Parser very slow with repeated parse calls

[andrewchambers]

bench.peg:

%prefix "asm"

%earlyheader {
typedef struct {int x; int y; int z;} Parsev;
}

%value "Parsev"

# Uncomment for 10X speed.
# file <- line+

line <- s:stmt eol
      / eol
      / .

stmt <- d:directive 
      / i:instr
      / l:label

directive <- ".glob" "o"? "l" ws i:ident
           / ".data"
           / ".text"
           / ".balign" ws n:number 
           / ".byte" ws n:number

label <- i:ident ':'

instr <- "nop"
       / "leave"
       / "ret"
       / i:jmp
       / i:add

jmp <- "jmp" ws i:ident

add <- "add" 'q'? ws s:m ws? ',' ws? d:r64
     / "add" 'q'? ws s:imm ws? ',' ws? d:r64
     / "add" 'q'? ws s:r64 ws? ',' ws? d:m
     / "add" 'q'? ws s:r64 ws? ',' ws? d:r64
     / "addq" ws s:imm ws? ',' ws? d:m

m <- '(' ws? r:r64 ws? ')'
   / <'-'?[0-9]+> ws? '(' ws? r:r64 ws? ')'
   / i:ident  ws? '(' ws? r:r64 ws? ')'

r64 <- "%rax"
     / "%rcx"
     / "%rdx"
     / "%rbx"
     / "%rsp"
     / "%rbp"
     / "%rsi"
     / "%rdi"
     / "%r8" 
     / "%r9" 
     / "%r10"
     / "%r11"
     / "%r12"
     / "%r13"
     / "%r14"
     / "%r15"

imm <- '$' i:ident
     / '$' <'-'?[0-9]+>

ident <- <[_a-zA-Z][_a-zA-Z0-9]*>

number <- <'-'?[0-9]+>

ws <- [ \t]+

eol <- ws? ("\n" / (! .))

%%
int main() {
    asm_context_t *ctx = asm_create(NULL);
    while (asm_parse(ctx, NULL));
    asm_destroy(ctx);
    return 0;
}

bench.txt:

for i in `seq 100000`;  do  echo "addq %rax, (%rax)"  >> bench.txt ; done

First run, one parse call per line:

$ ./packcc ./bench.peg && clang -O2 -g bench.c -o bench

$ time ./bench  < bench.txt
real    1m22.178s
user    1m21.425s
sys     0m0.428s

Now uncomment the file <- line+ part of the benchmark:

$ ./packcc ./bench.peg && clang -O2 -g bench.c -o bench
$ time ./bench  < bench.txt

real    0m1.221s
user    0m1.026s
sys     0m0.192s

If I benchmark the first case, I find the majority of work is memmove of the internal packcc arrays, my intuition is this extra work does not seem correct.

[arithy]

If I benchmark the first case, I find the majority of work is memmove of the internal packcc arrays, my intuition is this extra work does not seem correct.

You are right. The main cause is frequent call of memmove().
I have pushed the fixed version. Now, the both test cases show almost the same performance.
Please try it.

[arithy]

I solved this issue by 0ce1963.

[andrewchambers]

thank you.