I discovered a gcc bug in the -ffloat-store option.
This program:
#include <stdio.h>
int main(void) {
float reale = 1.0f;
int i = 0;
for (; (reale + 1.0f) != 1.0f; i++) {
reale=(float)reale/2.0f;
}
printf("\ni volte = %d\n",i);
}
This program must output 24 instead of 64 if compiled with -ffloat-store.
This other program instead workaround the gcc 2.7.2 bug:
#include <stdio.h>
int main(void)
{
float my_float = 1, result;
int i = 0;
do {
my_float /= 2;
result = my_float + 1;
i++;
} while ((float)(result != 1));
printf("i times = %d\n",i);
}
If I compile this program with gcc 2.8.1 and -ffloat-store and -O, instead
the bug is shown again and I get 64.
I started fixing the 2.7.2 to 2.8.1 inserted bug but gcc sources are a
mess. houndrid of lines per file functions very very long. All in the same
directory. It' s very very difficult to understand the gcc code for me.
The assembler code of the second program generated by gcc-2.8.1 -O2
-ffloat-store -S is this:
-------------------------------------
.file "p1.c"
.version "01.01"
gcc2_compiled.:
.section .rodata
.LC1:
.string "i times = %d\n"
.align 4
.LC0:
.long 0x3f000000
.text
.align 16
.globl main
.type main,@function
main:
pushl %ebp
movl %esp,%ebp
subl $8,%esp
movl $1065353216,-4(%ebp)
xorl %edx,%edx
flds .LC0
fld1
.align 4
.L2:
flds -4(%ebp)
fmul %st(2),%st
fsts -4(%ebp)
fadd %st(1),%st
fsts -8(%ebp)
^^^^^^^^^^^^^
incl %edx
fucomp %st(1)
fnstsw %ax
andb $68,%ah
xorb $64,%ah
jne .L2
fstp %st(0)
fstp %st(0)
pushl %edx
pushl $.LC1
call printf
movl %ebp,%esp
popl %ebp
ret
.Lfe1:
.size main,.Lfe1-main
.ident "GCC: (GNU) 2.8.1"
-------------------------------------
As you can see after the underlined line gcc run a test without reload the
register from memory and so it run the test with a very more precise data.
The assembler patch that fix the bug is this:
===================================================================
RCS file: p1.s,v
retrieving revision 1.1
diff -u -r1.1 p1.s
--- p1.s 1998/06/07 16:01:13 1.1
+++ p1.s 1998/06/07 16:17:42
@@ -25,7 +25,8 @@
fmul %st(2),%st
fsts -4(%ebp)
fadd %st(1),%st
- fsts -8(%ebp)
+ fstps -8(%ebp)
+ flds -8(%ebp)
incl %edx
fucomp %st(1)
fnstsw %ax
Does anybody known very well the gcc code to try to collaborate with me to
fix this bug? The problem should be in the file reload*.c.
I just sent the report to the gcc-bugs mailing-list but I got no answer
except from a guy that said me that he got the same problem in the past.
Andrea[s] Arcangeli
