[PATCHv4] Fix not 8-byte aligned ldrd/strd on ARMv5 (PR 89544)

Bernd Edlinger bernd.edlinger@hotmail.de
Thu Aug 15 12:38:00 GMT 2019


On 8/15/19 10:55 AM, Richard Biener wrote:
> On Wed, 14 Aug 2019, Bernd Edlinger wrote:
> 
>> On 8/14/19 2:00 PM, Richard Biener wrote:
>>
>> Well, yes, but I was scared away by the complexity of emit_move_insn_1.
>>
>> It could be done, but in the moment I would be happy to have these
>> checks of one major strict alignment target, ARM is a good candidate
>> since most instructions work even if they are accidentally
>> using unaligned arguments.  So middle-end errors do not always
>> visible by ordinary tests.  Nevertheless it is a blatant violation of the
>> contract between middle-end and back-end, which should be avoided.
> 
> Fair enough.
> 
>>>> Several struct-layout-1.dg testcase tripped over misaligned
>>>> complex_cst constants, fixed by varasm.c (align_variable).
>>>> This is likely a wrong code bug, because misaligned complex
>>>> constants, are expanded to misaligned MEM_REF, but the
>>>> expansion cannot handle misaligned constants, only packed
>>>> structure fields.
>>>
>>> Hmm.  So your patch overrides user-alignment here.  Woudln't it
>>> be better to do that more conciously by
>>>
>>>   if (! DECL_USER_ALIGN (decl)
>>>       || (align < GET_MODE_ALIGNMENT (DECL_MODE (decl))
>>>           && targetm.slow_unaligned_access (DECL_MODE (decl), align)))
>>>

? I don't know why that would be better?
If the value is underaligned no matter why, pretend it was declared as
naturally aligned if that causes wrong code otherwise.
That was the idea here.

>>> ?  And why is the movmisalign optab support missing here?
>>>
>>
>> Yes, I wanted to replicate what we have in assign_parm_adjust_stack_rtl:
>>
>>   /* If we can't trust the parm stack slot to be aligned enough for its
>>      ultimate type, don't use that slot after entry.  We'll make another
>>      stack slot, if we need one.  */
>>   if (stack_parm
>>       && ((GET_MODE_ALIGNMENT (data->nominal_mode) > MEM_ALIGN (stack_parm)
>>            && targetm.slow_unaligned_access (data->nominal_mode,
>>                                              MEM_ALIGN (stack_parm)))
>>
>> which also makes a variable more aligned than it is declared.
>> But maybe both should also check the movmisalign optab in
>> addition to slow_unaligned_access ?
> 
> Quite possible.
> 

Will do, see attached new version of the patch.

