGlobalManager::usesHWConstant(const kernel &krnl,
+                                        const llvm::StringRef &arg) {
+  const constPtr *curConst = getConstPtr(krnl, arg);
+  if (curConst) {
+    return curConst->usesHardware;
+  } else {
+    return false;
+  }
+}
+
+uint32_t AMDILGlobalManager::getConstPtrSize(const kernel &krnl,
+                                             const llvm::StringRef &arg)
+{
+  const constPtr *curConst = getConstPtr(krnl, arg);
+  if (curConst) {
+    return curConst->size;
+  } else {
+    return 0;
+  }
+}
+
+uint32_t AMDILGlobalManager::getConstPtrOff(const kernel &krnl,
+                                            const llvm::StringRef &arg)
+{
+  const constPtr *curConst = getConstPtr(krnl, arg);
+  if (curConst) {
+    return curConst->offset;
+  } else {
+    return 0;
+  }
+}
+
+uint32_t AMDILGlobalManager::getConstPtrCB(const kernel &krnl,
+                                           const llvm::StringRef &arg)
+{
+  const constPtr *curConst = getConstPtr(krnl, arg);
+  if (curConst) {
+    return curConst->cbNum;
+  } else {
+    return 0;
+  }
+}
+
+void AMDILGlobalManager::calculateCPOffsets(const MachineFunction *MF,
+                                            kernel &krnl)
+{
+  const MachineConstantPool *MCP = MF->getConstantPool();
+  if (!MCP) {
+    return;
+  }
+  const std::vector<MachineConstantPoolEntry> consts = MCP->getConstants();
+  size_t numConsts = consts.size();
+  for (size_t x = 0; x < numConsts; ++x) {
+    krnl.CPOffsets.push_back(
+        std::make_pair<uint32_t, const Constant*>(
+          mCurrentCPOffset, consts[x].Val.ConstVal));
+    size_t curSize = getTypeSize(consts[x].Val.ConstVal->getType(), true);
+    // Align the size to the vector boundary
+    curSize = (curSize + 15) & (~15);
+    mCurrentCPOffset += curSize;
+  }
+}
+
+bool AMDILGlobalManager::isConstPtrArray(const kernel &krnl,
+                                         const llvm::StringRef &arg) {
+  const constPtr *curConst = getConstPtr(krnl, arg);
+  if (curConst) {
+    return curConst->isArray;
+  } else {
+    return false;
+  }
+}
+
+bool AMDILGlobalManager::isConstPtrArgument(const kernel &krnl,
+                                            const llvm::StringRef &arg)
+{
+  const constPtr *curConst = getConstPtr(krnl, arg);
+  if (curConst) {
+    return curConst->isArgument;
+  } else {
+    return false;
+  }
+}
+
+const Value *AMDILGlobalManager::getConstPtrValue(const kernel &krnl,
+                                                  const llvm::StringRef &arg) {
+  const constPtr *curConst = getConstPtr(krnl, arg);
+  if (curConst) {
+    return curConst->base;
+  } else {
+    return NULL;
+  }
+}
+
+static void
+dumpZeroElements(const  StructType * const T, llvm::raw_ostream &O, bool asBytes);
+static void
+dumpZeroElements(const IntegerType * const T, llvm::raw_ostream &O, bool asBytes);
+static void
+dumpZeroElements(const   ArrayType * const T, llvm::raw_ostream &O, bool asBytes);
+static void
+dumpZeroElements(const  VectorType * const T, llvm::raw_ostream &O, bool asBytes);
+static void
+dumpZeroElements(const        Type * const T, llvm::raw_ostream &O, bool asBytes);
+
+void dumpZeroElements(const Type * const T, llvm::raw_ostream &O, bool asBytes) {
+  if (!T) {
+    return;
+  }
+  switch(T->getTypeID()) {
+  case Type::X86_FP80TyID:
+  case Type::FP128TyID:
+  case Type::PPC_FP128TyID:
+  case Type::LabelTyID:
