Optimizations (#4006)

* Use an arena for parser metadata

* Use the parser arena for the constant pool

* Speed up the constant hash function

* Small optimization for parser_lex_magic_comment

* Scan forward through inline whitespace to avoid writing to parser->current.end continuously

* Fast-paths for ASCII-only identifiers

* Avoid unnecessary zero-ing of memory

* Pre-size arena to avoid unnecessary growth

* Force the allocation to be inlined

* Inline pm_node_list_append, pm_char_is_whitespace, and pm_char_is_inline_whitespace

* Avoid redundant whitespace scanning in magic comment lexing

* Potentially skip whitespace scanning for speed

* Inline three more functions, and lower the hash threshold for locals

* Lex simple integer values as we are lexing

* Only dispatch to lex_optional_float_suffix when it is possible

* Optimize constant pool hash for short strings

* Include string in constant pool entry to avoid chasing pointer

* SIMD/SWAR for strpbrk

* Fix a bug where we removed the \r warning

* Use a bloom filter to quickly reject local lookups

* Cache strpbrk lookup tables

* Fix up rebase errors

* More correctly detect SIMD on MSVC

* Ensure allocations to the constant pool are through the arena

* Fix ASAN reading off end of strpbrk cache

ruby/ruby@968b999

Co-Authored-By: Kevin Newton <kddnewton@gmail.com>

* Do not use GCC-specific syntax for lookup tables

ruby/ruby@5026acf

Co-Authored-By: Kevin Newton <kddnewton@gmail.com>

* Fix infinite loop in parser_lex_magic_comment

ruby/ruby@ec3162c

Co-Authored-By: Kevin Newton <kddnewton@gmail.com>

* Fix C coverage by moving stuff slightly around

---------

Co-authored-by: Earlopain <14981592+Earlopain@users.noreply.github.com>

Author: Earlopain

include/prism/defines.h

@@ -91,6 +91,18 @@
 #   define inline __inline
 #endif
 
+/**
+ * Force a function to be inlined at every call site. Use sparingly — only for
+ * small, hot functions where the compiler's heuristics fail to inline.
+ */
+#if defined(_MSC_VER)
+#   define PRISM_FORCE_INLINE __forceinline
+#elif defined(__GNUC__) || defined(__clang__)
+#   define PRISM_FORCE_INLINE inline __attribute__((always_inline))
+#else
+#   define PRISM_FORCE_INLINE inline
+#endif
+
 /**
  * Old Visual Studio versions before 2015 do not implement sprintf, but instead
  * implement _snprintf. We standard that here.
@@ -264,6 +276,49 @@
     #define PRISM_UNLIKELY(x) (x)
 #endif
 
+/**
+ * Platform detection for SIMD / fast-path implementations. At most one of
+ * these macros is defined, selecting the best available vectorization strategy.
+ */
+#if (defined(__aarch64__) && defined(__ARM_NEON)) || (defined(_MSC_VER) && defined(_M_ARM64))
+    #define PRISM_HAS_NEON
+#elif (defined(__x86_64__) && defined(__SSSE3__)) || (defined(_MSC_VER) && defined(_M_X64))
+    #define PRISM_HAS_SSSE3
+#elif defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    #define PRISM_HAS_SWAR
+#endif
+
+/**
+ * Count trailing zero bits in a 64-bit value. Used by SWAR identifier scanning
+ * to find the first non-matching byte in a word.
+ *
+ * Precondition: v must be nonzero. The result is undefined when v == 0
+ * (matching the behavior of __builtin_ctzll and _BitScanForward64).
+ */
+#if defined(__GNUC__) || defined(__clang__)
+    #define pm_ctzll(v) ((unsigned) __builtin_ctzll(v))
+#elif defined(_MSC_VER)
+    #include <intrin.h>
+    static inline unsigned pm_ctzll(uint64_t v) {
+        unsigned long index;
+        _BitScanForward64(&index, v);
+        return (unsigned) index;
+    }
+#else
+    static inline unsigned
+    pm_ctzll(uint64_t v) {
+        unsigned c = 0;
+        v &= (uint64_t) (-(int64_t) v);
+        if (v & 0x00000000FFFFFFFFULL) c += 0;  else c += 32;
+        if (v & 0x0000FFFF0000FFFFULL) c += 0;  else c += 16;
+        if (v & 0x00FF00FF00FF00FFULL) c += 0;  else c += 8;
+        if (v & 0x0F0F0F0F0F0F0F0FULL) c += 0;  else c += 4;
+        if (v & 0x3333333333333333ULL) c += 0;  else c += 2;
+        if (v & 0x5555555555555555ULL) c += 0;  else c += 1;
+        return c;
+    }
+#endif
+
 /**
  * We use -Wimplicit-fallthrough to guard potentially unintended fall-through between cases of a switch.
  * Use PRISM_FALLTHROUGH to explicitly annotate cases where the fallthrough is intentional.