>>> IMHO whatever code later fails to properly use unaligned loads
>>> should be fixed instead rather than ignoring user requested alignment.
>>>
>>> Can you quote a short testcase that explains what exactly goes wrong?
>>> The struct-layout ones are awkward to look at...
>>>
>>
>> Sure,
>>
>> $ cat test.c
>> _Complex float __attribute__((aligned(1))) cf;
>>
>> void foo (void)
>> {
>>   cf = 1.0i;
>> }
>>
>> $ arm-linux-gnueabihf-gcc -S test.c 
>> during RTL pass: expand
>> test.c: In function 'foo':
>> test.c:5:6: internal compiler error: in gen_movsf, at config/arm/arm.md:7003
>>     5 |   cf = 1.0i;
>>       |   ~~~^~~~~~
>> 0x7ba475 gen_movsf(rtx_def*, rtx_def*)
>> 	../../gcc-trunk/gcc/config/arm/arm.md:7003
>> 0xa49587 insn_gen_fn::operator()(rtx_def*, rtx_def*) const
>> 	../../gcc-trunk/gcc/recog.h:318
>> 0xa49587 emit_move_insn_1(rtx_def*, rtx_def*)
>> 	../../gcc-trunk/gcc/expr.c:3695
>> 0xa49914 emit_move_insn(rtx_def*, rtx_def*)
>> 	../../gcc-trunk/gcc/expr.c:3791
>> 0xa494f7 emit_move_complex_parts(rtx_def*, rtx_def*)
>> 	../../gcc-trunk/gcc/expr.c:3490
>> 0xa49914 emit_move_insn(rtx_def*, rtx_def*)
>> 	../../gcc-trunk/gcc/expr.c:3791
>> 0xa5106f store_expr(tree_node*, rtx_def*, int, bool, bool)
>> 	../../gcc-trunk/gcc/expr.c:5855
>> 0xa51cc0 expand_assignment(tree_node*, tree_node*, bool)
>> 	../../gcc-trunk/gcc/expr.c:5441
> 
> Huh, so why didn't it trigger
> 
>   /* Handle misaligned stores.  */
>   mode = TYPE_MODE (TREE_TYPE (to));
>   if ((TREE_CODE (to) == MEM_REF
>        || TREE_CODE (to) == TARGET_MEM_REF)
>       && mode != BLKmode
>       && !mem_ref_refers_to_non_mem_p (to)
>       && ((align = get_object_alignment (to))
>           < GET_MODE_ALIGNMENT (mode))
>       && (((icode = optab_handler (movmisalign_optab, mode))
>            != CODE_FOR_nothing)
>           || targetm.slow_unaligned_access (mode, align)))
>     {
> 
> ?  (_Complex float is 32bit aligned it seems, the DECL_RTL for the
> var is (mem/c:SC (symbol_ref:SI ("cf") [flags 0x2] <var_decl 
> 0x2aaaaaad1240 cf>) [1 cf+0 S8 A8]), SCmode is 32bit aligned.
> 
> Ah, 'to' is a plain DECL here so the above handling is incomplete.
> IIRC component refs like __real cf = 0.f should be handled fine
> again(?).  So, does adding || DECL_P (to) fix the case as well?
> 

So I tried this instead of the varasm.c change:

Index: expr.c
===================================================================
--- expr.c	(revision 274487)
+++ expr.c	(working copy)
@@ -5002,9 +5002,10 @@ expand_assignment (tree to, tree from, bool nontem
   /* Handle misaligned stores.  */
   mode = TYPE_MODE (TREE_TYPE (to));
   if ((TREE_CODE (to) == MEM_REF
-       || TREE_CODE (to) == TARGET_MEM_REF)
+       || TREE_CODE (to) == TARGET_MEM_REF
+       || DECL_P (to))
       && mode != BLKmode
-      && !mem_ref_refers_to_non_mem_p (to)
+      && (DECL_P (to) || !mem_ref_refers_to_non_mem_p (to))
       && ((align = get_object_alignment (to))
 	  < GET_MODE_ALIGNMENT (mode))
       && (((icode = optab_handler (movmisalign_optab, mode))

Result, yes, it fixes this test case
but then I run all struct-layout-1.exp there are sill cases. where we have problems:

