Merge "Metadata support for general reduction."
diff --git a/bcinfo/MetadataExtractor.cpp b/bcinfo/MetadataExtractor.cpp
index add1ab1..f190c48 100644
--- a/bcinfo/MetadataExtractor.cpp
+++ b/bcinfo/MetadataExtractor.cpp
@@ -64,6 +64,13 @@
return c;
}
+const char *createStringFromOptionalValue(llvm::MDNode *n, unsigned opndNum) {
+ llvm::Metadata *opnd;
+ if (opndNum >= n->getNumOperands() || !(opnd = n->getOperand(opndNum)))
+ return nullptr;
+ return createStringFromValue(opnd);
+}
+
// Collect metadata from NamedMDNodes that contain a list of names
// (strings).
//
@@ -146,10 +153,14 @@
// (should be synced with slang_rs_metadata.h)
static const llvm::StringRef ExportForEachMetadataName = "#rs_export_foreach";
-// Name of metadata node where exported reduce name information resides
+// Name of metadata node where exported simple reduce name information resides
// (should be synced with slang_rs_metadata.h)
static const llvm::StringRef ExportReduceMetadataName = "#rs_export_reduce";
+// Name of metadata node where exported general reduce information resides
+// (should be synced with slang_rs_metadata.h)
+static const llvm::StringRef ExportReduceNewMetadataName = "#rs_export_reduce_new";
+
// Name of metadata node where RS object slot info resides (should be
// synced with slang_rs_metadata.h)
static const llvm::StringRef ObjectSlotMetadataName = "#rs_object_slots";
@@ -163,10 +174,11 @@
MetadataExtractor::MetadataExtractor(const char *bitcode, size_t bitcodeSize)
: mModule(nullptr), mBitcode(bitcode), mBitcodeSize(bitcodeSize),
mExportVarCount(0), mExportFuncCount(0), mExportForEachSignatureCount(0),
- mExportReduceCount(0), mExportVarNameList(nullptr),
+ mExportReduceCount(0), mExportReduceNewCount(0), mExportVarNameList(nullptr),
mExportFuncNameList(nullptr), mExportForEachNameList(nullptr),
mExportForEachSignatureList(nullptr),
mExportForEachInputCountList(nullptr), mExportReduceNameList(nullptr),
+ mExportReduceNewList(nullptr),
mPragmaCount(0), mPragmaKeyList(nullptr), mPragmaValueList(nullptr),
mObjectSlotCount(0), mObjectSlotList(nullptr),
mRSFloatPrecision(RS_FP_Full), mIsThreadable(true),
@@ -180,10 +192,11 @@
MetadataExtractor::MetadataExtractor(const llvm::Module *module)
: mModule(module), mBitcode(nullptr), mBitcodeSize(0),
mExportVarCount(0), mExportFuncCount(0), mExportForEachSignatureCount(0),
- mExportReduceCount(0), mExportVarNameList(nullptr),
+ mExportReduceCount(0), mExportReduceNewCount(0), mExportVarNameList(nullptr),
mExportFuncNameList(nullptr), mExportForEachNameList(nullptr),
mExportForEachSignatureList(nullptr),
mExportForEachInputCountList(nullptr), mExportReduceNameList(nullptr),
+ mExportReduceNewList(nullptr),
mPragmaCount(0), mPragmaKeyList(nullptr), mPragmaValueList(nullptr),
mObjectSlotCount(0), mObjectSlotList(nullptr),
mRSFloatPrecision(RS_FP_Full), mIsThreadable(true),
@@ -224,6 +237,9 @@
delete [] mExportForEachSignatureList;
mExportForEachSignatureList = nullptr;
+ delete [] mExportForEachInputCountList;
+ mExportForEachInputCountList = nullptr;
+
if (mExportReduceNameList) {
for (size_t i = 0; i < mExportReduceCount; i++) {
delete [] mExportReduceNameList[i];
@@ -233,6 +249,9 @@
delete [] mExportReduceNameList;
mExportReduceNameList = nullptr;
+ delete [] mExportReduceNewList;
+ mExportReduceNewList = nullptr;
+
for (size_t i = 0; i < mPragmaCount; i++) {
if (mPragmaKeyList) {
delete [] mPragmaKeyList[i];
@@ -468,6 +487,61 @@
}
+bool MetadataExtractor::populateReduceNewMetadata(const llvm::NamedMDNode *ReduceNewMetadata) {
+ mExportReduceNewCount = 0;
+ mExportReduceNewList = nullptr;
+
+ if (!ReduceNewMetadata || !(mExportReduceNewCount = ReduceNewMetadata->getNumOperands()))
+ return true;
+
+ ReduceNew *TmpReduceNewList = new ReduceNew[mExportReduceNewCount];
+
+ for (size_t i = 0; i < mExportReduceNewCount; i++) {
