yeptris 0.1.13.0 → 0.1.13.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 76bc800fd5065540397d957ca2162d2491474a086eda5b08436f016daa8f65e4
-  data.tar.gz: 683259ac45cfb084722337af80c066890bdf56f2608843343ebf8951589eafa8
+  metadata.gz: 288c8f3951e7cd252310e271ed666996442d66d90f04fbdefc8fde57fc122826
+  data.tar.gz: d52fb5127d673914f553b47449dbf2b57b2d1b581c5e0636c40c5faf99141222
 SHA512:
-  metadata.gz: 3b2be1bc057ee5a059e80ac087000afd2cee325473d834db34bd0f898318634e7b9be95dce9b523f909ae6884a56c9ede549a0e7142184eddf4cc96a13a11cb9
-  data.tar.gz: e95683f3b475a0dca70488a43cdf23a249d2bcd49c440b4c75bf31914265c0eae9fd2992d1bf6c0ebfa5b0487b796f1a24505e42b4715840c1731ce2af845a4c
+  metadata.gz: fc3a2f57041c9f28ec5133a98cf8573cfd820f9451d321ae6d333e507b1f2b9d7ce51814a0380eadee799904e056d013eea70142ebb6031d9c7480762f91634c
+  data.tar.gz: a6f0f5990a19597c1ac060e0dbb5abe2c921cdbdc2e49ec1fe70d04bfa91609c36c7b114e0d4e294e1175613a710fab722b1e524f6a1e5cd73f0340fbc2ead31

data/README.adoc

@@ -71,10 +71,20 @@ fastest first:
    ~29k-value corpus, Ruby 3.4:
 +
 ....
-JSON.parse        min 0.670  med 0.966  mean 0.910 ms
-Yeptris::JSON     min 0.637  med 0.674  mean 0.680 ms
-                  mean 0.75x faster, head-to-head 370/400 (92.5%)
+CI referee (gated at 1.05 on both platforms, order-alternating
+interleave, N=200; prefix-hashed token cache):
+  ubuntu (x86_64)   mean 0.745x  vs JSON.parse, 200/200 head-to-head
+  macos  (arm64)    mean 0.759x  vs JSON.parse, 169/200 head-to-head
 ....
+
+The parse-window GC strategy is per-arch and runtime-switchable:
+`Yeptris::Native.gc_mode` / `YEPTRIS_NATIVE_GC` = `:none`
+(steady-state: +0 heap pages, the stdlib's own GC cadence — the
+default on arm64) or `:disable` (no GC during the parse — the
+default on x86_64, faster where pages are free). On a LOADED x86_64
+box prefer `:none`: `:disable` grows ~478 heap pages per 50 parses,
+and page growth under contention is exactly the loaded-box
+regression. `:start` (in-window collection) measured 8x — dead.
 2. **Record-drain fallback** (always available): the strict-JSON
    validator gates, then the value records convert without the Psych
    quirk table — exact parity with engine 1, spec-pinned.

data/ext/yeptris_native/json_ruby.c

@@ -5,6 +5,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include "parse/scalars.h"
+#include <yeptris/visit.h>
 #include "scan/json.h"
 
 #define YEP_JR_MAX 1000
@@ -38,15 +39,26 @@ static void jr_ws(jr* j) {
     }
 }
 
+/* Token-cache knob (TODO.restructure/37): off = fresh strings
+ * everywhere (JSON.parse's shape: every key pays its own aset
+ * hashing); on = interned keys/tokens (shared frozen VALUEs
+ * memoize their hash). The CI referee A/Bs the net sign per arch. */
+enum { YEP_CACHE_ON = 0, YEP_CACHE_OFF = 1 };
+static int yep_cache_mode = YEP_CACHE_ON;
+
 static uint64_t jr_hash(const char* sp, long sl) {
-    /* FNV-1a 64 — cheap, good enough for short keys */
-    uint64_t h = 14695981039346656037ull;
-    for (long i = 0; i < sl; i++) {
-        h ^= (unsigned char)sp[i];
-        h *= 1099511628211ull;
-    }
-    h ^= (uint64_t)sl;
-    return h;
+    /* 8-byte-prefix key (the leptris nametab trick): one safe load
+     * of min(8, len) bytes + a multiply mix. Replaced byte-wise FNV
+     * (~15-20ns per token on the cached path, ~15k tokens per
+     * reference-corpus parse - the x86 materialization cost the
+     * decomposition isolated, TODO.restructure/37/38). */
+    uint64_t k = 0;
+    uint64_t take = (uint64_t)sl < 8 ? (uint64_t)sl : 8;
+    memcpy(&k, sp, (size_t)take);
+    k ^= (uint64_t)sl * 0x9E3779B97F4A7C15ull;
+    k *= 0xC2B2AE3D27D4EB4Full;
+    k ^= k >> 29;
+    return k;
 }
 
 static VALUE jr_cached(jr* j, const char* sp, long sl) {
@@ -86,7 +98,7 @@ static VALUE jr_str(jr* j, int as_key) {
         sp = j->p + start + 1;
         sl = (long)(close - start - 1);
     }
-    if (as_key || sl <= 24) return jr_cached(j, sp, sl);
+    if (yep_cache_mode == YEP_CACHE_ON && (as_key || sl <= 24)) return jr_cached(j, sp, sl);
     return rb_enc_str_new(sp, sl, j->enc);
 }
 