In file included from /home/ed/gnu/gcc-build-arm-linux-gnueabihf-linux64/gcc/testsuite/gcc/gcc.dg-struct-layout-1//t024_x.c:8:^M
/home/ed/gnu/gcc-build-arm-linux-gnueabihf-linux64/gcc/testsuite/gcc/gcc.dg-struct-layout-1//t024_test.h: In function 'test2112':^M
/home/ed/gnu/gcc-trunk/gcc/testsuite/gcc.dg/compat/struct-layout-1_x1.h:23:10: internal compiler error: in gen_movdf, at config/arm/arm.md:7107^M
/home/ed/gnu/gcc-trunk/gcc/testsuite/gcc.dg/compat/struct-layout-1_x1.h:62:3: note: in definition of macro 'TX'^M
/home/ed/gnu/gcc-build-arm-linux-gnueabihf-linux64/gcc/testsuite/gcc/gcc.dg-struct-layout-1//t024_test.h:113:1: note: in expansion of macro 'TCI'^M
/home/ed/gnu/gcc-build-arm-linux-gnueabihf-linux64/gcc/testsuite/gcc/gcc.dg-struct-layout-1//t024_test.h:113:294: note: in expansion of macro 'F'^M
0x7ba377 gen_movdf(rtx_def*, rtx_def*)^M
        ../../gcc-trunk/gcc/config/arm/arm.md:7107^M
0xa494c7 insn_gen_fn::operator()(rtx_def*, rtx_def*) const^M
        ../../gcc-trunk/gcc/recog.h:318^M
0xa494c7 emit_move_insn_1(rtx_def*, rtx_def*)^M
        ../../gcc-trunk/gcc/expr.c:3695^M
0xa49854 emit_move_insn(rtx_def*, rtx_def*)^M
        ../../gcc-trunk/gcc/expr.c:3791^M
0xa49437 emit_move_complex_parts(rtx_def*, rtx_def*)^M
        ../../gcc-trunk/gcc/expr.c:3490^M
0xa49854 emit_move_insn(rtx_def*, rtx_def*)^M
        ../../gcc-trunk/gcc/expr.c:3791^M
0xa50faf store_expr(tree_node*, rtx_def*, int, bool, bool)^M
        ../../gcc-trunk/gcc/expr.c:5856^M
0xa51f34 expand_assignment(tree_node*, tree_node*, bool)^M
        ../../gcc-trunk/gcc/expr.c:5302^M
0xa51f34 expand_assignment(tree_node*, tree_node*, bool)^M
        ../../gcc-trunk/gcc/expr.c:4983^M
0x9338af expand_gimple_stmt_1^M
        ../../gcc-trunk/gcc/cfgexpand.c:3777^M
0x9338af expand_gimple_stmt^M
        ../../gcc-trunk/gcc/cfgexpand.c:3875^M
0x939221 expand_gimple_basic_block^M
        ../../gcc-trunk/gcc/cfgexpand.c:5915^M
0x93af86 execute^M
        ../../gcc-trunk/gcc/cfgexpand.c:6538^M
Please submit a full bug report,^M

My personal gut feeling this will be more fragile than over-aligning the
constants.



>> 0xa51cc0 expand_assignment(tree_node*, tree_node*, bool)
>> 	../../gcc-trunk/gcc/expr.c:4983
>> 0x93396f expand_gimple_stmt_1
>> 	../../gcc-trunk/gcc/cfgexpand.c:3777
>> 0x93396f expand_gimple_stmt
>> 	../../gcc-trunk/gcc/cfgexpand.c:3875
>> 0x9392e1 expand_gimple_basic_block
>> 	../../gcc-trunk/gcc/cfgexpand.c:5915
>> 0x93b046 execute
>> 	../../gcc-trunk/gcc/cfgexpand.c:6538
>> Please submit a full bug report,
>> with preprocessed source if appropriate.
>> Please include the complete backtrace with any bug report.
>> See <https://gcc.gnu.org/bugs/> for instructions.
>>
>> Without the hunk in varasm.c of course.
>>
>> What happens is that expand_expr_real_2 returns a unaligned mem_ref here:
>>
>>     case COMPLEX_CST:
>>       /* Handle evaluating a complex constant in a CONCAT target.  */
>>       if (original_target && GET_CODE (original_target) == CONCAT)
>>         {
>>           [... this path not taken ...]

BTW: this code block executes when the other ICE happens.
 
>>         }
>>
>>       /* fall through */
>>
>>     case STRING_CST:
>>       temp = expand_expr_constant (exp, 1, modifier);
>>
>>       /* temp contains a constant address.
>>          On RISC machines where a constant address isn't valid,
>>          make some insns to get that address into a register.  */
>>       if (modifier != EXPAND_CONST_ADDRESS
>>           && modifier != EXPAND_INITIALIZER
>>           && modifier != EXPAND_SUM
>>           && ! memory_address_addr_space_p (mode, XEXP (temp, 0),
>>                                             MEM_ADDR_SPACE (temp)))
>>         return replace_equiv_address (temp,
>>                                       copy_rtx (XEXP (temp, 0)));
>>       return temp;
>>
>> The result of expand_expr_real(..., EXPAND_NORMAL) ought to be usable
>> by emit_move_insn, that is expected just *everywhere* and can't be changed.
>>
>> This could probably be fixed in an ugly way in the COMPLEX_CST, handler
>> but OTOH, I don't see any reason why this constant has to be misaligned
>> when it can be easily aligned, which avoids the need for a misaligned access.
> 
> If the COMPLEX_CST happends to end up in unaligned memory then that's
> of course a bug (unless the target requests that for all COMPLEX_CSTs).
> That is, if the unalignment is triggered because the store is to an
> unaligned decl.
> 
> But I think the issue is the above one?
> 

yes initially the constant seems to be unaligned. then it is expanded,
and there is no special handling for unaligned constants in expand_expr_real,
and then probably expand_assignment or store_expr seem not fully prepared for
this either.