+    assert(0 && "These types are not supported by this backend");
+  default:
+  case Type::DoubleTyID:
+    if (asBytes) {
+      O << ":0:0:0:0:0:0:0:0";
+    } else {
+      O << ":0";
+    }
+    break;
+  case Type::FloatTyID:
+  case Type::PointerTyID:
+  case Type::FunctionTyID:
+    if (asBytes) {
+      O << ":0:0:0:0";
+    } else {
+      O << ":0";
+    }
+    break;
+  case Type::IntegerTyID:
+    dumpZeroElements(dyn_cast<IntegerType>(T), O, asBytes);
+    break;
+  case Type::StructTyID:
+    {
+      const StructType *ST = cast<StructType>(T);
+      if (!ST->isOpaque()) {
+        dumpZeroElements(dyn_cast<StructType>(T), O, asBytes);
+      } else { // A pre-LLVM 3.0 opaque type
+        if (asBytes) {
+          O << ":0:0:0:0";
+        } else {
+          O << ":0";
+        }
+      }
+    }
+    break;
+  case Type::ArrayTyID:
+    dumpZeroElements(dyn_cast<ArrayType>(T), O, asBytes);
+    break;
+  case Type::VectorTyID:
+    dumpZeroElements(dyn_cast<VectorType>(T), O, asBytes);
+    break;
+  };
+}
+
+void
+dumpZeroElements(const StructType * const ST, llvm::raw_ostream &O, bool asBytes) {
+  if (!ST) {
+    return;
+  }
+  Type *curType;
+  StructType::element_iterator eib = ST->element_begin();
+  StructType::element_iterator eie = ST->element_end();
+  for (;eib != eie; ++eib) {
+    curType = *eib;
+    dumpZeroElements(curType, O, asBytes);
+  }
+}
+
+void
+dumpZeroElements(const IntegerType * const IT, llvm::raw_ostream &O, bool asBytes) {
+  if (asBytes) {
+    unsigned byteWidth = (IT->getBitWidth() >> 3);
+    for (unsigned x = 0; x < byteWidth; ++x) {
+      O << ":0";
+    }
+  }
+}
+
+void
+dumpZeroElements(const ArrayType * const AT, llvm::raw_ostream &O, bool asBytes) {
+  size_t size = AT->getNumElements();
+  for (size_t x = 0; x < size; ++x) {
+    dumpZeroElements(AT->getElementType(), O, asBytes);
+  }
+}
+
+void
+dumpZeroElements(const VectorType * const VT, llvm::raw_ostream &O, bool asBytes) {
+  size_t size = VT->getNumElements();
+  for (size_t x = 0; x < size; ++x) {
+    dumpZeroElements(VT->getElementType(), O, asBytes);
+  }
+}
+
+void AMDILGlobalManager::printConstantValue(const Constant *CAval,
+                                            llvm::raw_ostream &O, bool asBytes) {
+  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CAval)) {
+    bool isDouble = &CFP->getValueAPF().getSemantics()==&APFloat::IEEEdouble;
+    if (isDouble) {
+      double val = CFP->getValueAPF().convertToDouble();
+      union dtol_union {
+        double d;
+        uint64_t l;
+        char c[8];
+      } conv;
+      conv.d = val;
+      if (!asBytes) {
+        O << ":";
+        O.write_hex(conv.l);
+      } else {
+        for (int i = 0; i < 8; ++i) {
+          O << ":";
+          O.write_hex((unsigned)conv.c[i] & 0xFF);
+        }
+      }
+    } else {
+      float val = CFP->getValueAPF().convertToFloat();
+      union ftoi_union {
+        float f;
+        uint32_t u;
+        char c[4];
+      } conv;
+      conv.f = val;
+      if (!asBytes) {
+        O << ":";
+        O.write_hex(conv.u);
+      } else {
+        for (int i = 0; i < 4; ++i) {
+          O << ":";
+          O.write_hex((unsigned)conv.c[i] & 0xFF);
+        }
+      }
+    }
+  } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(CAval)) {
+    uint64_t zVal = CI->getValue().getZExtValue();
+    if (!asBytes) {
+      O << ":";
+      O.write_hex(zVal);
