Date:	Wed, 4 Nov 1998 11:04:40 +0200 (IST)
From:	Eli Zaretskii <eliz AT is DOT elta DOT co DOT il>
X-Sender:	eliz AT is
To:	Kbwms AT aol DOT com
cc:	djgpp-workers AT delorie DOT com
Subject:	Re: Inlining math functions and ANSI/Posix
In-Reply-To:	<abd34bd8.363f3b8d@aol.com>
Message-ID:	<Pine.SUN.3.91.981104110411.25255M-100000@is>
MIME-Version:	1.0
Reply-To:	djgpp-workers AT delorie DOT com

On Tue, 3 Nov 1998 Kbwms AT aol DOT com wrote:

> L10:
>         fldl -8(%ebp)
>         fsqrt
>         fucom %st(0)
>         fnstsw %ax
>         andb $69,%ah
>         cmpb $64,%ah
>         je L11
>         fstp %st(0)
>         pushl -4(%ebp)
>         pushl -8(%ebp)
>         fstpt -32(%ebp)
>         call _sqrt
>         addl $8,%esp
>         fldt -32(%ebp)
>         fxch %st(1)
> 
> I am not familiar with this brand of assembler code but, in the version of
> the assembler code that was optimized, it appears that the compiler checks
> the output of `fsqrt' and calls the library version of `sqrt' if something
> looks amiss.

Clever!  What that code does is this:

	 1. Call FSQRT.
	 2. FUCOM compares ST with the operand, in this case ST(0) is
	    compared to itself.  Depending on the result of FSQRT in
	    ST(0), this sets certain SW_Ci bits in the status word
	    (see below).
	 3. It then retrieves the FPU status word into AX and looks at
	    its high 8 bits.
	 4. AH is ANDed with 45h (69 decimal).  This masks off any
	    bits but SW_C0, SW_C2 and SW_C3 (see the docs of
	    `_status87' in libc.inf).
	 5. Next, if only SW_C3 is set, meaning that ST(0) is equal to
	    itself, everything's okay and it accepts the result of
	    FSQRT.
	 6. If not, this means SW_C0 and SW_C2 are also set, which can
	    only happen if ST(0) cannot be compared with itself, which
	    in turn means ST(0) is either a NaN or an Infinity, and
	    the code then falls back to the library function.

This is all perfectly okay, except that if some of the exceptions are
not masked, sqrt(-1) should trigger an FP exception and crash the
program before all this cleverness has a chance to execute.  v2.02
startup code masks all exceptions, but what if an application unmasks
some of them?  

Can you modify this program to unmask exceptions (by a call to
`_control87') before the call to `sqrt', and see what happens?

If it does crash, I wonder if we should do anything about it.  At
least some comment in the docs seems to be due.

- Raw text -