+ llvm::MDNode *Node = ReduceNewMetadata->getOperand(i);
+ if (!Node || Node->getNumOperands() < 3) {
+ ALOGE("Missing reduce metadata");
+ return false;
+ }
+
+ TmpReduceNewList[i].mReduceName = createStringFromValue(Node->getOperand(0));
+
+ if (!extractUIntFromMetadataString(&TmpReduceNewList[i].mAccumulatorDataSize,
+ Node->getOperand(1))) {
+ ALOGE("Non-integer accumulator data size value in reduce metadata");
+ return false;
+ }
+
+ llvm::MDNode *AccumulatorNode = llvm::dyn_cast<llvm::MDNode>(Node->getOperand(2));
+ if (!AccumulatorNode || AccumulatorNode->getNumOperands() != 2) {
+ ALOGE("Malformed accumulator node in reduce metadata");
+ return false;
+ }
+ TmpReduceNewList[i].mAccumulatorName = createStringFromValue(AccumulatorNode->getOperand(0));
+ if (!extractUIntFromMetadataString(&TmpReduceNewList[i].mSignature,
+ AccumulatorNode->getOperand(1))) {
+ ALOGE("Non-integer signature value in reduce metadata");
+ return false;
+ }
+ llvm::Function *Func =
+ mModule->getFunction(llvm::StringRef(TmpReduceNewList[i].mAccumulatorName));
+ if (!Func) {
+ ALOGE("reduce metadata names missing accumulator function");
+ return false;
+ }
+ // Why calculateNumInputs() - 1? The "-1" is because we don't
+ // want to treat the accumulator argument as an input.
+ TmpReduceNewList[i].mInputCount = calculateNumInputs(Func, TmpReduceNewList[i].mSignature) - 1;
+
+ TmpReduceNewList[i].mInitializerName = createStringFromOptionalValue(Node, 3);
+ TmpReduceNewList[i].mCombinerName = createStringFromOptionalValue(Node, 4);
+ TmpReduceNewList[i].mOutConverterName = createStringFromOptionalValue(Node, 5);
+ TmpReduceNewList[i].mHalterName = createStringFromOptionalValue(Node, 6);
+ }
+
+ mExportReduceNewList = TmpReduceNewList;
+ return true;
+}
+
void MetadataExtractor::readThreadableFlag(
const llvm::NamedMDNode *ThreadableMetadata) {
@@ -544,6 +618,8 @@
mModule->getNamedMetadata(ExportForEachMetadataName);
const llvm::NamedMDNode *ExportReduceMetadata =
mModule->getNamedMetadata(ExportReduceMetadataName);
+ const llvm::NamedMDNode *ExportReduceNewMetadata =
+ mModule->getNamedMetadata(ExportReduceNewMetadataName);
const llvm::NamedMDNode *PragmaMetadata =
mModule->getNamedMetadata(PragmaMetadataName);
const llvm::NamedMDNode *ObjectSlotMetadata =
@@ -577,6 +653,11 @@
return false;
}
+ if (!populateReduceNewMetadata(ExportReduceNewMetadata)) {
+ ALOGE("Could not populate export general reduction metadata");
+ return false;
+ }
+
populatePragmaMetadata(PragmaMetadata);
if (!populateObjectSlotMetadata(ObjectSlotMetadata)) {
diff --git a/bcinfo/tools/main.cpp b/bcinfo/tools/main.cpp
index c4e40f4..855199d 100644
--- a/bcinfo/tools/main.cpp
+++ b/bcinfo/tools/main.cpp
@@ -110,6 +110,11 @@
}
+static void dumpReduceNewInfo(FILE *info, const char *Kind, const char *Name) {
+ if (Name)
+ fprintf(info, " %s(%s)\n", Kind, Name);
+}
+
static int dumpInfo(bcinfo::MetadataExtractor *ME) {
if (!ME) {
return 1;
@@ -149,6 +154,20 @@
fprintf(info, "%s\n", reduceNameList[i]);
}
+ fprintf(info, "exportReduceNewCount: %zu\n", ME->getExportReduceNewCount());
+ const bcinfo::MetadataExtractor::ReduceNew *reduceNewList =
+ ME->getExportReduceNewList();
+ for (size_t i = 0; i < ME->getExportReduceNewCount(); i++) {
+ const bcinfo::MetadataExtractor::ReduceNew &reduceNew = reduceNewList[i];
+ fprintf(info, "%u - %s - %u - %u\n", reduceNew.mSignature, reduceNew.mReduceName,
+ reduceNew.mInputCount, reduceNew.mAccumulatorDataSize);
+ dumpReduceNewInfo(info, "initializer", reduceNew.mInitializerName);
+ dumpReduceNewInfo(info, "accumulator", reduceNew.mAccumulatorName);
+ dumpReduceNewInfo(info, "combiner", reduceNew.mCombinerName);
+ dumpReduceNewInfo(info, "outconverter", reduceNew.mOutConverterName);
+ dumpReduceNewInfo(info, "halter", reduceNew.mHalterName);
+ }
+
