diff options
3 files changed, 297 insertions, 0 deletions
diff --git a/research/experiments/045-proof-compiler/2026-10-19-range.jsonl b/research/experiments/045-proof-compiler/2026-10-19-range.jsonl new file mode 100644 index 0000000..d814ced --- /dev/null +++ b/research/experiments/045-proof-compiler/2026-10-19-range.jsonl @@ -0,0 +1,6 @@ +{"what": "range() in the executor and stitched: (variant v0r) - (variant v0), events.c, PURE_JIT=0, 3 rounds, two runs; v0 the same in both builds (0.63 instructions a point run)", "executor": {"instructions_per_point": [14487.9, 14489.199999999999], "cycles_per_point": [5512.7, 5503.6]}, "stitched": {"instructions_per_point": [8144.1, 8145.299999999999], "cycles_per_point": [3224.1, 3193.7]}, "answers": "range-out stitched (17,262 to 10,601 a point, so it ran) gave the executor's answers to range.c, which the five ways equal"} +{"way": "exact", "code_points": 61285, "rounds_of_halving_per_point": 11.4, "rounds": 5, "instructions_per_point": 2474.0, "cycles_per_point": 1315.3, "instructions_per_round": 217.06, "cycles_per_round": 115.4, "answers": "the executor's"} +{"way": "integer", "code_points": 61285, "rounds_of_halving_per_point": 11.4, "rounds": 5, "instructions_per_point": 1586.6, "cycles_per_point": 875.4, "instructions_per_round": 139.2, "cycles_per_round": 76.81, "answers": "the executor's"} +{"way": "flat", "code_points": 61285, "rounds_of_halving_per_point": 11.4, "rounds": 5, "instructions_per_point": 1083.0, "cycles_per_point": 633.2, "instructions_per_round": 95.01, "cycles_per_round": 55.55, "answers": "the executor's"} +{"way": "both", "code_points": 61285, "rounds_of_halving_per_point": 11.4, "rounds": 5, "instructions_per_point": 226.2, "cycles_per_point": 124.5, "instructions_per_round": 19.84, "cycles_per_round": 10.92, "answers": "the executor's"} +{"way": "c", "code_points": 61285, "rounds_of_halving_per_point": 11.4, "rounds": 5, "instructions_per_point": 129.0, "cycles_per_point": 92.7, "instructions_per_round": 11.32, "cycles_per_round": 8.13, "answers": "the executor's"} diff --git a/research/experiments/045-proof-compiler/range.c b/research/experiments/045-proof-compiler/range.c new file mode 100644 index 0000000..b358318 --- /dev/null +++ b/research/experiments/045-proof-compiler/range.c @@ -0,0 +1,255 @@ +/* Research 045, the next experiment: UAX #14's range() (aup's examples/linebreak.aup: the last range whose start is at + * or before a code point, by halving), 36% of the app's machine instructions in the executor, written by hand as an + * emitter would write it, in ways that keep or drop what the verifier's proof could license: + * + * exact Pure's meaning as the executor runs it: numbers are doubles and each sum and quotient is checked finite; + * // is floor_divide (fmod) and checks its divisor; ranges[middle] checks a whole number and the bounds + * (INDEX_AT), .start checks the bounds again and reads the stored list through pure_list_get (ITEM_GET); + * each round of the loop charges the work budget (iterate), and so does the call. Locals in registers. + * integer exact, but low, high and middle are integers (from 0, ranges.count and //): no fmod, no finite checks, + * no whole-number check; the bounds checks and pure_list_get kept. + * flat exact, but .start read from the list's start column held flat (the list does not change in the event). + * both integer and flat: integers read from a flat column, the bounds check and the budget kept. + * c the same halving in plain C on the flat column: no checks, no budget. + * + * Each answers every code point of the lines; the answers must equal the executor's (an app answering each point's + * range: variants/range-out.patch), read from its view first. Then machine instructions and cycles (perf_event_open, + * user mode) a code point and a round of halving, ROUNDS passes each, the ways interleaved. + * + * cc -O2 -std=c17 -I../../../core range.c ../../../build/libpure.a -lm -pthread -o range + * ./range range-out.pmod RANGES LINES ROUNDS one JSON line a way */ +#define _GNU_SOURCE +#include "module.h" +#include "module-list.h" +#include <math.