include/prism/node.h

@@ -17,14 +17,31 @@
 #define PM_NODE_LIST_FOREACH(list, index, node) \
     for (size_t index = 0; index < (list)->size && ((node) = (list)->nodes[index]); index++)
 
+/**
+ * Slow path for pm_node_list_append: grow the list and append the node.
+ * Do not call directly — use pm_node_list_append instead.
+ *
+ * @param arena The arena to allocate from.
+ * @param list The list to append to.
+ * @param node The node to append.
+ */
+void pm_node_list_append_slow(pm_arena_t *arena, pm_node_list_t *list, pm_node_t *node);
+
 /**
  * Append a new node onto the end of the node list.
  *
  * @param arena The arena to allocate from.
  * @param list The list to append to.
  * @param node The node to append.
  */
-void pm_node_list_append(pm_arena_t *arena, pm_node_list_t *list, pm_node_t *node);
+static PRISM_FORCE_INLINE void
+pm_node_list_append(pm_arena_t *arena, pm_node_list_t *list, pm_node_t *node) {
+    if (list->size < list->capacity) {
+        list->nodes[list->size++] = node;
+    } else {
+        pm_node_list_append_slow(arena, list, node);
+    }
+}
 
 /**
  * Prepend a new node onto the beginning of the node list.

include/prism/parser.h

@@ -100,6 +100,13 @@ typedef struct {
     pm_heredoc_indent_t indent;
 } pm_heredoc_lex_mode_t;
 
+/**
+ * The size of the breakpoints and strpbrk cache charset buffers. All
+ * breakpoint arrays and the strpbrk cache charset must share this size so
+ * that memcmp can safely compare the full buffer without overreading.
+ */
+#define PM_STRPBRK_CACHE_SIZE 16
+
 /**
  * When lexing Ruby source, the lexer has a small amount of state to tell which
  * kind of token it is currently lexing. For example, when we find the start of
@@ -169,7 +176,7 @@ typedef struct pm_lex_mode {
              * This is the character set that should be used to delimit the
              * tokens within the list.
              */
-            uint8_t breakpoints[11];
+            uint8_t breakpoints[PM_STRPBRK_CACHE_SIZE];
         } list;
 
         struct {
@@ -191,7 +198,7 @@ typedef struct pm_lex_mode {
              * This is the character set that should be used to delimit the
              * tokens within the regular expression.
              */
-            uint8_t breakpoints[7];
+            uint8_t breakpoints[PM_STRPBRK_CACHE_SIZE];
         } regexp;
 
         struct {
@@ -224,7 +231,7 @@ typedef struct pm_lex_mode {
              * This is the character set that should be used to delimit the
              * tokens within the string.
              */
-            uint8_t breakpoints[7];
+            uint8_t breakpoints[PM_STRPBRK_CACHE_SIZE];
         } string;
 
         struct {
@@ -556,6 +563,13 @@ typedef struct pm_locals {
     /** The capacity of the local variables set. */
     uint32_t capacity;
 
+    /**
+     * A bloom filter over constant IDs stored in this set. Used to quickly
+     * reject lookups for names that are definitely not present, avoiding the
+     * cost of a linear scan or hash probe.
+     */
+    uint32_t bloom;
+
     /** The nullable allocated memory for the local variables in the set. */
     pm_local_t *locals;
 } pm_locals_t;
@@ -639,6 +653,9 @@ struct pm_parser {
     /** The arena used for all AST-lifetime allocations. Caller-owned. */
     pm_arena_t *arena;
 
+    /** The arena used for parser metadata (comments, diagnostics, etc.). */
+    pm_arena_t metadata_arena;
+
     /**
      * The next node identifier that will be assigned. This is a unique
      * identifier used to track nodes such that the syntax tree can be dropped
@@ -790,12 +807,26 @@ struct pm_parser {
     pm_line_offset_list_t line_offsets;
 
     /**
-     * We want to add a flag to integer nodes that indicates their base. We only
-     * want to parse these once, but we don't have space on the token itself to
-     * communicate this information. So we store it here and pass it through
-     * when we find tokens that we need it for.
+     * State communicated from the lexer to the parser for integer tokens.
      */
-    pm_node_flags_t integer_base;
+    struct {
+        /**
+         * A flag indicating the base of the integer (binary, octal, decimal,
+         * hexadecimal). Set during lexing and read during node creation.
+         */
+        pm_node_flags_t base;
+
+        /**
+         * When lexing a decimal integer that fits in a uint32_t, we compute
+         * the value during lexing to avoid re-scanning the digits during
+         * parsing. If lexed is true, this holds the result and
+         * pm_integer_parse can be skipped.
+         */
+        uint32_t value;
+
+        /** Whether value holds a valid pre-computed integer. */
+        bool lexed;
+    } integer;
 
