RE: bug fix for hp-timing.h (aarch64)

["Tang, Jun" <juntangc@amazon.com>]

Zack,

After rounds of discussion, the suggested solution is the best option and works for all existing benches.  Let me know if I need to submit another request to get this merged

> -   (Diff) = ((End) - (Start)) * (UINT64_C(1000000000) / freq); \
> +   (Diff) = (((End) - (Start)) * UINT64_C(1000000000)) / freq; \

Regards,

Jun

-----Original Message-----
From: Zack Weinberg <zack@owlfolio.org> 
Sent: Wednesday, January 11, 2023 11:22 AM
To: GNU libc development <libc-alpha@sourceware.org>
Subject: Re: bug fix for hp-timing.h (aarch64)

On Wed, Jan 11, 2023, at 11:44 AM, Tang, Jun wrote:
> On Graviton 3 the frequency is 1050000000 and the original function 
> hp-timing.h no longer gives the correct answer. The following patch 
> can fix the bug:

Please try this instead:

> -   (Diff) = ((End) - (Start)) * (UINT64_C(1000000000) / freq); \
> +   (Diff) = (((End) - (Start)) * UINT64_C(1000000000)) / freq; \

This avoids using floating-point arithmetic but should still do the rescale correctly.
(The multiplier and divisor are both guaranteed to be less than 2**32, so, as long as we do the multiplication in 64-bit arithmetic, we cannot lose bits via overflow.)

zw