Re: [PATCH v3 04/18] Add string vectorized find and detection functions

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 12/01/2018 15:08, Joseph Myers wrote:
> On Thu, 11 Jan 2018, Paul Eggert wrote:
> 
>> On 01/11/2018 10:54 AM, Adhemerval Zanella wrote:
>>>> The Gnulib integer_length module has a faster implementation, at least for
>>>> 32-bit platforms. Do we still care about 32-bit platforms? If so, you
>>>> might want to take a look at it.
>>> Do you mean the version which uses double to integer, the one with 6
>>> comparisons
>>> or the naive one?
>>
>> I meant the one that converts int to double. It can be branchless since we
>> assume the int is nonzero.
> 
> Looking at glibc architectures (and architectures with recently proposed 
> ports):
> 
> * The following have clz patterns in GCC, unconditionally, meaning this 
> glibc patch will always use __builtin_clz functions and any fallback code 
> is irrelevant: aarch64 i386 ia64 powerpc tilegx x86_64.  (On ia64 the 
> pattern uses conversion to floating point.)
> 
> * The following have clz patterns in GCC, conditionally: alpha arm m68k 
> microblaze mips s390 sparc (and arc).  I have not checked whether in some 
> of those cases the conditions might in fact be true for every 
> configuration for which glibc can be built.
> 
> * The following lack clz patterns in GCC: hppa nios2 sh (and riscv).
> 
> If the configuration lacking clz is also soft-float, converting int to 
> double is an extremely inefficient way ending up calling the libgcc clz 
> implementation (both soft-fp and fp-bit use __builtin_clz).  I think 
> that's sufficient reason to avoid an approach involving conversion to 
> double unless an architecture has opted in to using it as an efficient 
> approach on that architecture.

Thanks for remind about soft-float, also for some architectures that does
have hardware floating pointer units the int to/from float is also a costly
operation.

Regarding index_{first,last}_ fallback implementation, maybe simpler 
implementation which just check the mask bits instead of fallback ones for
leading/trailing zero bit should better, I am open to suggestions here.

> 
> (For arm, for example, clz is supported if "TARGET_32BIT && arm_arch5", so 
> the only configurations without __builtin_clz expanded inline by the 
> compiler are armv4t ones - which are also all soft-float, so the expansion 
> using double can never make sense for arm.)
>