>>>> Furthermore gcc.dg/Warray-bounds-33.c was fixed by the
>>>> change in expr.c (expand_expr_real_1).  Certainly is it invalid
>>>> to read memory at a function address, but it should not ICE.
>>>> The problem here, is the MEM_REF has no valid MEM_ALIGN, it looks
>>>> like A32, so the misaligned code execution is not taken, but it is
>>>> set to A8 below, but then we hit an ICE if the result is used:
>>>
>>> So the user accessed it as A32.
>>>
>>>>         /* Don't set memory attributes if the base expression is
>>>>            SSA_NAME that got expanded as a MEM.  In that case, we should
>>>>            just honor its original memory attributes.  */
>>>>         if (TREE_CODE (tem) != SSA_NAME || !MEM_P (orig_op0))
>>>>           set_mem_attributes (op0, exp, 0);
>>>
>>> Huh, I don't understand this.  'tem' should never be SSA_NAME.
>>
>> tem is the result of get_inner_reference, why can't that be a SSA_NAME ?
> 
> We can't subset an SSA_NAME.  I have really no idea what this intended
> to do...
> 

Nice, so would you do a patch to change that to a
gcc_checking_assert (TREE_CODE (tem) != SSA_NAME) ?
maybe with a small explanation?

>>> But set_mem_attributes_minus_bitpos uses get_object_alignment_1
>>> and that has special treatment for FUNCTION_DECLs that is not
>>> covered by
>>>
>>>       /* When EXP is an actual memory reference then we can use
>>>          TYPE_ALIGN of a pointer indirection to derive alignment.
>>>          Do so only if get_pointer_alignment_1 did not reveal absolute
>>>          alignment knowledge and if using that alignment would
>>>          improve the situation.  */
>>>       unsigned int talign;
>>>       if (!addr_p && !known_alignment
>>>           && (talign = min_align_of_type (TREE_TYPE (exp)) * 
>>> BITS_PER_UNIT)
>>>           && talign > align)
>>>         align = talign;
>>>
>>> which could be moved out of the if-cascade.
>>>
>>> That said, setting A8 should eventually result into appropriate
>>> unaligned expansion, so it seems odd this triggers the assert...
>>>
>>
>> The function pointer is really 32-byte aligned in ARM mode to start
>> with...
>>
>> The problem is that the code that handles this misaligned access
>> is skipped because the mem_rtx has initially no MEM_ATTRS and therefore
>> MEM_ALIGN == 32, and therefore the code that handles the unaligned
>> access is not taken.  BUT before the mem_rtx is returned it is
>> set to MEM_ALIGN = 8 by set_mem_attributes, and we have an assertion,
>> because the result from expand_expr_real(..., EXPAND_NORMAL) ought to be
>> usable with emit_move_insn.
> 
> yes, as said the _access_ determines the address should be aligned
> so we shouldn't end up setting MEM_ALIGN to 8 but to 32 according
> to the access type/mode.  But we can't trust DECL_ALIGN of
> FUNCTION_DECLs but we _can_ trust users writing *(int *)fn
> (maybe for actual accesses we _can_ trust DECL_ALIGN, it's just
> we may not compute nonzero bits for the actual address because
> of function pointer mangling)
> (for accessing function code I'd say this would be premature
> optimization, but ...)
> 

Not a very nice solution, but it is not worth to spend much effort
in optimizing undefined behavior, I just want to avoid the ICE
at this time and would not trust the DECL_ALIGN either.

