add v1 + v2

Makefile

@@ -1,8 +1,8 @@
 
 CC = gcc
 
-CFLAGS = -std=c17 -Wall -Wextra -Wpedantic -Wconversion -g -pg
-LFLAGS = -lm -g -pg
+CFLAGS = -std=c17 -Wall -Wextra -Wpedantic -Wconversion -g
+LFLAGS = -lm -g
 
 TARGET= bf
 
@@ -18,7 +18,7 @@ disasm: $(TARGET)
 
 profile: $(TARGET)
 	$(RM) perf.data profile.perf
-	perf record -g -F max ./bf
+	perf record --call-graph dwarf -F 10000 ./$(TARGET) < bench.bf
 	perf script -F +pid > profile.perf
 
 $(TARGET): $(OBJECT_FILES)

bench.bf

@@ -0,0 +1 @@
+++++++++[->-[->-[->-[-]<]<]<]>++++++++[<++++++++++>-]<[>+>+<<-]>-.>-----.>

src/fastbf.h

@@ -0,0 +1,9 @@
+#ifndef FASTBF_H
+#define FASTBF_H
+
+#include <stddef.h>
+
+void v1(const char* program, size_t length);
+void v2(const char* program, size_t length);
+
+#endif

src/main.c

@@ -1,20 +1,12 @@
+#include "fastbf.h"
+#include <stdint.h>
 #include <stdio.h>
-
-int takes_long(int b)
-{
-    int a = 0;
-    for (int i = 0; i < b; i++) {
-        a = (a + 1) % 4;
-    }
-    return a;
-}
-
-int get_a(void) { return takes_long(100000001); }
-int get_b(void) { return takes_long(200100001); }
+#include <stdlib.h>
+#include <string.h>
 
 int main(void)
 {
-    puts("calculating...");
-    int c = get_a() + get_b();
-    printf("result = %d\n", c);
+    char program[30000];
+    fgets(program, 30000, stdin);
+    v2(program, strlen(program));
 }

src/v1.c

@@ -0,0 +1,63 @@
+#include "fastbf.h"
+#include <stdint.h>
+#include <stdio.h>
+
+/*
+simon@gamerarch:~/W/fastbf$ time ./bf < bench.bf
+OK
+real    0m0,680s
+user    0m0,676s
+sys     0m0,004s
+simon@gamerarch:~/W/fastbf$ time ./bf < bench.bf
+OK
+real    0m0,682s
+user    0m0,677s
+sys     0m0,003s
+simon@gamerarch:~/W/fastbf$ time ./bf < bench.bf
+OK
+real    0m0,679s
+user    0m0,679s
+sys     0m0,000s
+*/
+
+void v1(const char* program, size_t length)
+{
+    (void)length;
+    uint8_t memory[30000];
+    uint_fast16_t memory_index = 0;
+    for (uint_fast16_t i = 0; i < length; i++) {
+        char c = program[i];
+        switch (c) {
+            case '+':
+                memory[memory_index]++;
+                break;
+            case '-':
+                memory[memory_index]--;
+                break;
+            case '>':
+                memory_index++;
+                break;
+            case '<':
+                memory_index--;
+                break;
+            case '.':
+                putc(memory[memory_index], stdout);
+                break;
+            case ',':
+                memory[memory_index] = (uint8_t)getc(stdin);
+                break;
+            case ']':
+                if (memory[memory_index]) {
+                    int depth = 1;
+                    while (program[i] != '[' || depth != 0) {
+                        i--;
+                        if (program[i] == ']')
+                            depth++;
+                        else if (program[i] == '[')
+                            depth--;
+                    }
+                }
+                break;
+        }
+    }
+}

src/v2.c

@@ -0,0 +1,125 @@
+
+#include "fastbf.h"
+#include <stdint.h>
+#include <stdio.h>
+
+/*
+simon@gamerarch:~/W/fastbf$ time ./bf < bench.bf
+OK
+real    0m0,563s
+user    0m0,558s
+sys     0m0,003s
+simon@gamerarch:~/W/fastbf$ time ./bf < bench.bf
+OK
+real    0m0,562s
+user    0m0,562s
+sys     0m0,000s
+simon@gamerarch:~/W/fastbf$ time ./bf < bench.bf
+OK
+real    0m0,569s
+user    0m0,569s
+sys     0m0,001s
+*/
+
+typedef struct {
+    uint_fast16_t begin;
+    uint_fast16_t depth;
+} LoopEntry;
+
+typedef struct {
+    const char* program;
+    size_t length;
+    uint_fast16_t pc;
+    uint_fast16_t sp;
+    uint_fast16_t loop_depth;
+    uint_fast16_t loops_length;
+    uint_fast16_t loops_progress;
+    LoopEntry loops[1000];
+    uint8_t memory[30000];
+} Context;
+
+void v2_incr(Context* ctx) { ctx->memory[ctx->sp]++; }
+void v2_decr(Context* ctx) { ctx->memory[ctx->sp]--; }
+void v2_right(Context* ctx) { ctx->sp++; }
+void v2_left(Context* ctx) { ctx->sp--; }
+void v2_put(Context* ctx) { putc(ctx->memory[ctx->sp], stdout); }
+void v2_get(Context* ctx) { ctx->memory[ctx->sp] = (uint8_t)getc(stdout); }
+
+void v2_begin(Context* ctx)
+{
+    if (ctx->loops_progress < ctx->pc) {
+        ctx->loops_progress++;
+        ctx->loops[ctx->loops_length] = (LoopEntry) {
+            .begin = ctx->pc,
+            .depth = ctx->loop_depth,
+        };
+        ctx->loops_length++;
+    }
+    ctx->loop_depth++;
+}
+
+void v2_end(Context* ctx)
+{
+    if (ctx->memory[ctx->sp]) {
+        int entry_index = 0;
+        for (int i = (int)ctx->loops_length - 1; i >= 0; i--)
+            if (ctx->loops[i].depth == ctx->loop_depth)
+                entry_index = i;
+        ctx->pc = ctx->loops[entry_index].begin;
+    } else {
+        ctx->loop_depth--;
+    }
+}
+
+void v2_run_instruction(Context* ctx)
+{
+    printf(
+        "pc = %4ld, sp = %4ld, ins = %c\n",
+        ctx->pc,
+        ctx->sp,
+        ctx->program[ctx->pc]
+    );
+    switch (ctx->program[ctx->pc]) {
+        case '+':
+            v2_incr(ctx);
+            break;
+        case '-':
+            v2_decr(ctx);
+            break;
+        case '>':
+            v2_right(ctx);
+            break;
+        case '<':
+            v2_left(ctx);
+            break;
+        case '.':
+            v2_put(ctx);
+            break;
+        case ',':
+            v2_get(ctx);
+            break;
+        case '[':
+            v2_begin(ctx);
+            break;
+        case ']':
+            v2_end(ctx);
+            break;
+    }
+}
+
+void v2(const char* program, size_t length)
+{
+    (void)length;
+    Context ctx = {
+        .program = program,
+        .length = length,
+        .pc = 0,
+        .sp = 0,
+        .loop_depth = 0,
+        .loops_length = 0,
+        .loops_progress = 0,
+    };
+    for (; ctx.pc < length; ctx.pc++) {
+        v2_run_instruction(&ctx);
+    }
+}