RE: [Linux-ia64] unalinged access by loadpair instruction

[David Mosberger <davidm@napali.hpl.hp.com>]

>>>>> On Mon, 9 Dec 2002 14:59:30 -0800, "Luck, Tony" <tony.luck@intel.com> said:

  Tony> You do recall correctly ... checking the archives, I posted a
  Tony> patch on October 16, 2001.  I did fix the floating point case
  Tony> back then too ... in fact my post claims to include the patch
  Tony> for both integer and FP.  But I must have attached the wrong
  Tony> patch file.  I've long since deleted all my 2.4.10 trees, so
  Tony> I'd have to reconstruct from scratch (can't do it from memory,
  Tony> those neurons have been re-assigned :-(

OK, it looks like the fix is pretty straight-forward.  The patch below
_should_ work, though I haven't tested it extensively.

Hideki, can you try it out?  BTW: I think your test program is buggy.
The core-loop isn't right because br.ctop renames by one register
position, not two.  I attached a version of the test program which
does what you wanted.

	--david

=== arch/ia64/kernel/unaligned.c 1.6 vs edited ==--- 1.6/arch/ia64/kernel/unaligned.c	Thu Mar 14 00:28:41 2002
+++ edited/arch/ia64/kernel/unaligned.c	Mon Dec  9 18:24:54 2002
@@ -486,7 +486,21 @@
 	DPRINT("*0x%lx=0x%lx NaT=%d new unat: %p=%lx\n", addr, val, nat, (void *) unat,*unat);
 }
 
-#define IA64_FPH_OFFS(r) (r - IA64_FIRST_ROTATING_FR)
+/*
+ * Return the (rotated) index for floating point register REGNUM (REGNUM must be in the
+ * range from 32-127, result is in the range from 0-95.
+ */
+static inline unsigned long
+fph_index (struct pt_regs *regs, long regnum)
+{
+	unsigned long rrb_fr = (regs->cr_ifs >> 25) & 0x7f;
+
+	regnum -= IA64_FIRST_ROTATING_FR;
+	regnum += rrb_fr;
+	if (regnum >= 96)
+		regnum -= 96;
+	return regnum;
+}
 
 static void
 setfpreg (unsigned long regnum, struct ia64_fpreg *fpval, struct pt_regs *regs)
@@ -507,7 +521,7 @@
 	 */
 	if (regnum >= IA64_FIRST_ROTATING_FR) {
 		ia64_sync_fph(current);
-		current->thread.fph[IA64_FPH_OFFS(regnum)] = *fpval;
+		current->thread.fph[fph_index(regs, regnum)] = *fpval;
 	} else {
 		/*
 		 * pt_regs or switch_stack ?
@@ -566,7 +580,7 @@
 	 */
 	if (regnum >= IA64_FIRST_ROTATING_FR) {
 		ia64_flush_fph(current);
-		*fpval = current->thread.fph[IA64_FPH_OFFS(regnum)];
+		*fpval = current->thread.fph[fph_index(regs, regnum)];
 	} else {
 		/*
 		 * f0 = 0.0, f1= 1.0. Those registers are constant and are thus
----------------------------------------------------
#define n 100

double d[n],d2[n+1];

main() {
    int i,j;

    for (i = 0; i < n; i++)  {
	    d[i] = i;
	    d2[i] = 0.0;
    }
    copy_by_loadpair(&d, &d2, n/2-1);
    for (i = 0; i < n; i++) {
	    if (d2[i] != i)
		    printf("d2[%d] = %f, should be d[%d]=%f\n",
			   i, d2[i], i, d[i]);
    }
}

----------------------------------------------------
	.file	"a.c"
	.pred.safe_across_calls p1-p5,p16-p63
.text
	.align 16
	.global copy_by_loadpair
	.proc copy_by_loadpair
copy_by_loadpair:
        alloc   r8=ar.pfs,3,6,0,0 ;;
        mov     r15=r32
        mov     r2=r33
        add     r3=8,r33
        mov     ar.lc=r34
        mov     pr.rot=0x10000
        mov     ar.ec=5 ;;
L1:
  (p16) ldfpd   f32,f37=[r15],16
  (p20) stfd    [r2]ó6,16
  (p20) stfd    [r3]ô1,16
        br.ctop.sptk    L1;;
        br.ret.sptk.many        b0 ;;
	.endp get_by_loadpair