>>>>
>>>> Finally gcc.dg/torture/pr48493.c required the change
>>>> in assign_parm_setup_stack.  This is just not using the
>>>> correct MEM_ALIGN attribute value, while the memory is
>>>> actually aligned.
>>>
>>> But doesn't
>>>
>>>           int align = STACK_SLOT_ALIGNMENT (data->passed_type,
>>>                                             GET_MODE (data->entry_parm),
>>>                                             TYPE_ALIGN 
>>> (data->passed_type));
>>> +         if (align < (int)GET_MODE_ALIGNMENT (GET_MODE 
>>> (data->entry_parm))
>>> +             && targetm.slow_unaligned_access (GET_MODE 
>>> (data->entry_parm),
>>> +                                               align))
>>> +           align = GET_MODE_ALIGNMENT (GET_MODE (data->entry_parm));
>>>
>>> hint at that STACK_SLOT_ALIGNMENT is simply bogus for the target?
>>> That is, the target says, for natural alignment 64 the stack slot
>>> alignment can only be guaranteed 32.  You can't then simply up it
>>> but have to use unaligned accesses (or the target/middle-end needs
>>> to do dynamic stack alignment).
>>>
>> Yes, maybe, but STACK_SLOT_ALIGNMENT is used in a few other places as well,
>> and none of them have a problem, probably because they use expand_expr,
>> but here we use emit_move_insn:
>>
>>       if (MEM_P (src))
>>         {
>>           [...]
>>         }
>>       else
>>         {
>>           if (!REG_P (src))
>>             src = force_reg (GET_MODE (src), src);
>>           emit_move_insn (dest, src);
>>         }
>>
>> So I could restrict that to
>>
>>           if (!MEM_P (data->entry_parm)
>>               && align < (int)GET_MODE_ALIGNMENT (GET_MODE (data->entry_parm))
>>               && ((optab_handler (movmisalign_optab,
>> 				  GET_MODE (data->entry_parm))
>>                    != CODE_FOR_nothing)
>>                   || targetm.slow_unaligned_access (GET_MODE (data->entry_parm),
>>                                                     align)))
>>             align = GET_MODE_ALIGNMENT (GET_MODE (data->entry_parm));
>>
>> But OTOH even for arguments arriving in unaligned stack slots where
>> emit_block_move could handle it, that would just work against the
>> intention of assign_parm_adjust_stack_rtl.
>>
>> Of course there are limits how much alignment assign_stack_local
>> can handle, and that would result in an assertion in the emit_move_insn.
>> But in the end if that happens it is just an impossible target
>> configuration.
> 
> Still I think you can't simply override STACK_SLOT_ALIGNMENT just because
> of the mode of an entry param, can you?  If you can assume a bigger
> alignment then STACK_SLOT_ALIGNMENT should return it.
> 

I don't see a real problem here.  All target except i386 and gcn (whatever that is)
use the default for STACK_SLOT_ALIGNMENT which simply allows any (large) align value
to rule the effective STACK_SLOT_ALIGNMENT.  The user could have simply declared
the local variable with the alignment that results in better code FWIW.

If the stack alignment is too high that is capped in assign_stack_local:

  /* Ignore alignment if it exceeds MAX_SUPPORTED_STACK_ALIGNMENT.  */
  if (alignment_in_bits > MAX_SUPPORTED_STACK_ALIGNMENT)
    {
      alignment_in_bits = MAX_SUPPORTED_STACK_ALIGNMENT;
      alignment = MAX_SUPPORTED_STACK_ALIGNMENT / BITS_PER_UNIT;
    }

I for one, would just assume that MAX_SUPPORTED_STACK_ALIGNMENT should
be sufficient for all modes that need movmisalign_optab and friends.
If it is not, an ICE would be just fine.

