[v5,08/11] libio: Convert __vasprintf_internal to buffers

  The buffer resizing algorithm is slightly different.  The initial
buffer is on the stack, and small buffers are directly allocated
on the heap using the exact required size.  The overhead of the
additional copy is compensated by the lowered setup cost for buffers
compared to libio streams.
---
 include/printf_buffer.h            |   4 +
 libio/vasprintf.c                  | 141 ++++++++++++++++++++---------
 stdio-common/printf_buffer_flush.c |   4 +
 3 files changed, 104 insertions(+), 45 deletions(-)
  

On 12/12/22 12:23, Florian Weimer via Libc-alpha wrote:
> The buffer resizing algorithm is slightly different.  The initial
> buffer is on the stack, and small buffers are directly allocated
> on the heap using the exact required size.  The overhead of the
> additional copy is compensated by the lowered setup cost for buffers
> compared to libio streams.

Patch look good, some comments below.

> ---
>  include/printf_buffer.h            |   4 +
>  libio/vasprintf.c                  | 141 ++++++++++++++++++++---------
>  stdio-common/printf_buffer_flush.c |   4 +
>  3 files changed, 104 insertions(+), 45 deletions(-)
> 
> diff --git a/include/printf_buffer.h b/include/printf_buffer.h
> index 92393399fb..47e252b52e 100644
> --- a/include/printf_buffer.h
> +++ b/include/printf_buffer.h
> @@ -48,6 +48,7 @@ enum __printf_buffer_mode
>      __printf_buffer_mode_snprintf,
>      __printf_buffer_mode_sprintf_chk,
>      __printf_buffer_mode_to_file,
> +    __printf_buffer_mode_asprintf,
>      __printf_buffer_mode_strfmon,
>      __printf_buffer_mode_fp,         /* For __printf_fp_l_buffer.  */
>      __printf_buffer_mode_fp_to_wide, /* For __wprintf_fp_l_buffer.  */
> @@ -301,6 +302,9 @@ void __printf_buffer_flush_snprintf (struct __printf_buffer_snprintf *)
>  struct __printf_buffer_to_file;
>  void __printf_buffer_flush_to_file (struct __printf_buffer_to_file *)
>    attribute_hidden;
> +struct __printf_buffer_asprintf;
> +void __printf_buffer_flush_asprintf (struct __printf_buffer_asprintf *)
> +  attribute_hidden;
>  struct __printf_buffer_fp;
>  void __printf_buffer_flush_fp (struct __printf_buffer_fp *)
>    attribute_hidden;
> diff --git a/libio/vasprintf.c b/libio/vasprintf.c
> index 4430a266c6..f4ff7ec98b 100644
> --- a/libio/vasprintf.c
> +++ b/libio/vasprintf.c
> @@ -24,64 +24,115 @@
>     This exception applies to code released by its copyright holders
>     in files containing the exception.  */
>  
> -#include <string.h>
> +#include <array_length.h>
> +#include <errno.h>
> +#include <limits.h>
> +#include <math_ldbl_opt.h>
> +#include <printf.h>
> +#include <stdio.h>
>  #include <stdlib.h>
> -#include <strfile.h>
> +#include <string.h>
> +#include <printf_buffer.h>
> +
> +struct __printf_buffer_asprintf
> +{
> +  /* base.write_base points either to a heap-allocated buffer, or to
> +     the direct array below.  */
> +  struct __printf_buffer base;
> +
> +  /* Initial allocation.  200 should be large enough to copy almost
> +     all asprintf usages with just a single (final, correctly sized)
> +     heap allocation.  */
> +  char direct[200];
> +};

There are couple or more buffer sizes scattered in this patchset, maybe it
would be better to consolidate them in one place so it can be easily tuned.

For instance, since it is used with memcpy below, maybe it would be better
to use a size that compiler can inline (it would be arch-dependent, but 
using a size that most usual architecture inline is a net gain).

> +
> +void
> +__printf_buffer_flush_asprintf (struct __printf_buffer_asprintf *buf)
> +{
> +  size_t current_pos = buf->base.write_ptr - buf->base.write_base;
> +  if (current_pos >= INT_MAX)

Shouldn't it be SSIZE_MAX?