     /**
      * This string is used to pass information from the lexer to the parser. It
@@ -938,6 +969,27 @@ struct pm_parser {
      * toggled with a magic comment.
      */
     bool warn_mismatched_indentation;
+
+#if defined(PRISM_HAS_NEON) || defined(PRISM_HAS_SSSE3) || defined(PRISM_HAS_SWAR)
+    /**
+     * Cached lookup tables for pm_strpbrk's SIMD fast path. Avoids rebuilding
+     * the nibble-based tables on every call when the charset hasn't changed
+     * (which is the common case during string/regex/list lexing).
+     */
+    struct {
+        /** The cached charset (null-terminated, NUL-padded). */
+        uint8_t charset[PM_STRPBRK_CACHE_SIZE];
+
+        /** Nibble-based low lookup table for SIMD matching. */
+        uint8_t low_lut[16];
+
+        /** Nibble-based high lookup table for SIMD matching. */
+        uint8_t high_lut[16];
+
+        /** Scalar fallback table (4 x 64-bit bitmasks covering all ASCII). */
+        uint64_t table[4];
+    } strpbrk_cache;
+#endif
 };
 
 #endif

include/prism/util/pm_arena.h

@@ -44,16 +44,52 @@ typedef struct {
     size_t block_count;
 } pm_arena_t;
 
+/**
+ * Ensure the arena has at least `capacity` bytes available in its current
+ * block, allocating a new block if necessary. This allows callers to
+ * pre-size the arena to avoid repeated small block allocations.
+ *
+ * @param arena The arena to pre-size.
+ * @param capacity The minimum number of bytes to ensure are available.
+ */
+void pm_arena_reserve(pm_arena_t *arena, size_t capacity);
+
+/**
+ * Slow path for pm_arena_alloc: allocate a new block and return a pointer to
+ * the first `size` bytes. Do not call directly — use pm_arena_alloc instead.
+ *
+ * @param arena The arena to allocate from.
+ * @param size The number of bytes to allocate.
+ * @returns A pointer to the allocated memory.
+ */
+void * pm_arena_alloc_slow(pm_arena_t *arena, size_t size);
+
 /**
  * Allocate memory from the arena. The returned memory is NOT zeroed. This
  * function is infallible — it aborts on allocation failure.
  *
+ * The fast path (bump pointer within the current block) is inlined at each
+ * call site. The slow path (new block allocation) is out-of-line.
+ *
  * @param arena The arena to allocate from.
  * @param size The number of bytes to allocate.
  * @param alignment The required alignment (must be a power of 2).
  * @returns A pointer to the allocated memory.
  */
-void * pm_arena_alloc(pm_arena_t *arena, size_t size, size_t alignment);
+static PRISM_FORCE_INLINE void *
+pm_arena_alloc(pm_arena_t *arena, size_t size, size_t alignment) {
+    if (arena->current != NULL) {
+        size_t used_aligned = (arena->current->used + alignment - 1) & ~(alignment - 1);
+        size_t needed = used_aligned + size;
+
+        if (used_aligned >= arena->current->used && needed >= used_aligned && needed <= arena->current->capacity) {
+            arena->current->used = needed;
+            return arena->current->data + used_aligned;
+        }
+    }
+
+    return pm_arena_alloc_slow(arena, size);
+}
 
 /**
  * Allocate zero-initialized memory from the arena. This function is infallible
@@ -64,7 +100,12 @@ void * pm_arena_alloc(pm_arena_t *arena, size_t size, size_t alignment);
  * @param alignment The required alignment (must be a power of 2).
  * @returns A pointer to the allocated, zero-initialized memory.
  */
-void * pm_arena_zalloc(pm_arena_t *arena, size_t size, size_t alignment);
+static inline void *
+pm_arena_zalloc(pm_arena_t *arena, size_t size, size_t alignment) {
+    void *ptr = pm_arena_alloc(arena, size, alignment);
+    memset(ptr, 0, size);
+    return ptr;
+}
 
 /**
  * Allocate memory from the arena and copy the given data into it. This is a
@@ -76,7 +117,12 @@ void * pm_arena_zalloc(pm_arena_t *arena, size_t size, size_t alignment);
  * @param alignment The required alignment (must be a power of 2).
  * @returns A pointer to the allocated copy.
  */
-void * pm_arena_memdup(pm_arena_t *arena, const void *src, size_t size, size_t alignment);
+static inline void *
+pm_arena_memdup(pm_arena_t *arena, const void *src, size_t size, size_t alignment) {
+    void *dst = pm_arena_alloc(arena, size, alignment);
+    memcpy(dst, src, size);
+    return dst;
+}
 
 /**
  * Free all blocks in the arena. After this call, all pointers returned by