fprintf(info, "objectSlotCount: %zu\n", ME->getObjectSlotCount());
const uint32_t *slotList = ME->getObjectSlotList();
for (size_t i = 0; i < ME->getObjectSlotCount(); i++) {
@@ -210,6 +229,20 @@
}
printf("\n");
+ printf("exportReduceNewCount: %zu\n", ME->getExportReduceNewCount());
+ const bcinfo::MetadataExtractor::ReduceNew *reduceNewList = ME->getExportReduceNewList();
+ for (size_t i = 0; i < ME->getExportReduceNewCount(); i++) {
+ const bcinfo::MetadataExtractor::ReduceNew &reduceNew = reduceNewList[i];
+ printf("exportReduceNewList[%zu]: %s - 0x%08x - %u - %u\n", i, reduceNew.mReduceName,
+ reduceNew.mSignature, reduceNew.mInputCount, reduceNew.mAccumulatorDataSize);
+ dumpReduceNewInfo(stdout, "initializer", reduceNew.mInitializerName);
+ dumpReduceNewInfo(stdout, "accumulator", reduceNew.mAccumulatorName);
+ dumpReduceNewInfo(stdout, "combiner", reduceNew.mCombinerName);
+ dumpReduceNewInfo(stdout, "outconverter", reduceNew.mOutConverterName);
+ dumpReduceNewInfo(stdout, "halter", reduceNew.mHalterName);
+ }
+ printf("\n");
+
printf("pragmaCount: %zu\n", ME->getPragmaCount());
const char **keyList = ME->getPragmaKeyList();
const char **valueList = ME->getPragmaValueList();
diff --git a/include/bcinfo/MetadataExtractor.h b/include/bcinfo/MetadataExtractor.h
index 742346a..738c1c2 100644
--- a/include/bcinfo/MetadataExtractor.h
+++ b/include/bcinfo/MetadataExtractor.h
@@ -46,6 +46,40 @@
};
class MetadataExtractor {
+ public:
+ struct ReduceNew {
+ // These strings are owned by the ReduceNew instance, and deleted upon its destruction.
+ // They are assumed to have been allocated by "new []" and hence are deleted by "delete []".
+ const char *mReduceName;
+ const char *mInitializerName;
+ const char *mAccumulatorName;
+ const char *mCombinerName;
+ const char *mOutConverterName;
+ const char *mHalterName;
+
+ uint32_t mSignature; // of accumulator function
+ uint32_t mInputCount; // of accumulator function (and of kernel itself)
+ uint32_t mAccumulatorDataSize; // in bytes
+
+ ReduceNew() :
+ mReduceName(nullptr),
+ mInitializerName(nullptr), mAccumulatorName(nullptr), mCombinerName(nullptr),
+ mOutConverterName(nullptr), mHalterName(nullptr),
+ mSignature(0), mInputCount(0), mAccumulatorDataSize(0) {
+ }
+ ~ReduceNew() {
+ delete [] mReduceName;
+ delete [] mInitializerName;
+ delete [] mAccumulatorName;
+ delete [] mCombinerName;
+ delete [] mOutConverterName;
+ delete [] mHalterName;
+ }
+
+ ReduceNew(const ReduceNew &) = delete;
+ void operator=(const ReduceNew &) = delete;
+ };
+
private:
const llvm::Module *mModule;
const char *mBitcode;
@@ -55,12 +89,14 @@
size_t mExportFuncCount;
size_t mExportForEachSignatureCount;
size_t mExportReduceCount;
+ size_t mExportReduceNewCount;
const char **mExportVarNameList;
const char **mExportFuncNameList;
const char **mExportForEachNameList;
const uint32_t *mExportForEachSignatureList;
const uint32_t *mExportForEachInputCountList;
const char **mExportReduceNameList;
+ const ReduceNew *mExportReduceNewList;
size_t mPragmaCount;
const char **mPragmaKeyList;
@@ -83,6 +119,7 @@
// Helper functions for extraction
bool populateForEachMetadata(const llvm::NamedMDNode *Names,
const llvm::NamedMDNode *Signatures);
+ bool populateReduceNewMetadata(const llvm::NamedMDNode *ReduceNewMetadata);
bool populateObjectSlotMetadata(const llvm::NamedMDNode *ObjectSlotMetadata);
void populatePragmaMetadata(const llvm::NamedMDNode *PragmaMetadata);
void readThreadableFlag(const llvm::NamedMDNode *ThreadableMetadata);
@@ -183,20 +220,34 @@
}
/**
- * \return number of exported reduce kernels (slots) in this script/module.
+ * \return number of exported simple reduce kernels (slots) in this script/module.
*/
size_t getExportReduceCount() const {
return mExportReduceCount;
}
/**
- * \return array of exported reduce kernel names.