> +    {
> +      /* The result is not representable.  No need to continue.  */
> +      __set_errno (EOVERFLOW);
> +      __printf_buffer_mark_failed (&buf->base);
> +      return;
> +    }
> +
> +  size_t current_size = buf->base.write_end - buf->base.write_base;
> +  /* Implement an exponentiatial sizing policy.  Keep the size

s/exponentiatial/exponential

> +     congruent 8 (mod 16), to account for the footer in glibc
> +     malloc.  */
> +  size_t new_size = ((current_size + current_size / 2) & -15) | 8;

Maybe use ALIGN_UP ((current_size + current_size / 2), 16) here?

> +  char *new_buffer;
> +  if (buf->base.write_base == buf->direct)
> +    {
> +      new_buffer = malloc (new_size);
> +      if (new_buffer == NULL)
> +	{
> +	  __printf_buffer_mark_failed (&buf->base);
> +	  return;
> +	}
> +      memcpy (new_buffer, buf->direct, current_pos);
> +    }
> +  else
> +    {
> +      new_buffer = realloc (buf->base.write_base, new_size);
> +      if (new_buffer == NULL)
> +	{
> +	  __printf_buffer_mark_failed (&buf->base);
> +	  return;
> +	}
> +    }
> +
> +  /* Set up the new write area.  */
> +  buf->base.write_base = new_buffer;
> +  buf->base.write_ptr = new_buffer + current_pos;
> +  buf->base.write_end = new_buffer + new_size;
> +}
> +
>  

Ok.

>  int
>  __vasprintf_internal (char **result_ptr, const char *format, va_list args,
>  		      unsigned int mode_flags)
>  {
> -  /* Initial size of the buffer to be used.  Will be doubled each time an
> -     overflow occurs.  */
> -  const size_t init_string_size = 100;
> -  char *string;
> -  _IO_strfile sf;
> -  int ret;
> -  size_t needed;
> -  size_t allocated;
> -  /* No need to clear the memory here (unlike for open_memstream) since
> -     we know we will never seek on the stream.  */
> -  string = (char *) malloc (init_string_size);
> -  if (string == NULL)
> -    return -1;
> -#ifdef _IO_MTSAFE_IO
> -  sf._sbf._f._lock = NULL;
> -#endif
> -  _IO_no_init (&sf._sbf._f, _IO_USER_LOCK, -1, NULL, NULL);
> -  _IO_JUMPS (&sf._sbf) = &_IO_str_jumps;
> -  _IO_str_init_static_internal (&sf, string, init_string_size, string);
> -  sf._sbf._f._flags &= ~_IO_USER_BUF;
> -  sf._s._allocate_buffer_unused = (_IO_alloc_type) malloc;
> -  sf._s._free_buffer_unused = (_IO_free_type) free;
> -  ret = __vfprintf_internal (&sf._sbf._f, format, args, mode_flags);
> -  if (ret < 0)
> +  struct __printf_buffer_asprintf buf;
> +  __printf_buffer_init (&buf.base, buf.direct, array_length (buf.direct),
> +			__printf_buffer_mode_asprintf);
> +
> +  __printf_buffer (&buf.base, format, args, mode_flags);
> +  int done = __printf_buffer_done (&buf.base);
> +  if (done < 0)
>      {
> -      free (sf._sbf._f._IO_buf_base);
> -      return ret;
> +      if (buf.base.write_base != buf.direct)
> +	free (buf.base.write_base);
> +      return done;
> +    }
> +
> +  /* Transfer to the final buffer.  */
> +  char *result;
> +  size_t size = buf.base.write_ptr - buf.base.write_base;
> +  if (buf.base.write_base == buf.direct)
> +    {
> +      result = malloc (size + 1);
> +      if (result == NULL)
> +	return -1;
> +      memcpy (result, buf.direct, size);
>      }
> -  /* Only use realloc if the size we need is of the same (binary)
> -     order of magnitude then the memory we allocated.  */
> -  needed = sf._sbf._f._IO_write_ptr - sf._sbf._f._IO_write_base + 1;
> -  allocated = sf._sbf._f._IO_write_end - sf._sbf._f._IO_write_base;
> -  if ((allocated >> 1) <= needed)
> -    *result_ptr = (char *) realloc (sf._sbf._f._IO_buf_base, needed);
>    else
>      {
> -      *result_ptr = (char *) malloc (needed);
> -      if (*result_ptr != NULL)
> +      result = realloc (buf.base.write_base, size + 1);
> +      if (result == NULL)
>  	{
> -	  memcpy (*result_ptr, sf._sbf._f._IO_buf_base, needed - 1);
> -	  free (sf._sbf._f._IO_buf_base);
> +	  free (buf.base.write_base);
> +	  return -1;
>  	}
> -      else
> -	/* We have no choice, use the buffer we already have.  */
> -	*result_ptr = (char *) realloc (sf._sbf._f._IO_buf_base, needed);
>      }
> -  if (*result_ptr == NULL)
> -    *result_ptr = sf._sbf._f._IO_buf_base;
> -  (*result_ptr)[needed - 1] = '\0';
> -  return ret;
> +
> +  /* Add NUL termination.  */
> +  result[size] = '\0';
> +  *result_ptr = result;
> +
> +  return done;
>  }
>  
>  int

