Subzero. Moar performance tweaks.

src/IceCfg.cpp

@@ -15,6 +15,7 @@
 #include "IceCfg.h"
 
 #include "IceAssembler.h"
+#include "IceBitVector.h"
 #include "IceCfgNode.h"
 #include "IceClFlags.h"
 #include "IceDefs.h"
@@ -237,8 +238,8 @@ void Cfg::computeInOutEdges() {
 
   // Prune any unreachable nodes before computing in-edges.
   SizeT NumNodes = getNumNodes();
-  llvm::BitVector Reachable(NumNodes);
-  llvm::BitVector Pending(NumNodes);
+  BitVector Reachable(NumNodes);
+  BitVector Pending(NumNodes);
   Pending.set(getEntryNode()->getIndex());
   while (true) {
     int Index = Pending.find_first();
@@ -427,7 +428,7 @@ void Cfg::reorderNodes() {
 }
 
 namespace {
-void getRandomPostOrder(CfgNode *Node, llvm::BitVector &ToVisit,
+void getRandomPostOrder(CfgNode *Node, BitVector &ToVisit,
                         Ice::NodeList &PostOrder,
                         Ice::RandomNumberGenerator *RNG) {
   assert(ToVisit[Node->getIndex()]);
@@ -449,7 +450,7 @@ void Cfg::shuffleNodes() {
 
   NodeList ReversedReachable;
   NodeList Unreachable;
-  llvm::BitVector ToVisit(Nodes.size(), true);
+  BitVector ToVisit(Nodes.size(), true);
   // Create Random number generator for function reordering
   RandomNumberGenerator RNG(Ctx->getFlags().getRandomSeed(),
                             RPE_BasicBlockReordering, SequenceNumber);
@@ -813,7 +814,7 @@ void Cfg::liveness(LivenessMode Mode) {
   getVMetadata()->init(VMK_Uses);
   Live->init();
   // Initialize with all nodes needing to be processed.
-  llvm::BitVector NeedToProcess(Nodes.size(), true);
+  BitVector NeedToProcess(Nodes.size(), true);
   while (NeedToProcess.any()) {
     // Iterate in reverse topological order to speed up convergence.
     for (CfgNode *Node : reverse_range(Nodes)) {

src/IceDefs.h

@@ -21,7 +21,6 @@
 #include "IceTLS.h"
 
 #include "llvm/ADT/ArrayRef.h"
-#include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/ilist.h"
 #include "llvm/ADT/ilist_node.h"
 #include "llvm/ADT/iterator_range.h"
@@ -50,6 +49,7 @@
 namespace Ice {
 
 class Assembler;
+class BitVector;
 class Cfg;
 class CfgNode;
 class Constant;
@@ -155,7 +155,7 @@ using InstNumberT = int32_t;
 /// range.
 using LiveBeginEndMapEntry = std::pair<SizeT, InstNumberT>;
 using LiveBeginEndMap = CfgVector<LiveBeginEndMapEntry>;
-using LivenessBV = llvm::BitVector;
+using LivenessBV = BitVector;
 
 using TimerStackIdT = uint32_t;
 using TimerIdT = uint32_t;

src/IceLiveness.h

@@ -21,8 +21,9 @@
 #ifndef SUBZERO_SRC_ICELIVENESS_H
 #define SUBZERO_SRC_ICELIVENESS_H
 
-#include "IceCfgNode.h"
 #include "IceDefs.h"
+#include "IceBitVector.h"
+#include "IceCfgNode.h"
 #include "IceTypes.h"
 
 namespace Ice {
@@ -117,7 +118,7 @@ private:
   CfgVector<Variable *> LiveToVarMap;
   /// RangeMask[Variable::Number] indicates whether we want to track that
   /// Variable's live range.
-  llvm::BitVector RangeMask;
+  LivenessBV RangeMask;
 };
 
 } // end of namespace Ice

src/IceRNG.cpp

@@ -14,6 +14,7 @@
 
 #include "IceRNG.h"
 
+#include <climits>
 #include <ctime>
 
 namespace Ice {

src/IceTargetLowering.cpp

@@ -18,6 +18,7 @@
 
 #include "IceTargetLowering.h"
 
+#include "IceBitVector.h"
 #include "IceCfg.h" // setError()
 #include "IceCfgNode.h"
 #include "IceGlobalContext.h"
@@ -543,7 +544,7 @@ void TargetLowering::getVarStackSlotParams(
     uint32_t *SpillAreaAlignmentBytes, uint32_t *LocalsSlotsAlignmentBytes,
     std::function<bool(Variable *)> TargetVarHook) {
   const VariablesMetadata *VMetadata = Func->getVMetadata();
-  llvm::BitVector IsVarReferenced(Func->getNumVariables());
+  BitVector IsVarReferenced(Func->getNumVariables());
   for (CfgNode *Node : Func->getNodes()) {
     for (Inst &Instr : Node->getInsts()) {
       if (Instr.isDeleted())

src/IceTypes.cpp

@@ -16,6 +16,8 @@
 
 #include "IceDefs.h"
 
+#include <climits>
+
 namespace Ice {
 
 namespace {

src/main.cpp

@@ -17,6 +17,10 @@
 #include "IceBuildDefs.h"
 #include "IceCompileServer.h"
 
+#ifdef __pnacl__
+#include <malloc.h>
+#endif // __pnacl__
+
 /// Depending on whether we are building the compiler for the browser or
 /// standalone, we will end up creating a Ice::BrowserCompileServer or
 /// Ice::CLCompileServer object. Method
@@ -26,6 +30,15 @@
 /// We can only compile the Ice::BrowserCompileServer object with the PNaCl
 /// compiler toolchain, when building Subzero as a sandboxed translator.
 int main(int argc, char **argv) {
+#ifdef __pnacl__
+#define M_GRANULARITY (-2)
+  // PNaCl's default malloc implementation grabs small chunks of memory with
+  // mmap at a time, hence causing significant slowdowns. This call ensures that
+  // mmap is used to allocate 16MB at a time, to amortize the system call cost.
+  mallopt(M_GRANULARITY, 16 * 1024 * 1024);
+#undef M_GRANULARITY
+#endif // __pnacl__
+
   if (Ice::BuildDefs::browser()) {
     assert(argc == 1);
     return Ice::BrowserCompileServer().runAndReturnErrorCode();