+ * \return array of exported simple reduce kernel names.
*/
const char **getExportReduceNameList() const {
return mExportReduceNameList;
}
/**
+ * \return number of exported general reduce kernels (slots) in this script/module.
+ */
+ size_t getExportReduceNewCount() const {
+ return mExportReduceNewCount;
+ }
+
+ /**
+ * \return array of exported general reduce kernel descriptions.
+ */
+ const ReduceNew *getExportReduceNewList() const {
+ return mExportReduceNewList;
+ }
+
+ /**
* \return number of pragmas contained in pragmaKeyList and pragmaValueList.
*/
size_t getPragmaCount() const {
diff --git a/lib/Core/Compiler.cpp b/lib/Core/Compiler.cpp
index 5c769b4..71c0ed2 100644
--- a/lib/Core/Compiler.cpp
+++ b/lib/Core/Compiler.cpp
@@ -330,10 +330,12 @@
size_t exportFuncCount = me.getExportFuncCount();
size_t exportForEachCount = me.getExportForEachSignatureCount();
size_t exportReduceCount = me.getExportReduceCount();
+ size_t exportReduceNewCount = me.getExportReduceNewCount();
const char **exportVarNameList = me.getExportVarNameList();
const char **exportFuncNameList = me.getExportFuncNameList();
const char **exportForEachNameList = me.getExportForEachNameList();
const char **exportReduceNameList = me.getExportReduceNameList();
+ const bcinfo::MetadataExtractor::ReduceNew *exportReduceNewList = me.getExportReduceNewList();
size_t i;
for (i = 0; i < exportVarCount; ++i) {
@@ -349,7 +351,7 @@
// functions around until createInternalizePass() is finished making
// its own copy of the visible symbols.
std::vector<std::string> expanded_funcs;
- expanded_funcs.reserve(exportForEachCount + exportReduceCount);
+ expanded_funcs.reserve(exportForEachCount + exportReduceCount + exportReduceNewCount);
for (i = 0; i < exportForEachCount; ++i) {
expanded_funcs.push_back(std::string(exportForEachNameList[i]) + ".expand");
@@ -357,6 +359,9 @@
for (i = 0; i < exportReduceCount; ++i) {
expanded_funcs.push_back(std::string(exportReduceNameList[i]) + ".expand");
}
+ for (i = 0; i < exportReduceNewCount; ++i) {
+ expanded_funcs.push_back(std::string(exportReduceNewList[i].mAccumulatorName) + ".expand");
+ }
for (auto &symbol_name : expanded_funcs) {
export_symbols.push_back(symbol_name.c_str());
diff --git a/lib/Renderscript/RSEmbedInfo.cpp b/lib/Renderscript/RSEmbedInfo.cpp
index b0c2767..54e0acb 100644
--- a/lib/Renderscript/RSEmbedInfo.cpp
+++ b/lib/Renderscript/RSEmbedInfo.cpp
@@ -74,6 +74,7 @@
size_t exportFuncCount = me.getExportFuncCount();
size_t exportForEachCount = me.getExportForEachSignatureCount();
size_t exportReduceCount = me.getExportReduceCount();
+ size_t exportReduceNewCount = me.getExportReduceNewCount();
size_t objectSlotCount = me.getObjectSlotCount();
size_t pragmaCount = me.getPragmaCount();
const char **exportVarNameList = me.getExportVarNameList();
@@ -82,6 +83,8 @@
const char **exportReduceNameList = me.getExportReduceNameList();
const uint32_t *exportForEachSignatureList =
me.getExportForEachSignatureList();
+ const bcinfo::MetadataExtractor::ReduceNew *exportReduceNewList =
+ me.getExportReduceNewList();
const uint32_t *objectSlotList = me.getObjectSlotList();
const char **pragmaKeyList = me.getPragmaKeyList();
const char **pragmaValueList = me.getPragmaValueList();
@@ -90,13 +93,22 @@
size_t i;
- // We use a simple text format here that the compatibility library can
- // easily parse. Each section starts out with its name followed by a count.
- // The count denotes the number of lines to parse for that particular
- // category. Variables and Functions merely put the appropriate identifier
- // on the line, while ForEach kernels have the encoded int signature,
- // followed by a hyphen followed by the identifier (function to look up).
- // Object Slots are just listed as one integer per line.
+ // We use a simple text format here that the compatibility library
+ // can easily parse. Each section starts out with its name
+ // followed by a count. The count denotes the number of lines to
+ // parse for that particular category. Variables and Functions and
+ // simple reduce kernels merely put the appropriate identifier on
+ // the line. ForEach kernels have the encoded int signature,
+ // followed by a hyphen followed by the identifier (function to
+ // look up). General reduce kernels have the encoded int
+ // signature, followed by a hyphen followed by the accumulator
+ // data size, followed by a hyphen followed by the identifier
+ // (reduction name); and then for each possible constituent
+ // function, a hyphen followed by the identifier (function name)
+ // -- in the case where the function is omitted, "." is used in
+ // place of the identifier. Object Slots are just listed as one
+ // integer per line.