@@ -119,14 +131,18 @@ static VALUE jr_num(jr* j) {
     return DBL2NUM(dv);
 }
 
+/* Insert strategy (TODO.restructure/34): bulk lands all pairs in one
+ * rb_hash_bulk_insert (skips per-pair dispatch) but costs +1.4k
+ * intermediate allocations on the reference corpus — the CI referee
+ * rules per platform. aset = the per-pair rb_hash_aset loop. */
+enum { YEP_INS_BULK = 0, YEP_INS_ASET = 1 };
+static int yep_ins_mode = YEP_INS_BULK;
+
 static VALUE jr_object(jr* j) {
     j->i++; j->depth++;
     VALUE h = rb_hash_new_capa(8);
     jr_ws(j);
     if (j->i < j->len && j->p[j->i] == '}') { j->i++; j->depth--; return h; }
-    /* pairs collect into a flat buffer and land in ONE bulk insert:
-     * rb_hash_bulk_insert skips the per-pair method dispatch the
-     * aset loop pays (TODO.restructure/26's margin work) */
     VALUE pairs[64];
     VALUE* pv = pairs;
     size_t pcap = 64, pn = 0, heap_cap = 0;
@@ -140,23 +156,29 @@ static VALUE jr_object(jr* j) {
         j->i++;
         VALUE val = jr_value(j);
         if (j->err) goto out;
-        if (pn + 2 > pcap) {
-            size_t ncap = pcap * 2;
-            VALUE* nv = malloc(ncap * sizeof(VALUE));
-            if (!nv) { j->err = -1; goto out; }
-            memcpy(nv, pv, pn * sizeof(VALUE));
-            if (pv != pairs) { free(pv); heap_cap = 1; }
-            pv = nv; pcap = ncap;
+        if (yep_ins_mode == YEP_INS_ASET) {
+            rb_hash_aset(h, key, val);
+        } else {
+            if (pn + 2 > pcap) {
+                size_t ncap = pcap * 2;
+                VALUE* nv = malloc(ncap * sizeof(VALUE));
+                if (!nv) { j->err = -1; goto out; }
+                memcpy(nv, pv, pn * sizeof(VALUE));
+                if (pv != pairs) { free(pv); heap_cap = 1; }
+                pv = nv; pcap = ncap;
+            }
+            pv[pn++] = key;
+            pv[pn++] = val;
         }
-        pv[pn++] = key;
-        pv[pn++] = val;
         jr_ws(j);
         if (j->i >= j->len) { j->err = -2; goto out; }
         if (j->p[j->i] == ',') { j->i++; continue; }
         if (j->p[j->i] == '}') { j->i++; break; }
         j->err = -2; goto out;
     }
-    rb_hash_bulk_insert((long)pn, (const VALUE*)pv, h);
+    if (yep_ins_mode == YEP_INS_BULK) {
+        rb_hash_bulk_insert((long)pn, (const VALUE*)pv, h);
+    }
 out:
     if (pv != pairs) { free(pv); (void)heap_cap; }
     if (j->err) return Qnil;
@@ -205,16 +227,75 @@ static VALUE jr_value(jr* j) {
     j->err = -2; return Qnil;
 }
 
