question about malloc's bin_index

question about malloc's bin_index

[Weiming Zhao]

[-- Attachment #1: Type: text/plain, Size: 760 bytes --]

Hi,

 

I'm trying to use MUSL on a hardware without floating point unit, however,
the "union" in the bin_index brings some issues.

I'm using -mgeneral-regs-only and gcc refuses to compile it due to the
floating point conversion.

 

I'm wondering what's the purpose of the union? How should I rewrite it to
use integer only?

 

static int bin_index(size_t x)

{

                x = x / SIZE_ALIGN - 1;

                if (x <= 32) return x;

                if (x > 0x1c00) return 63;

                return ((union { float v; uint32_t r; }){(int)x}.r>>21) -
496;  

}

 

Thanks a lot,

Weiming

 

Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, 
a Linux Foundation Collaborative Project

 


[-- Attachment #2: Type: text/html, Size: 3856 bytes --]

Re: question about malloc's bin_index

[Rich Felker]

On Wed, Nov 19, 2014 at 06:43:23PM -0800, Weiming Zhao wrote:
> Hi,
> 
>  
> 
> I'm trying to use MUSL on a hardware without floating point unit, however,
> the "union" in the bin_index brings some issues.
> 
> I'm using -mgeneral-regs-only and gcc refuses to compile it due to the
> floating point conversion.

This is odd. It should use soft-float if there's no fpu.

> I'm wondering what's the purpose of the union? How should I rewrite it to
> use integer only?

The value we want, for categorizing sizes into bins, is exactly the
floating point exponent and first 2 bits of the mantissa. It's
basically a sort of base-2 log with 4 linear steps between successive
logarithmic-scale points. I think nsz has an integer-only version of
this code that might be useful; it performs mildly better on archs
without fpu and mildly worse on ones with fpu and it's somewhat larger
(but of course not as large as soft-float code). I'll see if I can
find a copy or if he can post it.

Rich

Re: question about malloc's bin_index

[Szabolcs Nagy]

* Rich Felker <dalias@libc.org> [2014-11-19 23:06:33 -0500]:
> On Wed, Nov 19, 2014 at 06:43:23PM -0800, Weiming Zhao wrote:
> > I'm wondering what's the purpose of the union? How should I rewrite it to
> > use integer only?
> 
> The value we want, for categorizing sizes into bins, is exactly the
> floating point exponent and first 2 bits of the mantissa. It's
> basically a sort of base-2 log with 4 linear steps between successive
> logarithmic-scale points. I think nsz has an integer-only version of
> this code that might be useful; it performs mildly better on archs
> without fpu and mildly worse on ones with fpu and it's somewhat larger
> (but of course not as large as soft-float code). I'll see if I can
> find a copy or if he can post it.

i have this in src/malloc/malloc.c with #ifdef NOFPU:

/* non-float bin index */

static const unsigned char bintab[32]={
	 0, 0, 0, 0,32,33,34,35,36,36,37,37,38,38,39,39,
	40,40,40,40,41,41,41,41,42,42,42,42,43,43,43,43
};

static int bin_index(size_t x)
{
	x = x / SIZE_ALIGN - 1;
	if (x <= 32) return x;
	x /= 8;
	if (x < 32)
		return bintab[x];
	x /= 8;
	if (x < 32)
		return bintab[x]+12;
	x /= 8;
	if (x < 16)
		return bintab[x]+24;
	return 63;
}

static int bin_index_up(size_t x)
{
	x = x / SIZE_ALIGN - 1;
	if (x <= 32) return x;
	x--;
	x /= 8;
	if (x < 32)
		return bintab[x]+1;
	x /= 8;
	if (x < 32)
		return bintab[x]+13;
	x /= 8;
	return bintab[x]+25;
}

Re: question about malloc's bin_index

[weimingz]

I think I need a float-to-int-bits routine. But still I'm not clear why we
need such conversion.

Thanks,
Weiming

> Hi,
>
>
>
> I'm trying to use MUSL on a hardware without floating point unit, however,
> the "union" in the bin_index brings some issues.
>
> I'm using -mgeneral-regs-only and gcc refuses to compile it due to the
> floating point conversion.
>
>
>
> I'm wondering what's the purpose of the union? How should I rewrite it to
> use integer only?
>
>
>
> static int bin_index(size_t x)
>
> {
>
>                 x = x / SIZE_ALIGN - 1;
>
>                 if (x <= 32) return x;
>
>                 if (x > 0x1c00) return 63;
>
>                 return ((union { float v; uint32_t r; }){(int)x}.r>>21) -
> 496;
>
> }
>
>
>
> Thanks a lot,
>
> Weiming
>
>
>
> Qualcomm Innovation Center, Inc.
> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
> a Linux Foundation Collaborative Project
>
>
>
>