Ok.

> diff --git a/stdio-common/printf_buffer_flush.c b/stdio-common/printf_buffer_flush.c
> index f44c8b887f..14fe1b2df4 100644
> --- a/stdio-common/printf_buffer_flush.c
> +++ b/stdio-common/printf_buffer_flush.c
> @@ -27,6 +27,7 @@
>  #ifndef SHARED
>  # pragma weak __printf_buffer_flush_snprintf
>  # pragma weak __printf_buffer_flush_to_file
> +# pragma weak __printf_buffer_flush_asprintf
>  # pragma weak __printf_buffer_flush_fp
>  # pragma weak __printf_buffer_flush_fp_to_wide
>  # pragma weak __printf_buffer_flush_fphex_to_wide
> @@ -49,6 +50,9 @@ __printf_buffer_do_flush (struct __printf_buffer *buf)
>      case __printf_buffer_mode_to_file:
>        __printf_buffer_flush_to_file ((struct __printf_buffer_to_file *) buf);
>        return;
> +    case __printf_buffer_mode_asprintf:
> +      __printf_buffer_flush_asprintf ((struct __printf_buffer_asprintf *) buf);
> +      return;
>      case __printf_buffer_mode_strfmon:
>        __set_errno (E2BIG);
>        __printf_buffer_mark_failed (buf);
  

* Adhemerval Zanella Netto:

> On 12/12/22 12:23, Florian Weimer via Libc-alpha wrote:
>> The buffer resizing algorithm is slightly different.  The initial
>> buffer is on the stack, and small buffers are directly allocated
>> on the heap using the exact required size.  The overhead of the
>> additional copy is compensated by the lowered setup cost for buffers
>> compared to libio streams.
>
> Patch look good, some comments below.

Thanks.

>> +struct __printf_buffer_asprintf
>> +{
>> +  /* base.write_base points either to a heap-allocated buffer, or to
>> +     the direct array below.  */
>> +  struct __printf_buffer base;
>> +
>> +  /* Initial allocation.  200 should be large enough to copy almost
>> +     all asprintf usages with just a single (final, correctly sized)
>> +     heap allocation.  */
>> +  char direct[200];
>> +};
>
> There are couple or more buffer sizes scattered in this patchset, maybe it
> would be better to consolidate them in one place so it can be easily tuned.

I see what I can do.

> For instance, since it is used with memcpy below, maybe it would be better
> to use a size that compiler can inline (it would be arch-dependent, but 
> using a size that most usual architecture inline is a net gain).

It would need additional optimization hints for GCC, I believe.

>> +void
>> +__printf_buffer_flush_asprintf (struct __printf_buffer_asprintf *buf)
>> +{
>> +  size_t current_pos = buf->base.write_ptr - buf->base.write_base;
>> +  if (current_pos >= INT_MAX)
>
> Shouldn't it be SSIZE_MAX?

Nope, asprintf has the same incorrect return type as printf. 8-(

>
>> +    {
>> +      /* The result is not representable.  No need to continue.  */
>> +      __set_errno (EOVERFLOW);
>> +      __printf_buffer_mark_failed (&buf->base);
>> +      return;
>> +    }
>> +
>> +  size_t current_size = buf->base.write_end - buf->base.write_base;
>> +  /* Implement an exponentiatial sizing policy.  Keep the size
>
> s/exponentiatial/exponential

Fixed.

>> +     congruent 8 (mod 16), to account for the footer in glibc
>> +     malloc.  */
>> +  size_t new_size = ((current_size + current_size / 2) & -15) | 8;
>
> Maybe use ALIGN_UP ((current_size + current_size / 2), 16) here?

Changed.

Thanks,
Florian