+/* GC strategy for the parse window (TODO.restructure/34): JSON.parse
+ * pays minor GCs mid-parse and recycles slots continuously; a blanket
+ * disable defers every collection — fresh pages each iteration, which
+ * is cheap idle and expensive exactly under memory contention (the
+ * loaded-box regression). The strategy is a runtime choice so the CI
+ * referee can A/B without rebuilds:
+ *   disable (default) — pause GC for the window
+ *   none               — never pause
+ *   start              — pause, then one gc_start before returning
+ *                        (pay the minor GC in-window, like JSON.parse)
+ */
+enum { YEP_GC_DISABLE = 0, YEP_GC_NONE = 1, YEP_GC_START = 2 };
+/* Default per arch (TODO.restructure/35, two CI rounds of evidence):
+ * - aarch64/darwin: NONE — +0 heap pages, the stdlib's own GC cadence
+ *   (0.745x mean, h2h 91% on mac runners; disable was 0.899x/48%).
+ * - x86_64: DISABLE — fresh CI VMs have free pages and cheaper
+ *   page faults than minor GCs (0.911-0.922x vs none's 1.06-1.20x).
+ * The mechanism cuts both ways under load: disable's +478 pages/50
+ * iters is exactly what a LOADED x86 box punishes — set
+ * YEPTRIS_NATIVE_GC=none there (documented in README). */
+#if defined(__aarch64__) || defined(__arm__) || defined(__ARMEL__) || defined(_M_ARM64)
+#define YEP_GC_DEFAULT YEP_GC_NONE
+#else
+#define YEP_GC_DEFAULT YEP_GC_DISABLE
+#endif
+static int yep_gc_mode = YEP_GC_DEFAULT;
+
+int yep_rb_gc_mode(void) { return yep_gc_mode; }
+int yep_rb_ins_mode(void) { return yep_ins_mode; }
+int yep_rb_cache_mode(void) { return yep_cache_mode; }
+void yep_rb_set_cache_mode(int mode) {
+    if (mode == YEP_CACHE_ON || mode == YEP_CACHE_OFF) yep_cache_mode = mode;
+}
+void yep_rb_set_ins_mode(int mode) {
+    if (mode == YEP_INS_BULK || mode == YEP_INS_ASET) yep_ins_mode = mode;
+}
+void yep_rb_set_gc_mode(int mode) {
+    if (mode >= YEP_GC_DISABLE && mode <= YEP_GC_START) {
+        yep_gc_mode = mode;
+    }
+}
+
+/* Pure grammar walk, no materialization (TODO.restructure/37): the
+ * same scan kernels through the null vtable. Decomposes scan vs
+ * materialize cost. Returns seconds for n iterations. */
+#include <time.h>
+double yep_rb_scan_time(const char* p, size_t len, int n) {
+    static const YeptrisVisitVTable none = {0};
+    struct timespec t0, t1;
+    clock_gettime(CLOCK_MONOTONIC, &t0);
+    for (int i = 0; i < n; i++) {
+        (void)yeptris_visit_json(p, len, &none, NULL);
+    }
+    clock_gettime(CLOCK_MONOTONIC, &t1);
+    return (double)(t1.tv_sec - t0.tv_sec) + (double)(t1.tv_nsec - t0.tv_nsec) / 1e9;
+}
+
 VALUE yep_rb_parse_json(const char* p, size_t len) {
     jr j;
     memset(&j, 0, sizeof(j));
     j.p = p; j.len = len; j.enc = rb_utf8_encoding();
-    VALUE already = rb_gc_disable();
+    VALUE already = Qtrue;
+    if (yep_gc_mode != YEP_GC_NONE) already = rb_gc_disable();
     VALUE v = jr_value(&j);
     jr_ws(&j);
     if (j.i != len && j.err == 0) j.err = -2;
     free(j.scratch);
     if (already == Qfalse) rb_gc_enable();
+    if (yep_gc_mode == YEP_GC_START && j.err == 0) rb_gc_start();
     if (j.err == -1) rb_raise(rb_eNoMemError, "yeptris native json");
     if (j.err != 0) rb_raise(rb_path2class("Yeptris::ParseError"), "native json parse failed");
     return v;

data/ext/yeptris_native/yeptris_native.c

@@ -322,6 +322,75 @@ static VALUE native_load_stream(VALUE self, VALUE input, VALUE schema) {
     return ctx_result(&s.inner, st == YEPTRIS_OK ? YEPTRIS_ERROR_INTERNAL : st);
 }
 
