[Clang] [Lexer] Optimize CharInfo.h utilities

[Thibault-Monnier]

This PR optimizes some of the utilities in Charinfo.h by replacing lookup table checks with a simple bounds or mask check when possible. This reduces instruction latency, allowing for a faster compilation overall.

This change does increase instruction count, but seems to be faster nonetheless: llvm-compile-time-tracker (that's a different commit, but basically the same changes).

@cor3ntin @AaronBallman

[llvmbot]

@llvm/pr-subscribers-clang

Author: Thibault Monnier (Thibault-Monnier)

Changes

This PR optimizes some of the utilities in Charinfo.h by replacing lookup table checks with a simple bounds or mask check when possible. This reduces instruction latency, allowing for a faster compilation overall.

This change does increase instruction count, but seems to be faster nonetheless: llvm-compile-time-tracker (that's a different commit, but basically the same changes).

@cor3ntin @AaronBallman

---

Full diff: https://github.com/llvm/llvm-project/pull/171052.diff

1 Files Affected:

• (modified) clang/include/clang/Basic/CharInfo.h (+9-16)

diff --git a/clang/include/clang/Basic/CharInfo.h b/clang/include/clang/Basic/CharInfo.h
index 87626eeb8a700..c34bcf4fbf88e 100644
--- a/clang/include/clang/Basic/CharInfo.h
+++ b/clang/include/clang/Basic/CharInfo.h
@@ -89,16 +89,15 @@ LLVM_READONLY inline bool isAsciiIdentifierContinue(unsigned char c,
 ///
 /// Note that this returns false for '\\0'.
 LLVM_READONLY inline bool isHorizontalWhitespace(unsigned char c) {
-  using namespace charinfo;
-  return (InfoTable[c] & (CHAR_HORZ_WS|CHAR_SPACE)) != 0;
+  constexpr unsigned long long Mask = 0b100000000000000000001101000000000;
+  return (c <= 32) && (Mask >> c) & 1;
 }
 
 /// Returns true if this character is vertical ASCII whitespace: '\\n', '\\r'.
 ///
 /// Note that this returns false for '\\0'.
 LLVM_READONLY inline bool isVerticalWhitespace(unsigned char c) {
-  using namespace charinfo;
-  return (InfoTable[c] & CHAR_VERT_WS) != 0;
+  return c == '\n' || c == '\r';
 }
 
 /// Return true if this character is horizontal or vertical ASCII whitespace:
@@ -106,26 +105,23 @@ LLVM_READONLY inline bool isVerticalWhitespace(unsigned char c) {
 ///
 /// Note that this returns false for '\\0'.
 LLVM_READONLY inline bool isWhitespace(unsigned char c) {
-  using namespace charinfo;
-  return (InfoTable[c] & (CHAR_HORZ_WS|CHAR_VERT_WS|CHAR_SPACE)) != 0;
+  constexpr unsigned long long Mask = 0b100000000000000000011111000000000;
+  return (c <= 32) && (Mask >> c) & 1;
 }
 
 /// Return true if this character is an ASCII digit: [0-9]
 LLVM_READONLY inline bool isDigit(unsigned char c) {
-  using namespace charinfo;
-  return (InfoTable[c] & CHAR_DIGIT) != 0;
+  return c >= '0' && c <= '9';
 }
 
 /// Return true if this character is a lowercase ASCII letter: [a-z]
 LLVM_READONLY inline bool isLowercase(unsigned char c) {
-  using namespace charinfo;
-  return (InfoTable[c] & CHAR_LOWER) != 0;
+  return c >= 'a' && c <= 'z';
 }
 
 /// Return true if this character is an uppercase ASCII letter: [A-Z]
 LLVM_READONLY inline bool isUppercase(unsigned char c) {
-  using namespace charinfo;
-  return (InfoTable[c] & CHAR_UPPER) != 0;
+  return c >= 'A' && c <= 'Z';
 }
 
 /// Return true if this character is an ASCII letter: [a-zA-Z]
@@ -158,9 +154,7 @@ LLVM_READONLY inline bool isPunctuation(unsigned char c) {
 /// character that should take exactly one column to print in a fixed-width
 /// terminal.
 LLVM_READONLY inline bool isPrintable(unsigned char c) {
-  using namespace charinfo;
-  return (InfoTable[c] & (CHAR_UPPER | CHAR_LOWER | CHAR_PERIOD | CHAR_PUNCT |
-                          CHAR_DIGIT | CHAR_UNDER | CHAR_SPACE)) != 0;
+  return c >= 32 && c <= 126;
 }
 
 /// Return true if this is the body character of a C preprocessing number,
@@ -236,7 +230,6 @@ LLVM_READONLY inline char toUppercase(char c) {
   return c;
 }
 
-
 /// Return true if this is a valid ASCII identifier.
 ///
 /// Note that this is a very simple check; it does not accept UCNs as valid

[AaronBallman]

This change does increase instruction count, but seems to be faster nonetheless: llvm-compile-time-tracker (that's a different commit, but basically the same changes).

I'd like to understand this a bit more because as best I can tell, this looks to regress performance. CC @nikic for more opinions on the perf side of things.

As far as the changes go, I prefer the original code because it was more clear what was going on. So unless this measurably improves performance, I'm not certain it's a direction we should head (though @cor3ntin is the code owner there, so maybe he feels differently).