>>>
>>>>  Note that set_mem_attributes does not
>>>> always preserve the MEM_ALIGN of the ref, since:
>>>
>>> set_mem_attributes sets _all_ attributes from an expression or type.
>>>
>>
>> Not really:
>>
>>   refattrs = MEM_ATTRS (ref);
>>   if (refattrs)
>>     {
>>       /* ??? Can this ever happen?  Calling this routine on a MEM that
>>          already carries memory attributes should probably be invalid.  */
>>       [...]
>>       attrs.align = refattrs->align;
>>     }
>>   else
>>     [...]
>>
>>   if (objectp || TREE_CODE (t) == INDIRECT_REF)
>>     attrs.align = MAX (attrs.align, TYPE_ALIGN (type));
>>
>>>>   /* Default values from pre-existing memory attributes if present.  */
>>>>   refattrs = MEM_ATTRS (ref);
>>>>   if (refattrs)
>>>>     {
>>>>       /* ??? Can this ever happen?  Calling this routine on a MEM that
>>>>          already carries memory attributes should probably be invalid.  */
>>>>       attrs.expr = refattrs->expr;
>>>>       attrs.offset_known_p = refattrs->offset_known_p;
>>>>       attrs.offset = refattrs->offset;
>>>>       attrs.size_known_p = refattrs->size_known_p;
>>>>       attrs.size = refattrs->size;
>>>>       attrs.align = refattrs->align;
>>>>     }
>>>>
>>>> but if we happen to set_mem_align to _exactly_ the MODE_ALIGNMENT
>>>> the MEM_ATTRS are zero, and a smaller alignment may result.
>>>
>>> Not sure what you are saying here.  That
>>>
>>> set_mem_align (MEM:SI A32, 32)
>>>
>>> produces a NULL MEM_ATTRS and thus set_mem_attributes not inheriting
>>> the A32 but eventually computing sth lower?  Yeah, that's probably
>>> an interesting "hole" here.  I'm quite sure that if we'd do
>>>
>>> refattrs = MEM_ATTRS (ref) ? MEM_ATTRS (ref) : mem_mode_attrs[(int) GET_MODE (ref)];
>>>
>>> we run into issues exactly on strict-align targets ...
>>>
>>
>> Yeah, that's scary...
>>
>>>
>>> @@ -3291,6 +3306,23 @@ assign_parm_setup_reg (struct assign_parm_data_all
>>>
>>>        did_conversion = true;
>>>      }
>>> +  else if (MEM_P (data->entry_parm)
>>> +          && GET_MODE_ALIGNMENT (promoted_nominal_mode)
>>> +             > MEM_ALIGN (data->entry_parm)
>>> +          && (((icode = optab_handler (movmisalign_optab,
>>> +                                       promoted_nominal_mode))
>>> +               != CODE_FOR_nothing)
>>> +              || targetm.slow_unaligned_access (promoted_nominal_mode,
>>> +                                                MEM_ALIGN 
>>> (data->entry_parm))))
>>> +    {
>>> +      if (icode != CODE_FOR_nothing)
>>> +       emit_insn (GEN_FCN (icode) (parmreg, validated_mem));
>>> +      else
>>> +       rtl = parmreg = extract_bit_field (validated_mem,
>>> +                       GET_MODE_BITSIZE (promoted_nominal_mode), 0,
>>> +                       unsignedp, parmreg,
>>> +                       promoted_nominal_mode, VOIDmode, false, NULL);
>>> +    }
>>>    else
>>>      emit_move_insn (parmreg, validated_mem);
>>>
>>> This hunk would be obvious to me if we'd use MEM_ALIGN (validated_mem) /
>>> GET_MODE (validated_mem) instead of MEM_ALIGN (data->entry_parm)
>>> and promoted_nominal_mode.
>>>
>>
>> Yes, the idea is just to save some cycles, since
>>
>> parmreg = gen_reg_rtx (promoted_nominal_mode);
>> we know that parmreg will also have that mode, plus
>> emit_move_insn (parmreg, validated_mem) which would be called here
>> asserts that:
>>
>>   gcc_assert (mode != BLKmode
>>               && (GET_MODE (y) == mode || GET_MODE (y) == VOIDmode));
>>
>> so GET_MODE(validated_mem) == GET_MODE (parmreg) == promoted_nominal_mode
>>
>> I still like the current version with promoted_nominal_mode slighhtly
>> better both because of performance, and the 80-column restriction. :)
> 
> So if you say they are 1:1 equivalent then go for it (for this hunk,
> approved as "obvious").
> 

Okay.  Thanks, so I committed that hunk as r274531.

Here is what I have right now, boot-strapped and reg-tested on x86_64-pc-linux-gnu
and arm-linux-gnueabihf (still running, but looks good so far).

Is it OK for trunk?


Thanks
Bernd.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: patch-arm-align-abi.diff
Type: text/x-patch
Size: 11293 bytes
Desc: patch-arm-align-abi.diff
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20190815/da3c0cdb/attachment.bin>


More information about the Gcc-patches mailing list