+
s << "exportVarCount: " << exportVarCount << "\n";
for (i = 0; i < exportVarCount; ++i) {
s << exportVarNameList[i] << "\n";
@@ -118,6 +130,21 @@
s << exportReduceNameList[i] << "\n";
}
+ s << "exportReduceNewCount: " << exportReduceNewCount << "\n";
+ auto reduceNewFnName = [](const char *Name) { return Name ? Name : "."; };
+ for (i = 0; i < exportReduceNewCount; ++i) {
+ const bcinfo::MetadataExtractor::ReduceNew &reduceNew = exportReduceNewList[i];
+ s << reduceNew.mSignature << " - "
+ << reduceNew.mAccumulatorDataSize << " - "
+ << reduceNew.mReduceName << " - "
+ << reduceNewFnName(reduceNew.mInitializerName) << " - "
+ << reduceNewFnName(reduceNew.mAccumulatorName) << " - "
+ << reduceNewFnName(reduceNew.mCombinerName) << " - "
+ << reduceNewFnName(reduceNew.mOutConverterName) << " - "
+ << reduceNewFnName(reduceNew.mHalterName)
+ << "\n";
+ }
+
s << "objectSlotCount: " << objectSlotCount << "\n";
for (i = 0; i < objectSlotCount; ++i) {
s << objectSlotList[i] << "\n";
diff --git a/tests/libbcc/test_reduce_general_metadata.ll b/tests/libbcc/test_reduce_general_metadata.ll
new file mode 100644
index 0000000..d23854a
--- /dev/null
+++ b/tests/libbcc/test_reduce_general_metadata.ll
@@ -0,0 +1,333 @@
+; Check that the #rs_export_reduce_new node is recognized.
+
+; RUN: llvm-rs-as %s -o %t
+; RUN: bcinfo %t | FileCheck %s
+
+; CHECK: exportReduceNewCount: 8
+; CHECK: exportReduceNewList[0]: addint - 0x00000001 - 1 - 4
+; CHECK: accumulator(aiAccum)
+; CHECK: exportReduceNewList[1]: mpyint - 0x00000001 - 1 - 4
+; CHECK: initializer(mpyInit)
+; CHECK: accumulator(mpyAccum)
+; CHECK: exportReduceNewList[2]: dp - 0x00000001 - 2 - 4
+; CHECK: accumulator(dpAccum)
+; CHECK: combiner(dpSum)
+; CHECK: exportReduceNewList[3]: findMinAndMax - 0x00000009 - 1 - 16
+; CHECK: initializer(fMMInit)
+; CHECK: accumulator(fMMAccumulator)
+; CHECK: combiner(fMMCombiner)
+; CHECK: outconverter(fMMOutConverter)
+; CHECK: exportReduceNewList[4]: fz - 0x00000009 - 1 - 4
+; CHECK: initializer(fzInit)
+; CHECK: accumulator(fzAccum)
+; CHECK: combiner(fzCombine)
+; CHECK: halter(fzFound)
+; CHECK: exportReduceNewList[5]: fz2 - 0x00000019 - 1 - 8
+; CHECK: initializer(fz2Init)
+; CHECK: accumulator(fz2Accum)
+; CHECK: combiner(fz2Combine)
+; CHECK: halter(fz2Found)
+; CHECK: exportReduceNewList[6]: histogram - 0x00000001 - 1 - 1024
+; CHECK: accumulator(hsgAccum)
+; CHECK: combiner(hsgCombine)
+; CHECK: exportReduceNewList[7]: mode - 0x00000001 - 1 - 1024
+; CHECK: accumulator(hsgAccum)
+; CHECK: combiner(hsgCombine)
+; CHECK: outconverter(modeOutConvert)
+
+; ModuleID = 'reduce_general_examples.bc'
+target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-none-linux-gnueabi"
+
+%struct.MinAndMax.0 = type { %struct.IndexedVal.1, %struct.IndexedVal.1 }
+%struct.IndexedVal.1 = type { float, i32 }
+
+@fMMInit.r = internal unnamed_addr constant %struct.MinAndMax.0 { %struct.IndexedVal.1 { float 0.000000e+00, i32 -1 }, %struct.IndexedVal.1 { float -0.000000e+00, i32 -1 } }, align 4
