[PATCH] ia64: fix ld.a emulation

This patch fixes a corner-case of ld.a emulation.  ld.a should be
emulated such that it always loads a misaligned value and clears the
corresponding ALAT entry.  The old emulation was correct for the case
where ld.a was followed by ld.c/chk.a (since ALAT entry was cleared),
but wrong for the case where it _wasn't_ followed by ld.c/chk.a.  In
that case, the misaligned value wasn't read from memory, as it should
have been.

arch/ia64/kernel/unaligned.c

@@ -740,6 +740,7 @@ static int
 emulate_load_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs)
 {
 	unsigned int len = 1 << ld.x6_sz;
+	unsigned long val = 0;
 
 	/*
 	 * r0, as target, doesn't need to be checked because Illegal Instruction
@@ -750,21 +751,18 @@ emulate_load_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs)
 	 */
 
 	/*
-	 * ldX.a we don't try to emulate anything but we must invalidate the ALAT entry.
+	 * ldX.a we will emulate load and also invalidate the ALAT entry.
 	 * See comment below for explanation on how we handle ldX.a
 	 */
-	if (ld.x6_op != 0x2) {
-		unsigned long val = 0;
 
-		if (len != 2 && len != 4 && len != 8) {
-			DPRINT("unknown size: x6=%d\n", ld.x6_sz);
-			return -1;
-		}
-		/* this assumes little-endian byte-order: */
-		if (copy_from_user(&val, (void *) ifa, len))
-		    return -1;
-		setreg(ld.r1, val, 0, regs);
+	if (len != 2 && len != 4 && len != 8) {
+		DPRINT("unknown size: x6=%d\n", ld.x6_sz);
+		return -1;
 	}
+	/* this assumes little-endian byte-order: */
+	if (copy_from_user(&val, (void *) ifa, len))
+		return -1;
+	setreg(ld.r1, val, 0, regs);
 
 	/*
 	 * check for updates on any kind of loads
@@ -817,7 +815,7 @@ emulate_load_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs)
 	 *		store & shift to temporary;
 	 *		r1=temporary
 	 *
-	 *	  So int this case, you would get the right value is r1 but the wrong info in
+	 *	  So in this case, you would get the right value is r1 but the wrong info in
 	 *	  the ALAT.  Notice that you could do it in reverse to finish with address 3
 	 *	  but you would still get the size wrong.  To get the size right, one needs to
 	 *	  execute exactly the same kind of load. You could do it from a aligned
@@ -826,9 +824,12 @@ emulate_load_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs)
 	 *	  So no matter what, it is not possible to emulate an advanced load
 	 *	  correctly. But is that really critical ?
 	 *
+	 *	  We will always convert ld.a into a normal load with ALAT invalidated.  This
+	 *	  will enable compiler to do optimization where certain code path after ld.a
+	 *	  is not required to have ld.c/chk.a, e.g., code path with no intervening stores.
 	 *
-	 *	  Now one has to look at how ld.a is used, one must either do a ld.c.* or
-	 *	  chck.a.* to reuse the value stored in the ALAT. Both can "fail" (meaning no
+	 *	  If there is a store after the advanced load, one must either do a ld.c.* or
+	 *	  chk.a.* to reuse the value stored in the ALAT. Both can "fail" (meaning no
 	 *	  entry found in ALAT), and that's perfectly ok because:
 	 *
 	 *		- ld.c.*, if the entry is not present a  normal load is executed
@@ -836,19 +837,8 @@ emulate_load_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs)
 	 *
 	 *	  In either case, the load can be potentially retried in another form.
 	 *
-	 *	  So it's okay NOT to do any actual load on an unaligned ld.a. However the ALAT
-	 *	  must be invalidated for the register (so that's chck.a.*,ld.c.* don't pick up
-	 *	  a stale entry later) The register base update MUST also be performed.
-	 *
-	 *	  Now what is the content of the register and its NaT bit in the case we don't
-	 *	  do the load ?  EAS2.4, says (in case an actual load is needed)
-	 *
-	 *		- r1 = [r3], Nat = 0 if succeeds
-	 *		- r1 = 0 Nat = 0 if trying to access non-speculative memory
-	 *
-	 *	  For us, there is nothing to do, because both ld.c.* and chk.a.* are going to
-	 *	  retry and thus eventually reload the register thereby changing Nat and
-	 *	  register content.
+	 *	  ALAT must be invalidated for the register (so that chk.a or ld.c don't pick
+	 *	  up a stale entry later). The register base update MUST also be performed.
 	 */
 
 	/*