On Tue, Nov 25, 2014 at 7:04 AM, Richard Biener
<richard.guent...@gmail.com> wrote:
> On Tue, Nov 25, 2014 at 4:01 PM, Richard Biener
> <richard.guent...@gmail.com> wrote:
>> On Tue, Nov 25, 2014 at 1:57 PM, H.J. Lu <hongjiu...@intel.com> wrote:
>>> Hi,
>>>
>>> The enclosed testcase fails on x86 when compiled with -Os since we pass
>>> a byte parameter with a byte load in caller and read it as an int in
>>> callee. The reason it only shows up with -Os is x86 backend encodes
>>> a byte load with an int load if -O isn't used. When a byte load is
>>> used, the upper 24 bits of the register have random value for none
>>> WORD_REGISTER_OPERATIONS targets.
>>>
>>> It happens because setup_incoming_promotions in combine.c has
>>>
>>> /* The mode and signedness of the argument before any promotions
>>> happen
>>> (equal to the mode of the pseudo holding it at that stage). */
>>> mode1 = TYPE_MODE (TREE_TYPE (arg));
>>> uns1 = TYPE_UNSIGNED (TREE_TYPE (arg));
>>>
>>> /* The mode and signedness of the argument after any source language
>>> and
>>> TARGET_PROMOTE_PROTOTYPES-driven promotions. */
>>> mode2 = TYPE_MODE (DECL_ARG_TYPE (arg));
>>> uns3 = TYPE_UNSIGNED (DECL_ARG_TYPE (arg));
>>>
>>> /* The mode and signedness of the argument as it is actually passed,
>>> after any TARGET_PROMOTE_FUNCTION_ARGS-driven ABI promotions. */
>>> mode3 = promote_function_mode (DECL_ARG_TYPE (arg), mode2, &uns3,
>>> TREE_TYPE (cfun->decl), 0);
>>>
>>> while they are actually passed in register by assign_parm_setup_reg in
>>> function.c:
>>>
>>> /* Store the parm in a pseudoregister during the function, but we may
>>> need to do it in a wider mode. Using 2 here makes the result
>>> consistent with promote_decl_mode and thus expand_expr_real_1. */
>>> promoted_nominal_mode
>>> = promote_function_mode (data->nominal_type, data->nominal_mode,
>>> &unsignedp,
>>> TREE_TYPE (current_function_decl), 2);
>>>
>>> where nominal_type and nominal_mode are set up with TREE_TYPE (parm)
>>> and TYPE_MODE (nominal_type). TREE_TYPE here is
>>
>> I think the bug is here, not in combine.c. Can you try going back in history
>> for both snippets and see if they matched at some point?
>
> Oh, and note that I think DECL_ARG_TYPE is sth dangerous - it's meant
> to be a source language "ABI" kind-of-thing. Or rather an optimization
> hit. For example in C when integral promotions happen to call arguments
> this can be used to optimize sign-/zero-extensions in the callee. Unless
> something else overrides this (like the target which specifies the real ABI).
> IIRC.
>
PROMOTE_MODE is a performance hint, not an ABI requirement.
i386.h has
#define PROMOTE_MODE(MODE, UNSIGNEDP, TYPE) \
do { \
if (((MODE) == HImode && TARGET_PROMOTE_HI_REGS) \
|| ((MODE) == QImode && TARGET_PROMOTE_QI_REGS)) \
(MODE) = SImode; \
} while (0)
We may promote QI/HI to SI, depending on optimization.
On the other hand, TARGET_PROMOTE_FUNCTION_MODE is
determined by psABI.
I am enclosing the missing ChangeLog entries.
--
H.J.
---
gcc/
PR rtl-optimization/64037
* combine.c (setup_incoming_promotions): Pass the argument
before any promotions happen to promote_function_mode.
gcc/testsuite/
PR rtl-optimization/64037
* g++.dg/pr64037.C: New test.