+@llvm.used = appending global [20 x i8*] [i8* bitcast (void (<2 x i32>*)* @fz2Init to i8*), i8* bitcast (void ([256 x i32]*, [256 x i32]*)* @hsgCombine to i8*), i8* bitcast (i1 (<2 x i32>*)* @fz2Found to i8*), i8* bitcast (void (i32*, i32)* @mpyAccum to i8*), i8* bitcast (void (%struct.MinAndMax.0*)* @fMMInit to i8*), i8* bitcast (void (float*, float, float)* @dpAccum to i8*), i8* bitcast (void (<2 x i32>*, [256 x i32]*)* @modeOutConvert to i8*), i8* bitcast (void ([256 x i32]*, i8)* @hsgAccum to i8*), i8* bitcast (void (i32*)* @mpyInit to i8*), i8* bitcast (void (%struct.MinAndMax.0*, float, i32)* @fMMAccumulator to i8*), i8* bitcast (void (float*, float*)* @dpSum to i8*), i8* bitcast (void (%struct.MinAndMax.0*, %struct.MinAndMax.0*)* @fMMCombiner to i8*), i8* bitcast (void (i32*, i32*)* @fzCombine to i8*), i8* bitcast (void (i32*, i32)* @aiAccum to i8*), i8* bitcast (void (i32*)* @fzInit to i8*), i8* bitcast (void (i32*, i32, i32)* @fzAccum to i8*), i8* bitcast (i1 (i32*)* @fzFound to i8*), i8* bitcast (void (<2 x i32>*, i32, i32, i32)* @fz2Accum to i8*), i8* bitcast (void (<2 x i32>*, %struct.MinAndMax.0*)* @fMMOutConverter to i8*), i8* bitcast (void (<2 x i32>*, <2 x i32>*)* @fz2Combine to i8*)], section "llvm.metadata"
+
+; Function Attrs: nounwind
+define internal void @aiAccum(i32* nocapture %accum, i32 %val) #0 {
+ %1 = load i32, i32* %accum, align 4, !tbaa !18
+ %2 = add nsw i32 %1, %val
+ store i32 %2, i32* %accum, align 4, !tbaa !18
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @mpyInit(i32* nocapture %accum) #0 {
+ store i32 1, i32* %accum, align 4, !tbaa !18
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @mpyAccum(i32* nocapture %accum, i32 %val) #0 {
+ %1 = load i32, i32* %accum, align 4, !tbaa !18
+ %2 = mul nsw i32 %1, %val
+ store i32 %2, i32* %accum, align 4, !tbaa !18
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @dpAccum(float* nocapture %accum, float %in1, float %in2) #0 {
+ %1 = fmul float %in1, %in2
+ %2 = load float, float* %accum, align 4, !tbaa !22
+ %3 = fadd float %1, %2
+ store float %3, float* %accum, align 4, !tbaa !22
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @dpSum(float* nocapture %accum, float* nocapture %val) #0 {
+ %1 = load float, float* %val, align 4, !tbaa !22
+ %2 = load float, float* %accum, align 4, !tbaa !22
+ %3 = fadd float %1, %2
+ store float %3, float* %accum, align 4, !tbaa !22
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @fMMInit(%struct.MinAndMax.0* nocapture %accum) #0 {
+ %1 = bitcast %struct.MinAndMax.0* %accum to i8*
+ tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %1, i8* bitcast (%struct.MinAndMax.0* @fMMInit.r to i8*), i64 16, i32 4, i1 false), !tbaa.struct !24
+ ret void
+}
+
+; Function Attrs: nounwind
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i32, i1) #0
+
+; Function Attrs: nounwind
+define internal void @fMMAccumulator(%struct.MinAndMax.0* nocapture %accum, float %in, i32 %x) #0 {
+ %1 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %accum, i64 0, i32 0, i32 0
+ %2 = load float, float* %1, align 4, !tbaa !22
+ %3 = fcmp ogt float %2, %in
+ br i1 %3, label %4, label %6
+
+; <label>:4 ; preds = %0
+ store float %in, float* %1, align 4
+ %5 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %accum, i64 0, i32 0, i32 1
+ store i32 %x, i32* %5, align 4
+ br label %6
+
+; <label>:6 ; preds = %4, %0
+ %7 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %accum, i64 0, i32 1, i32 0
+ %8 = load float, float* %7, align 4, !tbaa !22
+ %9 = fcmp olt float %8, %in
+ br i1 %9, label %10, label %12
+
+; <label>:10 ; preds = %6
+ store float %in, float* %7, align 4
+ %11 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %accum, i64 0, i32 1, i32 1
+ store i32 %x, i32* %11, align 4
+ br label %12
+
+; <label>:12 ; preds = %10, %6
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @fMMCombiner(%struct.MinAndMax.0* nocapture %accum, %struct.MinAndMax.0* nocapture %val) #0 {
+ %1 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %val, i64 0, i32 0, i32 0
