[clang] Speedup LineOffsetMapping::get

[serge-sans-paille]

Summary

LineOffsetMapping::get is a critical function that consistently appears
in the top 5 more computation intensive functions when running the
preprocessor.

This change brings consistent speedup of ~.5% on, preprocessing time,
see

https://llvm-compile-time-tracker.com/compare.php?from=0745b0c0354a0c8e1fefb68a3876d15db6c2e27a&to=460f3f04dac025e6952d78fce104a88151508a29&stat=instructions:u

for detailed statistics.

[serge-sans-paille]

Gentle ping :-)

[aaron.ballman]

LGTM, thank you!

[tschuett]

error: comparison of integers of different signs: 'long' and 'unsigned long' [-Werror,-Wsign-compare]
 if (End - Start > sizeof(Word)) {

clang/lib/Basic/SourceManager.cpp

1292

Sorry for being late. This comparison has issues.

[dyung]

@serge-sans-paille, one of our internal tests hit an assertion failure which I bisected back to your change. I have filed issue 59321 with a repro. Can you take a look?

clang/lib/Basic/SourceManager.cpp

LineOffsetMapping LineOffsetMapping::get(llvm::MemoryBufferRef Buffer,
 
   // Find the file offsets of all of the *physical* source lines.  This does
   // not look at trigraphs, escaped newlines, or anything else tricky.
   SmallVector<unsigned, 256> LineOffsets;
 
   // Line #1 starts at char 0.
   LineOffsets.push_back(0);
 
-  const unsigned char *Buf = (const unsigned char *)Buffer.getBufferStart();
+  const unsigned char *Start = (const unsigned char *)Buffer.getBufferStart();
   const unsigned char *End = (const unsigned char *)Buffer.getBufferEnd();
-  const std::size_t BufLen = End - Buf;
+  const unsigned char *Buf = Start;
 
-  unsigned I = 0;
   uint64_t Word;
 
   // scan sizeof(Word) bytes at a time for new lines.
   // This is much faster than scanning each byte independently.
-  if (BufLen > sizeof(Word)) {
+  if (End - Start > sizeof(Word)) {
     do {
-      Word = llvm::support::endian::read64(Buf + I, llvm::support::little);
+      Word = llvm::support::endian::read64(Buf, llvm::support::little);
       // no new line => jump over sizeof(Word) bytes.
       auto Mask = likelyhasbetween(Word, '\n', '\r');
       if (!Mask) {
-        I += sizeof(Word);
+        Buf += sizeof(Word);
         continue;
       }
 
       // At that point, Mask contains 0x80 set at each byte that holds a value
       // in [\n, \r + 1 [
 
       // Scan for the next newline - it's very likely there's one.
       unsigned N =
           llvm::countTrailingZeros(Mask) - 7; // -7 because 0x80 is the marker
       Word >>= N;
-      I += N / 8 + 1;
+      Buf += N / 8 + 1;
       unsigned char Byte = Word;
-      if (Byte == '\n') {
-        LineOffsets.push_back(I);
-      } else if (Byte == '\r') {
+      switch (Byte) {
+      case 'r':
         // If this is \r\n, skip both characters.
-        if (Buf[I] == '\n')
-          ++I;
-        LineOffsets.push_back(I);
+        if (*Buf == '\n') {
+          ++Buf;
         }
-    } while (I < BufLen - sizeof(Word) - 1);
+        LLVM_FALLTHROUGH;
+      case '\n':
+        LineOffsets.push_back(Buf - Start);
+      };
+    } while (Buf < End - sizeof(Word) - 1);
   }
 
   // Handle tail using a regular check.
-  while (I < BufLen) {
-    if (Buf[I] == '\n') {
-      LineOffsets.push_back(I + 1);
-    } else if (Buf[I] == '\r') {
+  while (Buf < End) {
+    if (*Buf == '\n') {
+      LineOffsets.push_back(Buf - Start + 1);
+    } else if (*Buf == '\r') {
       // If this is \r\n, skip both characters.
-      if (I + 1 < BufLen && Buf[I + 1] == '\n')
-        ++I;
-      LineOffsets.push_back(I + 1);
+      if (Buf + 1 < End && Buf[1] == '\n') {
+        ++Buf;
+      }
+      LineOffsets.push_back(Buf - Start + 1);
     }
-    ++I;
+    ++Buf;
   }
 
   return LineOffsetMapping(LineOffsets, Alloc);
 }
 
 LineOffsetMapping::LineOffsetMapping(ArrayRef<unsigned> LineOffsets,
                                      llvm::BumpPtrAllocator &Alloc)
     : Storage(Alloc.Allocate<unsigned>(LineOffsets.size() + 1)) {