🚀 go-pugleaf

Gcc-1.35 generates incorrect assembler code when subtracting an int
from a double. Configuration is:

	Linked `config.h' to `config/xm-i386v.h'.
	Linked `tm.h' to `config/tm-i386gas.h'.
	Linked `md' to `config/i386.md'.
	Linked `aux-output.c' to `config/out-i386.c'.
	Links are now set up for use with a i386g.

OS is ISC 386/ix version 2.0.1 (sysVr3.2). Machine is ITT Xtra 386.

The following file prints -32101.000000, it should print 32101.000000:

main()
{
  double dval;
  int ival;

  dval = 32467.0;
  ival = 366;
  dval -= (double)ival;
  printf("value is %f\n", dval);
}


----------------------
ksh1 [424]$ gcc -v t.c
gcc version 1.35
 /gnu/lib/gcc-cpp -v -undef -D__GNUC__ -Dunix -Di386 -D__unix__ -D__i386__ t.c /tmp/cca07742.cpp
GNU CPP version 1.35
 /gnu/lib/gcc-cc1 /tmp/cca07742.cpp -quiet -dumpbase t.c -version -o /tmp/cca07742.s
GNU C version 1.35 (80386, BSD syntax) compiled by GNU C version 1.35.
 /gnu/lib/gcc-as -o t.o /tmp/cca07742.s
 /gnu/lib/gcc-ld /gnu/lib/crt0.o t.o /gnu/lib/gcc-gnulib -lc
ksh1 [425]$ a.out
value is -32101.000000
ksh1 [426]$ gcc -S t.c
ksh1 [427]$ cat t.s
	.file	"t.c"
gcc_compiled.:
.text
	.align 2
LC0:
	.double 0d3.24670000000000000000e+04
LC1:
	.ascii "value is %f\12\0"
	.align 2
.globl _main
_main:
	pushl %ebp
	movl %esp,%ebp
	subl $12,%esp
	pushl LC0
	popl -8(%ebp)
	pushl LC0+4
	popl -4(%ebp)
	movl $366,-12(%ebp)
	fildl -12(%ebp)		<========
	fldl -8(%ebp)		<========
	fsubp %st,%st(1)
	fstpl -8(%ebp)
	pushl -4(%ebp)
	pushl -8(%ebp)
	pushl $LC1
	call _printf
L1:
	leave
	ret
ksh1 [428]$ 
----------------------
The problem is with the marked lines. The load of the double should
occur before the load of the int.

Bill
--
William G. Bunton            wgb@tntdev.tnt.com     {uunet,natinst}!tntdev!wgb
Tools & Techniques, Inc. Austin, TX