From:	"Evgeni V. Sorokin" <SOROKIN AT ps1 DOT iaee DOT tuwien DOT ac DOT at>
To:	djgpp AT sun DOT soe DOT clarkson DOT edu
Date:	Fri, 30 Dec 1994 15:41:24 MEZ
Subject:	80x86 specific code?
Reply-To:	sorokin AT ps1 DOT iaee DOT tuwien DOT ac DOT at

Happy New Year to everybody!

Optimizing a calculation program, I confronted the following problem:
the innermost loop contained a number of floating-point tests (which
were included in an attempt to reduce the number of operations), like
if (X > 0.0) ...; .
Looking at assembler, I found that this comparison is performed as
    fldz
    fcomp [X]
    fnstsw ax
    and [dummy var],ah
    cmp [another var],ah
    je ...
[translated into INTEL notations - E.S.]

This is a VERY inefficient code, and consumes a lot of time in my
program. Other opportunities include, e.g. considering the highest
word of X as signed integer since the sign of a IEEE float is stored
in its highest bit:
    cmpb [X+sizeof(X)-sizeof(byte)],0
    jl ...
[Equivalent in C should be possible by some ugly type-casting]

Such code is of course purely platform-specific, and ideally should
be generated by the compiler itself.

Before I start playing with non-portable codes, I'd like to ask, if
there is some switch, which will enable generating of such 80x86
specific code? I tried -m486 (which should be in DJGPP by default
anyway) and -O3, without success.

Zhenya.
----
E.Sorokin                                TU Vienna, IAEE
Tel +43(1)58801 -3703, -3948       Gusshausstr. 27/359-9
Fax +43(1)504 2477                A-1040 Vienna, Austria

- Raw text -