+/* GC-strategy surface (TODO.restructure/34): ENV at load sets the
+ * default; the setter re-picks at runtime so the CI referee can A/B
+ * in-process. Symbols: :disable, :none, :start. */
+extern int yep_rb_gc_mode(void);
+extern void yep_rb_set_gc_mode(int mode);
+extern int yep_rb_ins_mode(void);
+extern void yep_rb_set_ins_mode(int mode);
+extern int yep_rb_cache_mode(void);
+extern void yep_rb_set_cache_mode(int mode);
+extern double yep_rb_scan_time(const char* p, size_t len, int n);
+
+static VALUE native_gc_mode(VALUE self) {
+    (void)self;
+    switch (yep_rb_gc_mode()) {
+    case 1: return ID2SYM(rb_intern("none"));
+    case 2: return ID2SYM(rb_intern("start"));
+    default: return ID2SYM(rb_intern("disable"));
+    }
+}
+
+static VALUE native_ins_mode(VALUE self) {
+    (void)self;
+    return yep_rb_ins_mode() == 1 ? ID2SYM(rb_intern("aset")) : ID2SYM(rb_intern("bulk"));
+}
+
+static VALUE native_ins_mode_set(VALUE self, VALUE mode) {
+    (void)self;
+    Check_Type(mode, T_SYMBOL);
+    ID id = rb_sym2id(mode);
+    if (id == rb_intern("bulk")) yep_rb_set_ins_mode(0);
+    else if (id == rb_intern("aset")) yep_rb_set_ins_mode(1);
+    else rb_raise(rb_eArgError, "ins_mode must be :bulk or :aset");
+    return mode;
+}
+
+static VALUE native_cache_mode(VALUE self) {
+    (void)self;
+    return yep_rb_cache_mode() == 1 ? ID2SYM(rb_intern("off")) : ID2SYM(rb_intern("on"));
+}
+
+static VALUE native_cache_mode_set(VALUE self, VALUE mode) {
+    (void)self;
+    Check_Type(mode, T_SYMBOL);
+    ID id = rb_sym2id(mode);
+    if (id == rb_intern("on")) yep_rb_set_cache_mode(0);
+    else if (id == rb_intern("off")) yep_rb_set_cache_mode(1);
+    else rb_raise(rb_eArgError, "cache_mode must be :on or :off");
+    return mode;
+}
+
+static VALUE native_scan_time(VALUE self, VALUE input, VALUE count) {
+    (void)self;
+    StringValue(input);
+    int n = NUM2INT(count);
+    double secs = yep_rb_scan_time(RSTRING_PTR(input), (size_t)RSTRING_LEN(input), n);
+    return DBL2NUM(secs / (double)n);
+}
+
+static VALUE native_gc_mode_set(VALUE self, VALUE mode) {
+    (void)self;
+    Check_Type(mode, T_SYMBOL);
+    ID id = rb_sym2id(mode);
+    if (id == rb_intern("disable")) yep_rb_set_gc_mode(0);
+    else if (id == rb_intern("none")) yep_rb_set_gc_mode(1);
+    else if (id == rb_intern("start")) yep_rb_set_gc_mode(2);
+    else rb_raise(rb_eArgError, "gc_mode must be :disable, :none, or :start");
+    return mode;
+}
+
 RUBY_FUNC_EXPORTED void Init_native(void) {
     utf8_enc = rb_utf8_encoding();
     VALUE mYep = rb_define_module("Yeptris");
@@ -329,5 +398,20 @@ RUBY_FUNC_EXPORTED void Init_native(void) {
     rb_define_singleton_method(mNat, "load_json", native_load_json, 1);
     rb_define_singleton_method(mNat, "load", native_load, 2);
     rb_define_singleton_method(mNat, "load_stream", native_load_stream, 2);
+    rb_define_singleton_method(mNat, "gc_mode", native_gc_mode, 0);
+    rb_define_singleton_method(mNat, "gc_mode=", native_gc_mode_set, 1);
+    rb_define_singleton_method(mNat, "ins_mode", native_ins_mode, 0);
+    rb_define_singleton_method(mNat, "ins_mode=", native_ins_mode_set, 1);
+    rb_define_singleton_method(mNat, "cache_mode", native_cache_mode, 0);
+    rb_define_singleton_method(mNat, "cache_mode=", native_cache_mode_set, 1);
+    rb_define_singleton_method(mNat, "scan_time", native_scan_time, 2);
     rb_define_const(mNat, "AVAILABLE", Qtrue);
+    const char* env = getenv("YEPTRIS_NATIVE_GC");
+    if (env != NULL && strcmp(env, "none") == 0) yep_rb_set_gc_mode(1);
+    else if (env != NULL && strcmp(env, "start") == 0) yep_rb_set_gc_mode(2);
+    else if (env != NULL && strcmp(env, "disable") == 0) yep_rb_set_gc_mode(0);
+    const char* ins = getenv("YEPTRIS_NATIVE_INSERT");
+    if (ins != NULL && strcmp(ins, "aset") == 0) yep_rb_set_ins_mode(1);
+    const char* cache = getenv("YEPTRIS_NATIVE_CACHE");
+    if (cache != NULL && strcmp(cache, "off") == 0) yep_rb_set_cache_mode(1);
 }

data/lib/yeptris.rb

@@ -3,7 +3,7 @@
 module Yeptris
   # The gem's version lives in the parent namespace's file — the last
   # internal require (yeptris/version) retired with it.
-  VERSION = "0.1.13.0".freeze
+  VERSION = "0.1.13.2".freeze
   # The error hierarchy lives in THIS file (the parent namespace's
   # own file): nested constants do not trigger a parent-constant
   # autoload, and the law forbids internal requires — defining the

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: yeptris
 version: !ruby/object:Gem::Version
-  version: 0.1.13.0
+  version: 0.1.13.2
 platform: ruby
 authors:
 - Ribose Inc.