donn@sdchema.UUCP (06/04/84)
Subject: f(x) ** 2 produces f(x) * f(x) in f77
Index: usr.bin/f77/src/f77pass1/optim.c 4.2BSD
Description:
When an f77 program is compiled with the optimizer enabled,
side effects of an expression are evaluated twice if that
expression is squared. This was originally pointed out to
me by Jerry Berkman at UC Berkeley.
Repeat-By:
Put the following f77 subroutine in a file fsq.f and compile
it to assembly language with the optimizer turned on:
----------------------------------------------------------------
subroutine fsq( x )
real f, x
x = f( x ) ** 2
return
end
----------------------------------------------------------------
If the bug is present in your compiler you should see the
following (although probably not so pretty):
----------------------------------------------------------------
.globl _fsq_
.set LF1,4
_fsq_:
.word LWM1
subl2 $LF1,sp
jmp L12
L15:
pushl 4(ap)
calls $1,_f_
movl r0,-4(fp)
pushl 4(ap)
calls $1,_f_
mulf2 -4(fp),r0
movl r0,*4(ap)
ret
.align 1
L12:
jmp L15
----------------------------------------------------------------
Fix:
In the code for buffpower() in optim.c, the comment mentions
that the routine was ripped off from putpower() in putpcc.c;
unfortunately the thieves did not also steal the bug fix when
it was installed in putpower()... Sigh. Here is the fix:
----------------------------------------------------------------
*** /tmp/,RCSt1009277 Mon May 28 00:02:50 1984
--- optim.c Fri May 4 21:03:06 1984
***************
*** 632,638
fatal("buffpower: bad call");
base = p->exprblock.leftp;
! if (k == 2)
{
expptr prod;
prod = mkexpr (OPSTAR,cpexpr(base),cpexpr(base));
--- 660,666 -----
fatal("buffpower: bad call");
base = p->exprblock.leftp;
! if (k == 2 && base->tag == TADDR && ISCONST(base->addrblock.memoffset))
{
expptr prod;
prod = mkexpr (OPSTAR,cpexpr(base),cpexpr(base));
----------------------------------------------------------------
Another one bites the dust,
Donn Seeley UCSD Chemistry Dept. ucbvax!sdcsvax!sdchema!donn
32 52' 30"N 117 14' 25"W (619) 452-4016 sdcsvax!sdchema!donn@nosc.ARPA