This is the mail archive of the
mailing list for the GCC project.
[PATCH] Volatile bitfields vs. inline asm memory constraints
- From: Julian Brown <julian at codesourcery dot com>
- To: gcc-patches at gcc dot gnu dot org
- Cc: DJ Delorie <dj at redhat dot com>
- Date: Mon, 22 Nov 2010 13:28:54 +0000
- Subject: [PATCH] Volatile bitfields vs. inline asm memory constraints
This patch fixes the issue in the (Launchpad, not GCC) bug tracker:
The problem was introduced by the patch from DJ to honour volatile
but not exposed (on ARM) until the option was made the default (on the
Linaro branch) -- it's not yet the default on mainline.
The issue is as follows: after DJ's patch and with
-fstrict-volatile-bitfields, in expr.c:expand_expr_real_1, the if
condition with the comment "In cases where an aligned union has an
unaligned object as a field, we might be extracting a BLKmode value
from an integer-mode (e.g., SImode) object [...]" triggers for a normal
(non-bitfield) volatile field of a struct/class.
But, this appears to be over-eager: in the particular case mentioned
above, when expanding a "volatile int" struct field used as a memory
constraint for an inline asm, we end up with something which is no
longer addressable (I think because of the actions of
extract_bit_field). So, compilation aborts.
My proposed fix is to restrict the conditional by only making it execute
for -fstrict-volatile-bitfields only for non-naturally-aligned accesses:
this appears to work (fixes test in question, and no regressions for
cross to ARM Linux, gcc/g++/libstdc++, with -fstrict-volatile-bitfields
turned on), but I don't know if there will be unintended consequences.
DJ, does it look sane to you?
Incidentally the constraints in the inline asm in the Launchpad
testcase might be slightly dubious (attempting to force (mem (reg)) by
using both "+m" (var) and "r" (&var) constraints), but replacing
them with e.g.:
"ldrex %[newValue], %[_q_value]\n"
"sub %[newValue], %[newValue], #1\n"
"strex %[result], %[newValue], %[_q_value]\n"
"teq %[result], #0\n"
: [newValue] "=&r" (newValue),
[result] "=&r" (result)
: [_q_value] "Q" (_q_value)
: "cc", "memory");
still leads to a warning (not an error) with trunk and
atomic-changed.cc:24:35: warning: use of memory input without lvalue in
asm operand 2 is deprecated [enabled by default]
The warning goes away with the attached patch. So, I don't think the
problem is purely that the original inline asm is invalid.
OK to apply, or any comments?
* expr.c (expand_expr_real_1): Only use BLKmode for volatile
accesses which are not naturally aligned.
--- gcc/expr.c (revision 166945)
+++ gcc/expr.c (working copy)
@@ -9124,7 +9124,8 @@ expand_expr_real_1 (tree exp, rtx target
&& modifier != EXPAND_INITIALIZER)
/* If the field is volatile, we always want an aligned
- || (volatilep && flag_strict_volatile_bitfields > 0)
+ || (volatilep && flag_strict_volatile_bitfields > 0
+ && (bitpos % GET_MODE_ALIGNMENT (mode) != 0))
/* If the field isn't aligned enough to fetch as a memref,
fetch it as a bit field. */
|| (mode1 != BLKmode