Le lundi 23 janvier 2012 01:53:21, Daniel Glöckner a écrit :
> On Mon, Jan 23, 2012 at 12:14:58AM +0100, grischka wrote:
> > Thomas Preud'homme wrote:
> > >>Similar bug happens for i386 for example with
> > >>
> > >> double bar(double a, double b, double c, double d);
> > >> double foo (double *p)
> > >> {
> > >>
> > >> return bar(p[1], p[2], p[3], p[4]);
> > >>
> > >> }
> > >>
> > >>which produces
> > >>
> > >> 49: 8b 5d fc mov 0xfffffffc(%ebp),%ebx
> > >> 4c: dd 03 fldl (%ebx)
> > >>
> > >>It should never use %ebx. Hope this helps.
> > >
> > >Are you working on a fix? I looked for some use of ebx in i386-*
> > >and didn't see any reference to ebx or rbx which looked
> > >suspicious. Would it be a wrong construction of an instruction?
> >
> > Good question ;)
>
> Fixed in mob
Great, thanks a lot.
>
> Daniel
Shouldn't the same fix (Cf attached file) be applied for x86-64? As I'm not
sure
about the answer I didn't dare to commit the change.Best regards, Thomas Preud'homme > > _______________________________________________ > Tinycc-devel mailing list > Tinycc-devel@nongnu.org > https://lists.nongnu.org/mailman/listinfo/tinycc-devel
diff --git a/x86_64-gen.c b/x86_64-gen.c
index 4d2521d..0b3db22 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -367,8 +367,10 @@ void load(int r, SValue *sv)
v1.type.t = VT_PTR;
v1.r = VT_LOCAL | VT_LVAL;
v1.c.ul = fc;
- load(r, &v1);
fr = r;
+ if (!(reg_classes[fr] & RC_INT))
+ fr = get_reg(RC_INT);
+ load(fr, &v1);
}
ll = 0;
if ((ft & VT_BTYPE) == VT_FLOAT) {
signature.asc
Description: This is a digitally signed message part.
_______________________________________________ Tinycc-devel mailing list Tinycc-devel@nongnu.org https://lists.nongnu.org/mailman/listinfo/tinycc-devel