+ %2 = load float, float* %1, align 4, !tbaa !22
+ %3 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %val, i64 0, i32 0, i32 1
+ %4 = load i32, i32* %3, align 4, !tbaa !18
+ tail call void @fMMAccumulator(%struct.MinAndMax.0* %accum, float %2, i32 %4)
+ %5 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %val, i64 0, i32 1, i32 0
+ %6 = load float, float* %5, align 4, !tbaa !22
+ %7 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %val, i64 0, i32 1, i32 1
+ %8 = load i32, i32* %7, align 4, !tbaa !18
+ tail call void @fMMAccumulator(%struct.MinAndMax.0* %accum, float %6, i32 %8)
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @fMMOutConverter(<2 x i32>* nocapture %result, %struct.MinAndMax.0* nocapture %val) #0 {
+ %1 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %val, i64 0, i32 0, i32 1
+ %2 = load i32, i32* %1, align 4, !tbaa !18
+ %3 = load <2 x i32>, <2 x i32>* %result, align 8
+ %4 = insertelement <2 x i32> %3, i32 %2, i64 0
+ store <2 x i32> %4, <2 x i32>* %result, align 8
+ %5 = getelementptr inbounds %struct.MinAndMax.0, %struct.MinAndMax.0* %val, i64 0, i32 1, i32 1
+ %6 = load i32, i32* %5, align 4, !tbaa !18
+ %7 = insertelement <2 x i32> %4, i32 %6, i64 1
+ store <2 x i32> %7, <2 x i32>* %result, align 8
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @fzInit(i32* nocapture %accumIdx) #0 {
+ store i32 -1, i32* %accumIdx, align 4, !tbaa !18
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @fzAccum(i32* nocapture %accumIdx, i32 %inVal, i32 %x) #0 {
+ %1 = icmp eq i32 %inVal, 0
+ br i1 %1, label %2, label %3
+
+; <label>:2 ; preds = %0
+ store i32 %x, i32* %accumIdx, align 4, !tbaa !18
+ br label %3
+
+; <label>:3 ; preds = %2, %0
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @fzCombine(i32* nocapture %accumIdx, i32* nocapture %accumIdx2) #0 {
+ %1 = load i32, i32* %accumIdx2, align 4, !tbaa !18
+ %2 = icmp sgt i32 %1, -1
+ br i1 %2, label %3, label %4
+
+; <label>:3 ; preds = %0
+ store i32 %1, i32* %accumIdx, align 4, !tbaa !18
+ br label %4
+
+; <label>:4 ; preds = %3, %0
+ ret void
+}
+
+; Function Attrs: nounwind readonly
+define internal i1 @fzFound(i32* nocapture %accumIdx) #1 {
+ %1 = load i32, i32* %accumIdx, align 4, !tbaa !18
+ %2 = icmp sgt i32 %1, -1
+ ret i1 %2
+}
+
+; Function Attrs: nounwind
+define internal void @fz2Init(<2 x i32>* nocapture %accum) #0 {
+ store <2 x i32> <i32 -1, i32 -1>, <2 x i32>* %accum, align 8
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @fz2Accum(<2 x i32>* nocapture %accum, i32 %inVal, i32 %x, i32 %y) #0 {
+ %1 = icmp eq i32 %inVal, 0
+ br i1 %1, label %2, label %5
+
+; <label>:2 ; preds = %0
+ %3 = insertelement <2 x i32> undef, i32 %x, i64 0
+ %4 = insertelement <2 x i32> %3, i32 %y, i64 1
+ store <2 x i32> %4, <2 x i32>* %accum, align 8
+ br label %5
+
+; <label>:5 ; preds = %2, %0
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @fz2Combine(<2 x i32>* nocapture %accum, <2 x i32>* nocapture %accum2) #0 {
+ %1 = load <2 x i32>, <2 x i32>* %accum2, align 8
+ %2 = extractelement <2 x i32> %1, i64 0
+ %3 = icmp sgt i32 %2, -1
+ br i1 %3, label %4, label %5
+
+; <label>:4 ; preds = %0
+ store <2 x i32> %1, <2 x i32>* %accum, align 8, !tbaa !25
+ br label %5
+
+; <label>:5 ; preds = %4, %0
+ ret void
+}
+
+; Function Attrs: nounwind readonly
+define internal i1 @fz2Found(<2 x i32>* nocapture %accum) #1 {
+ %1 = load <2 x i32>, <2 x i32>* %accum, align 8
+ %2 = extractelement <2 x i32> %1, i64 0
+ %3 = icmp sgt i32 %2, -1
+ ret i1 %3
+}
+
+; Function Attrs: nounwind