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <linux/perf_event.h> +#include <sys/ioctl.h> +#include <sys/syscall.h> +#include <unistd.h> + +static unsigned char *whole(const char *path, size_t *size) { + FILE *f = fopen(path, "rb"); + if (!f) return NULL; + fseek(f, 0, SEEK_END); + *size = (size_t)ftell(f); + rewind(f); + unsigned char *bytes = malloc(*size ? *size : 1); + if (fread(bytes, 1, *size, f) != *size) { free(bytes); bytes = NULL; } + fclose(f); + return bytes; +} + +/* What the emitted code is given: the stored list, its start column held flat, and the work budget (as PureApp's). */ +typedef struct { const PureList *ranges; const double *starts; size_t count; uint64_t work, budget; } Context; +enum { ROUND = 24, CALL = 8 }; /* what a round of the loop and a call charge (their instructions as written) */ + +static inline bool charge(Context *x, uint64_t instructions) { + if (instructions > x->budget - x->work) return false; + x->work += instructions; + return true; +} + +/* core/module.c's, as it is. */ +static double floor_divide(double a, double b) { + double mod = fmod(a, b), quotient = (a - mod) / b; + if (mod != 0 && (b < 0) != (mod < 0)) quotient -= 1.0; + if (quotient == 0) return copysign(0.0, a / b); + double whole = floor(quotient); + return quotient - whole > 0.5 ? whole + 1.0 : whole; +} + +/* -1: a fault (the executor's would stop the event). */ +static __attribute__((noinline)) double range_exact(Context *x, double c) { + if (!charge(x, CALL)) return -1; + double low = 0, high = (double)pure_list_count(x->ranges); + for (;;) { + double span = high - low; + if (!isfinite(span)) return -1; + if (!(span > 1)) break; + if (!charge(x, ROUND)) return -1; + double sum = low + high; + if (!isfinite(sum)) return -1; + double middle = floor_divide(sum, 2); + if (!isfinite(middle)) return -1; + if (middle != floor(middle) || middle < 0 || middle >= (double)pure_list_count(x->ranges)) return -1; /* INDEX_AT */ + size_t at = (size_t)middle; + if (at >= pure_list_count(x->ranges)) return -1; /* ITEM_GET */ + double start = pure_list_get(x->ranges, at, 0).num; + if (start <= c) low = middle; + else high = middle; + } + return low; +} + +static __attribute__((noinline)) double range_integer(Context *x, double c) { + if (!charge(x, CALL)) return -1; + int64_t low = 0, high = (int64_t)pure_list_count(x->ranges); + while (high - low > 1) { + if (!charge(x, ROUND)) return -1; + int64_t middle = (low + high) / 2; /* both whole and at or above 0: // is / */ + if ((uint64_t)middle >= pure_list_count(x->ranges)) return -1; + if (pure_list_get(x->ranges, (size_t)middle, 0).num <= c) low = middle; + else high = middle; + } + return (double)low; +} + +static __attribute__((noinline)) double range_flat(Context *x, double c) { + if (!charge(x, CALL)) return -1; + double low = 0, high = (double)x->count; + for (;;) { + double span = high - low; + if (!isfinite(span)) return -1; + if (!