Re: RFC: changing precision control setting in initial FPU context

From: buhr@stat.wisc.edu (Kevin Buhr)
To: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: linux-kernel@vger.kernel.org
Subject: Re: RFC: changing precision control setting in initial FPU context
Date: 03 Mar 2001 18:27:50 -0600	[thread overview]
Message-ID: <vban1b2ped5.fsf@mozart.stat.wisc.edu> (raw)
In-Reply-To: <E14ZLat-0004Js-00@the-village.bc.nu>
In-Reply-To: Alan Cox's message of "Sat, 3 Mar 2001 23:37:45 +0000 (GMT)"

Alan Cox <alan@lxorguk.ukuu.org.uk> writes:
> 
> You want peoples existing applications to suddenely and magically change
> their results. Umm problem.

So, how would you feel about a mechanism whereby the kernel could be
passed a default FPU control word by the binary (with old binaries, by
default, using the old default control word)?  There's already an ELF
AT_FPUCW auxv entry type.  What if this was used by the kernel, rather
than the C library (as it is now), to set a default to be used in
"init_fpu()" when and if the program executed a floating point
instruction?

Then, a compiler startup-code writer would be able to specify a
default control word for binaries that was appropriate for (new)
programs generated by that compiler *WITHOUT* worrying about whether
he was accidentally turning a non-FP program into an FP program by
introducing "fnstcw" as its only FPU instruction.

The C library is already trying to do this (setting the CW based on
the AT_FPUCW vector).  It just can't do it *right* because it doesn't
know if the program is really FP.  It just guesses that if the
AT_FPUCW vector contains something other than the hard-coded
_FPU_DEFAULT (which is supposed to be equal to the kernel default: it
isn't, but it's close enough), it must be set; otherwise, it's left
alone.

Kevin <buhr@stat.wisc.edu>