+define internal void @hsgAccum([256 x i32]* nocapture %h, i8 %in) #0 {
+ %1 = zext i8 %in to i64
+ %2 = getelementptr inbounds [256 x i32], [256 x i32]* %h, i64 0, i64 %1
+ %3 = load i32, i32* %2, align 4, !tbaa !18
+ %4 = add i32 %3, 1
+ store i32 %4, i32* %2, align 4, !tbaa !18
+ ret void
+}
+
+; Function Attrs: nounwind
+define internal void @hsgCombine([256 x i32]* nocapture %accum, [256 x i32]* nocapture %addend) #0 {
+ br label %2
+
+; <label>:1 ; preds = %2
+ ret void
+
+; <label>:2 ; preds = %2, %0
+ %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %2 ]
+ %3 = getelementptr inbounds [256 x i32], [256 x i32]* %addend, i64 0, i64 %indvars.iv
+ %4 = load i32, i32* %3, align 4, !tbaa !18
+ %5 = getelementptr inbounds [256 x i32], [256 x i32]* %accum, i64 0, i64 %indvars.iv
+ %6 = load i32, i32* %5, align 4, !tbaa !18
+ %7 = add i32 %6, %4
+ store i32 %7, i32* %5, align 4, !tbaa !18
+ %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+ %exitcond = icmp eq i64 %indvars.iv.next, 256
+ br i1 %exitcond, label %1, label %2
+}
+
+; Function Attrs: nounwind
+define internal void @modeOutConvert(<2 x i32>* nocapture %result, [256 x i32]* nocapture %h) #0 {
+ br label %8
+
+; <label>:1 ; preds = %8
+ %2 = load <2 x i32>, <2 x i32>* %result, align 8
+ %3 = insertelement <2 x i32> %2, i32 %i.0.mode.0, i64 0
+ store <2 x i32> %3, <2 x i32>* %result, align 8
+ %4 = zext i32 %i.0.mode.0 to i64
+ %5 = getelementptr inbounds [256 x i32], [256 x i32]* %h, i64 0, i64 %4
+ %6 = load i32, i32* %5, align 4, !tbaa !18
+ %7 = insertelement <2 x i32> %3, i32 %6, i64 1
+ store <2 x i32> %7, <2 x i32>* %result, align 8
+ ret void
+
+; <label>:8 ; preds = %8, %0
+ %indvars.iv = phi i64 [ 1, %0 ], [ %indvars.iv.next, %8 ]
+ %mode.01 = phi i32 [ 0, %0 ], [ %i.0.mode.0, %8 ]
+ %9 = getelementptr inbounds [256 x i32], [256 x i32]* %h, i64 0, i64 %indvars.iv
+ %10 = load i32, i32* %9, align 4, !tbaa !18
+ %11 = zext i32 %mode.01 to i64
+ %12 = getelementptr inbounds [256 x i32], [256 x i32]* %h, i64 0, i64 %11
+ %13 = load i32, i32* %12, align 4, !tbaa !18
+ %14 = icmp ugt i32 %10, %13
+ %15 = trunc i64 %indvars.iv to i32
+ %i.0.mode.0 = select i1 %14, i32 %15, i32 %mode.01
+ %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+ %exitcond = icmp eq i64 %indvars.iv.next, 256
+ br i1 %exitcond, label %1, label %8
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readonly }
+
+!llvm.ident = !{!0}
+!\23pragma = !{!1, !2}
+!\23rs_export_reduce_new = !{!3, !5, !7, !9, !11, !13, !15, !17}
+
+!0 = !{!"clang version 3.6 "}
+!1 = !{!"version", !"1"}
+!2 = !{!"java_package_name", !"examples"}
+!3 = !{!"addint", !"4", !4}
+!4 = !{!"aiAccum", !"1"}
+!5 = !{!"mpyint", !"4", !6, !"mpyInit"}
+!6 = !{!"mpyAccum", !"1"}
+!7 = !{!"dp", !"4", !8, null, !"dpSum"}
+!8 = !{!"dpAccum", !"1"}
+!9 = !{!"findMinAndMax", !"16", !10, !"fMMInit", !"fMMCombiner", !"fMMOutConverter"}
+!10 = !{!"fMMAccumulator", !"9"}
+!11 = !{!"fz", !"4", !12, !"fzInit", !"fzCombine", null, !"fzFound"}
+!12 = !{!"fzAccum", !"9"}
+!13 = !{!"fz2", !"8", !14, !"fz2Init", !"fz2Combine", null, !"fz2Found"}
+!14 = !{!"fz2Accum", !"25"}
+!15 = !{!"histogram", !"1024", !16, null, !"hsgCombine"}
+!16 = !{!"hsgAccum", !"1"}
+!17 = !{!"mode", !"1024", !16, null, !"hsgCombine", !"modeOutConvert"}
+!18 = !{!19, !19, i64 0}
+!19 = !{!"int", !20, i64 0}
+!20 = !{!"omnipotent char", !21, i64 0}
+!21 = !{!"Simple C/C++ TBAA"}
+!22 = !{!23, !23, i64 0}
+!23 = !{!"float", !20, i64 0}
+!24 = !{i64 0, i64 4, !22, i64 4, i64 4, !18, i64 8, i64 4, !22, i64 12, i64 4, !18}
+!25 = !{!20, !20, i64 0}