(span > 1)) break; + if (!charge(x, ROUND)) return -1; + double sum = low + high; + if (!isfinite(sum)) return -1; + double middle = floor_divide(sum, 2); + if (!isfinite(middle)) return -1; + if (middle != floor(middle) || middle < 0 || middle >= (double)x->count) return -1; + size_t at = (size_t)middle; + if (at >= x->count) return -1; + if (x->starts[at] <= c) low = middle; + else high = middle; + } + return low; +} + +static __attribute__((noinline)) double range_both(Context *x, double c) { + if (!charge(x, CALL)) return -1; + int64_t low = 0, high = (int64_t)x->count; + while (high - low > 1) { + if (!charge(x, ROUND)) return -1; + int64_t middle = (low + high) / 2; + if ((uint64_t)middle >= x->count) return -1; + if (x->starts[middle] <= c) low = middle; + else high = middle; + } + return (double)low; +} + +static __attribute__((noinline)) double range_c(Context *x, double c) { + size_t low = 0, high = x->count; + while (high - low > 1) { + size_t middle = (low + high) / 2; + if (x->starts[middle] <= c) low = middle; + else high = middle; + } + return (double)low; +} + +static double rounds_of(const Context *x, double c) { /* the rounds of halving a code point takes */ + size_t low = 0, high = x->count, n = 0; + while (high - low > 1) { + size_t middle = (low + high) / 2; + ++n; + if (x->starts[middle] <= c) low = middle; + else high = middle; + } + return (double)n; +} + +static int counter(uint64_t config) { + struct perf_event_attr a; + memset(&a, 0, sizeof(a)); + a.size = sizeof(a), a.type = PERF_TYPE_HARDWARE, a.config = config, a.disabled = 1, a.exclude_kernel = 1, a.exclude_hv = 1; + return (int)syscall(SYS_perf_event_open, &a, 0, -1, -1, 0); +} + +static uint64_t read_counter(int fd) { + uint64_t v = 0; + if (read(fd, &v, sizeof(v)) != sizeof(v)) return 0; + return v; +} + +/* The executor's answers: each point's range as two bytes, low first. */ +typedef struct { unsigned char *bytes; size_t at, room; } Answers; +static bool take(void *context, PureSpan span) { + Answers *a = context; + for (size_t i = 0; i < span.size && a->at < a->room; ++i) a->bytes[a->at++] = span.data[i]; + return true; +} + +typedef double (*Way)(Context *, double); + +int main(int argc, char **argv) { + if (argc != 5) { fprintf(stderr, "usage: %s range-out.pmod RANGES LINES ROUNDS\n", argv[0]); return 2; } + size_t ranges_size, lines_size; + unsigned char *ranges = whole(argv[2], &ranges_size), *bytes = whole(argv[3], &lines_size); + int rounds = atoi(argv[4]); + if (!ranges || !bytes) return 1; + PureModule module = {0}; + char error[160]; + if (pure_module_read(argv[1], &module, error, sizeof(error)) != PURE_OK) { fprintf(stderr, "%s\n", error); return 1; } + PureApp app; + if (pure_app_start(&app, &module, (PureData){0}) != PURE_OK) return 1; + int load = pure_module_input(&module, "load"), line = pure_module_input(&module, "line"); + int state = pure_module_state(&module, "ranges"), view = pure_module_view(&module, "boundaries"); + if (load < 0 || line < 0 || state < 0 || view < 0) return 1; + if (pure_app_message(&app, (uint32_t)load, (PureSpan){ranges, ranges_size}) != PURE_OK) return 1; + Context x = {.ranges = app.lists[state], .budget = UINT64_MAX}; + x.count = pure_list_count(x.ranges); + double *starts = malloc(x.count * sizeof(double)); + for (size_t i = 0; i < x.count; ++i) starts[i] = pure_list_get(x.ranges, i, 0).num; + x.starts = starts; + /* The code points, and the executor's answer for each (a line at a time, as the app takes them). */ + size_t points = 0, wrong = 0; + double *codes = NULL, *want = NULL, iterations = 0; + for (size_t at = 0; at + 4 <= lines_size;) { + uint32_t count; + memcpy(&count, bytes + at, 4); + const uint32_t *line_codes = (const uint32_t *)(bytes + at + 4); + at += 4 + 4 * (size_t)count + count + 1; + codes = realloc(codes, (points + count) * sizeof(double)), want = realloc(want, (points + count) * sizeof(double)); + unsigned char *payload = malloc(8 * (size_t)count + 1), answer[2 * 4096]; + for (uint32_t i = 0; i < count; ++i) { + double c = line_codes[i]; + memcpy(payload + 8 * i, &c, 8); + codes[points + i] = c; + } + Answers a = {answer, 0, sizeof(answer)}; + if (pure_app_message(&app, (uint32_t)line, (PureSpan){payload, 8 * (size_t)count}) != PURE_OK || + pure_app_view_text(&app, (uint32_t)view, take, &a) != PURE_OK || a.at != 2 * (size_t)count) + return 1; + for (uint32_t i = 0; i < count; ++i) want[points + i] = answer[2 * i] + 256.0 * answer[2 * i + 1]; + free(payload); + points += count; + } + for (size_t i = 0; i < points; ++i) iterations += rounds_of(&x, codes[i]); + const char *names[] = {"exact", "integer", "flat", "both", "c"}; + Way ways[] = {range_exact, range_integer, range_flat, range_both, range_c}; + enum { WAYS = 5 }; + for (int w = 0; w < WAYS; ++w) + for (size_t i = 0; i < points; ++i) + if (ways[w](&x, codes[i]) != want[i]) { + if (wrong++ < 5) fprintf(stderr, "%s: point %zu (%.0f): %.0f, the executor %.0f\n", names[w], i, codes[i], + ways[w](&x, codes[i]), want[i]); + } + if (wrong) { fprintf(stderr, "%zu answers differ\n", wrong); return 1; } + int ins = counter(PERF_COUNT_HW_INSTRUCTIONS), cyc = counter(PERF_COUNT_HW_CPU_CYCLES); + if (ins < 0 || cyc < 0) { perror("perf_event_open"); return 1; } + uint64_t total_ins[WAYS] = {0}, total_cyc[WAYS] = {0}; + volatile double sink = 0; + for (int r = 0; r < rounds; ++r) + for (int w = 0; w < WAYS; ++w) { + x.work = 0; + ioctl(ins, PERF_EVENT_IOC_RESET, 0), ioctl(cyc, PERF_EVENT_IOC_RESET, 0); + ioctl(ins, PERF_EVENT_IOC_ENABLE, 0), ioctl(cyc, PERF_EVENT_IOC_ENABLE, 0); + double s = 0; + for (size_t i = 0; i < points; ++i) s += ways[w](&x, codes[i]); + ioctl(ins, PERF_EVENT_IOC_DISABLE, 0), ioctl(cyc, PERF_EVENT_IOC_DISABLE, 0); + sink += s; + total_ins[w] += read_counter(ins), total_cyc[w] += read_counter(cyc); + } + double n = (double)points * rounds; + for (int w = 0; w < WAYS; ++w) + printf("{\"way\":\"%s\",\"code_points\":%zu,\"rounds_of_halving_per_point\":%.2f,\"rounds\":%d,\"instructions_per_point\":%.1f," + "\"cycles_per_point\":%.1f,\"instructions_per_round\":%.2f,\"cycles_per_round\":%.2f,\"answers\":\"the executor's\"}\n", + names[w], points, iterations / points, rounds, total_ins[w] / n, total_cyc[w] / n, + total_ins[w] / (iterations * rounds), total_cyc[w] / (iterations * rounds)); + pure_app_stop(&app); + return 0; +} diff --git a/research/experiments/045-proof-compiler/variants/range-out.patch b/research/experiments/045-proof-compiler/variants/range-out.patch new file mode 100644 index 0000000..5cad619 --- /dev/null +++ b/research/experiments/045-proof-compiler/variants/range-out.patch @@ -0,0 +1,36 @@ +--- linebreak.aup ++++ range-out.aup +@@ -153,30 +153,11 @@ + } + + on line { +- let codes, cls, eff, start, nonspace, ri, at: list of number +- let joined, numeric: list of bool ++ let out: list of i64 + for p in line.points { +- let i = codes.count + let r = range(p.code) +- let c = ranges[r].class +- let mark = c == CM || c == ZWJ +- let previous = if i > 0 { eff[i - 1] } else { SP } +- let joins = mark && i > 0 && !end(previous) && previous != SP && previous != ZW # LB9, LB10 +- let e = if joins { previous } else if mark { AL } else { c } +- codes.append(p.code) +- cls.append(c) +- at.append(r) +- joined.append(joins) +- eff.append(e) +- start.append(if joins { start[i - 1] } else { i }) +- nonspace.append(if e != SP { start[i] } else if i > 0 { nonspace[i - 1] } else { -1 }) +- numeric.append(if joins { numeric[i - 1] } else { e == NU || ((e == SY || e == IS) && i > 0 && numeric[i - 1]) }) +- ri.append(if joins { ri[i - 1] } else if e == RI { (if i > 0 { ri[i - 1] } else { 0 }) + 1 } else { 0 }) ++ out.append(i64(r % 256)) ++ out.append(i64(r // 256)) + } +- let t = Text { codes: codes, cls: cls, eff: eff, joined: joined, start: start, nonspace: nonspace, numeric: numeric, ri: ri, +- at: at } +- let out: list of i64 = [0] # LB2 +- for i in 1 .. codes.count { out.append(i64(decide(t, i))) } +- if codes.count > 0 { out.append(2) } # LB3 + breaks = bytes(out) + } |