+    } else {
+      switch (CI->getBitWidth()) {
+      default:
+        {
+          union ltob_union {
+            uint64_t l;
+            char c[8];
+          } conv;
+          conv.l = zVal;
+          for (int i = 0; i < 8; ++i) {
+            O << ":";
+            O.write_hex((unsigned)conv.c[i] & 0xFF);
+          }
+        }
+        break;
+      case 8:
+        O << ":";
+        O.write_hex(zVal & 0xFF);
+        break;
+      case 16:
+        {
+          union stob_union {
+            uint16_t s;
+            char c[2];
+          } conv;
+          conv.s = (uint16_t)zVal;
+          O << ":";
+          O.write_hex((unsigned)conv.c[0] & 0xFF);
+          O << ":";
+          O.write_hex((unsigned)conv.c[1] & 0xFF);
+        }
+        break;
+      case 32:
+        {
+          union itob_union {
+            uint32_t i;
+            char c[4];
+          } conv;
+          conv.i = (uint32_t)zVal;
+          for (int i = 0; i < 4; ++i) {
+            O << ":";
+            O.write_hex((unsigned)conv.c[i] & 0xFF);
+          }
+        }
+        break;
+      }
+    }
+  } else if (const ConstantVector *CV = dyn_cast<ConstantVector>(CAval)) {
+    int y = CV->getNumOperands()-1;
+    int x = 0;
+    for (; x < y; ++x) {
+      printConstantValue(CV->getOperand(x), O, asBytes);
+    }
+    printConstantValue(CV->getOperand(x), O, asBytes);
+  } else if (const ConstantStruct *CS = dyn_cast<ConstantStruct>(CAval)) {
+    int y = CS->getNumOperands();
+    int x = 0;
+    for (; x < y; ++x) {
+      printConstantValue(CS->getOperand(x), O, asBytes);
+    }
+  } else if (const ConstantAggregateZero *CAZ
+      = dyn_cast<ConstantAggregateZero>(CAval)) {
+    int y = CAZ->getNumOperands();
+    if (y > 0) {
+      int x = 0;
+      for (; x < y; ++x) {
+        printConstantValue((llvm::Constant *)CAZ->getOperand(x),
+            O, asBytes);
+      }
+    } else {
+      if (asBytes) {
+        dumpZeroElements(CAval->getType(), O, asBytes);
+      } else {
+        int y = getNumElements(CAval->getType())-1;
+        for (int x = 0; x < y; ++x) {
+          O << ":0";
+        }
+        O << ":0";
+      }
+    }
+  } else if (const ConstantArray *CA = dyn_cast<ConstantArray>(CAval)) {
+    int y = CA->getNumOperands();
+    int x = 0;
+    for (; x < y; ++x) {
+      printConstantValue(CA->getOperand(x), O, asBytes);
+    }
+  } else if (dyn_cast<ConstantPointerNull>(CAval)) {
+    O << ":0";
+    //assert(0 && "Hit condition which was not expected");
+  } else if (dyn_cast<ConstantExpr>(CAval)) {
+    O << ":0";
+    //assert(0 && "Hit condition which was not expected");
+  } else if (dyn_cast<UndefValue>(CAval)) {
+    O << ":0";
+    //assert(0 && "Hit condition which was not expected");
+  } else {
+    assert(0 && "Hit condition which was not expected");
+  }
+}
+
+static bool isStruct(Type * const T)
+{
+  if (!T) {
+    return false;
+  }
+  switch (T->getTypeID()) {
+  default:
+    return false;
+  case Type::PointerTyID:
+    return isStruct(T->getContainedType(0));
+  case Type::StructTyID:
+    return true;
+  case Type::ArrayTyID:
+  case Type::VectorTyID:
+    return isStruct(dyn_cast<SequentialType>(T)->getElementType());
+  };
+
+}
+
+void AMDILGlobalManager::dumpDataToCB(llvm::raw_ostream &O, AMDILKernelManager *km,
+                                      uint32_t id) {
+  uint32_t size = 0;
+  for (StringMap<constPtr>::iterator cmb = consts_begin(),
