The following patch' purpose is to reduce compile time for compilation of small

programs on targets with large register files. The root of the compile time
overhead was in the use of llvm::SmallVector to hold PhysRegEntries, which
resulted in slow-down from calling llvm::SmallVector::assign(N, 0). In contrast
std::vector uses the faster __platform_bzero to zero out primitive buffers when
assign is called, while SmallVector uses an iterator.

The fix for this was simply to replace the SmallVector with a dynamically
allocated buffer and to initialize or reinitialize the buffer based on the
total registers that the target architecture requires. The changes support
cases where a pass manager may be reused for different targets, and note that
the PhysRegEntries is allocated using calloc mainly for good for, and also to
quite tools like Valgrind (see comments for more info on this).

There is an rdar to track the fact that SmallVector doesn't have platform
specific speedup optimizations inside of it for things like this, and I'll
create a bugzilla entry at some point soon as well.

TL;DR: This fix replaces the expensive llvm::SmallVector<unsigned
char>::assign(N, 0) with a call to calloc for N bytes which is much faster
because SmallVector's assign uses iterators.

llvm-svn: 200917

lib/CodeGen/InterferenceCache.cpp

@@ -22,6 +22,22 @@ using namespace llvm;
 // Static member used for null interference cursors.
 InterferenceCache::BlockInterference InterferenceCache::Cursor::NoInterference;
 
+// Initializes PhysRegEntries (instead of a SmallVector, PhysRegEntries is a
+// buffer of size NumPhysRegs to speed up alloc/clear for targets with large
+// reg files). Calloced memory is used for good form, and quites tools like
+// Valgrind too, but zero initialized memory is not required by the algorithm:
+// this is because PhysRegEntries works like a SparseSet and its entries are
+// only valid when there is a corresponding CacheEntries assignment. There is
+// also support for when pass managers are reused for targets with different
+// numbers of PhysRegs: in this case PhysRegEntries is freed and reinitialized.
+void InterferenceCache::reinitPhysRegEntries() {
+  if (PhysRegEntriesCount == TRI->getNumRegs()) return;
+  free(PhysRegEntries);
+  PhysRegEntriesCount = TRI->getNumRegs();
+  PhysRegEntries = (unsigned char*)
+    calloc(PhysRegEntriesCount, sizeof(unsigned char));
+}
+
 void InterferenceCache::init(MachineFunction *mf,
                              LiveIntervalUnion *liuarray,
                              SlotIndexes *indexes,
@@ -30,7 +46,7 @@ void InterferenceCache::init(MachineFunction *mf,
   MF = mf;
   LIUArray = liuarray;
   TRI = tri;
-  PhysRegEntries.assign(TRI->getNumRegs(), 0);
+  reinitPhysRegEntries();
   for (unsigned i = 0; i != CacheEntries; ++i)
     Entries[i].clear(mf, indexes, lis);
 }

lib/CodeGen/InterferenceCache.h

@@ -135,7 +135,8 @@ class InterferenceCache {
 
   // Point to an entry for each physreg. The entry pointed to may not be up to
   // date, and it may have been reused for a different physreg.
-  SmallVector<unsigned char, 2> PhysRegEntries;
+  unsigned char* PhysRegEntries;
+  size_t PhysRegEntriesCount;
 
   // Next round-robin entry to be picked.
   unsigned RoundRobin;
@@ -147,7 +148,14 @@ class InterferenceCache {
   Entry *get(unsigned PhysReg);
 
 public:
-  InterferenceCache() : TRI(0), LIUArray(0), MF(0), RoundRobin(0) {}
+  InterferenceCache() : TRI(0), LIUArray(0), MF(0), PhysRegEntries(NULL),
+                        PhysRegEntriesCount(0), RoundRobin(0) {}
+
+  ~InterferenceCache() {
+    free(PhysRegEntries);
+  }
+
+  void reinitPhysRegEntries();
 
   /// init - Prepare cache for a new function.
   void init(MachineFunction*, LiveIntervalUnion*, SlotIndexes*, LiveIntervals*,