+      cme = consts_end(); cmb != cme; ++cmb) {
+    if (id == cmb->second.cbNum) {
+      size += (cmb->second.size + 15) & (~15);
+    }
+  }
+  if (id == 0) {
+    O << ";#DATASTART:" << (size + mCurrentCPOffset) << "\n";
+    if (mCurrentCPOffset) {
+      for (StringMap<kernel>::iterator kcpb = mKernels.begin(),
+          kcpe = mKernels.end(); kcpb != kcpe; ++kcpb) {
+        const kernel& k = kcpb->second;
+        size_t numConsts = k.CPOffsets.size();
+        for (size_t x = 0; x < numConsts; ++x) {
+          size_t offset = k.CPOffsets[x].first;
+          const Constant *C = k.CPOffsets[x].second;
+          Type *Ty = C->getType();
+          size_t size = (isStruct(Ty) ? getTypeSize(Ty, true)
+                                      : getNumElements(Ty));
+          O << ";#" << km->getTypeName(Ty, symTab) << ":";
+          O << offset << ":" << size ;
+          printConstantValue(C, O, isStruct(Ty));
+          O << "\n";
+        }
+      }
+    }
+  } else {
+    O << ";#DATASTART:" << id << ":" << size << "\n";
+  }
+
+  for (StringMap<constPtr>::iterator cmb = consts_begin(), cme = consts_end();
+       cmb != cme; ++cmb) {
+    if (cmb->second.cbNum != id) {
+      continue;
+    }
+    const GlobalVariable *G = dyn_cast<GlobalVariable>(cmb->second.base);
+    Type *Ty = (G) ? G->getType() : NULL;
+    size_t offset = cmb->second.offset;
+    const Constant *C = G->getInitializer();
+    size_t size = (isStruct(Ty)
+        ? getTypeSize(Ty, true)
+        : getNumElements(Ty));
+    O << ";#" << km->getTypeName(Ty, symTab) << ":";
+    if (!id) {
+      O << (offset + mCurrentCPOffset) << ":" << size;
+    } else {
+      O << offset << ":" << size;
+    }
+    if (C) {
+      printConstantValue(C, O, isStruct(Ty));
+    } else {
+      assert(0 && "Cannot have a constant pointer"
+          " without an initializer!");
+    }
+    O <<"\n";
+  }
+  if (id == 0) {
+    O << ";#DATAEND\n";
+  } else {
+    O << ";#DATAEND:" << id << "\n";
+  }
+}
+
+void
+AMDILGlobalManager::dumpDataSection(llvm::raw_ostream &O, AMDILKernelManager *km) {
+  if (mConstMems.empty() && !mCurrentCPOffset) {
+    return;
+  } else {
+    llvm::DenseSet<uint32_t> const_set;
+    for (StringMap<constPtr>::iterator cmb = consts_begin(), cme = consts_end();
+         cmb != cme; ++cmb) {
+      const_set.insert(cmb->second.cbNum);
+    }
+    if (mCurrentCPOffset) {
+      const_set.insert(0);
+    }
+    for (llvm::DenseSet<uint32_t>::iterator setb = const_set.begin(),
+           sete = const_set.end(); setb != sete; ++setb) {
+      dumpDataToCB(O, km, *setb);
+    }
+  }
+}
+
+/// Create a function ID if it is not known or return the known
+/// function ID.
+uint32_t AMDILGlobalManager::getOrCreateFunctionID(const GlobalValue* func) {
+  if (func->getName().size()) {
+    return getOrCreateFunctionID(func->getName());
+  } 
+  uint32_t id;
+  if (mFuncPtrNames.find(func) == mFuncPtrNames.end()) {
+    id = mFuncPtrNames.size() + RESERVED_FUNCS + mFuncNames.size();
+    mFuncPtrNames[func] = id;
+  } else {
+    id = mFuncPtrNames[func];
+  }
+  return id;
+}
+uint32_t AMDILGlobalManager::getOrCreateFunctionID(const std::string &func) {
+  uint32_t id;
+  if (mFuncNames.find(func) == mFuncNames.end()) {
+    id = mFuncNames.size() + RESERVED_FUNCS + mFuncPtrNames.size();
+    mFuncNames[func] = id;
+  } else {
+    id = mFuncNames[func];
+  }
+  return id;
+}