Sorry for my late reply (I just came back from vacation last night).
> On Aug 23, 2021, at 8:55 AM, Richard Biener <rguent...@suse.de> wrote:
>
>>>>
>>>>
>>>> Looks like for the following code:
>>>>
>>>> 3026 if (!reg_lhs)
>>>> 3027 {
>>>> 3028 /* If this is a VLA or the variable is not in register,
>>>> 3029 expand to a memset to initialize it. */
>>>> 3030 mark_addressable (lhs);
>>>> 3031 tree var_addr = build_fold_addr_expr (lhs);
>>>> 3032
>>>> 3033 tree value = (init_type == AUTO_INIT_PATTERN) ?
>>>> 3034 build_int_cst (integer_type_node,
>>>> 3035 INIT_PATTERN_VALUE) :
>>>> 3036 integer_zero_node;
>>>> 3037 tree m_call = build_call_expr (builtin_decl_implicit
>>>> (BUILT_IN_MEMSET),
>>>> 3038 3, var_addr, value, var_size);
>>>> 3039 /* Expand this memset call. */
>>>> 3040 expand_builtin_memset (m_call, NULL_RTX, TYPE_MODE (var_type));
>>>> 3041 }
>>>>
>>>> At line 3030, “lhs” could be a SSA_NAME.
>>>>
>>>> My questions are:
>>>>
>>>> 1. Could the routine “mark_addressable” and “build_fold_addr_expr” be
>>>> applied on SSA_NAME?
>>>
>>> No.
>>>
>>>> 2. Could the routine “expand_builtin_memset” be applied on the memset call
>>>> whose “DEST” is
>>>> an address expression on SSA_NAME?
>>>
>>> No.
>>>
>>>> 3. Within “expand_DEFERRED_INIT”, can I call “expand_builtin_memset” to
>>>> expand .DEFERRED_INIT?
>>>
>>> Well, not with "invalid" GENERIC I fear (address of a SSA name).
>>>
>>>> I suspect that one of the above 3 might be the issue, but not sure which
>>>> one?
>>>
>>> All of the above ;) So while reg_lhs is now precise as to how the
>>> variable will end up (the SSA name will end up as a stack variable in this
>>> case, for whatever reason), expansion via memcpy only works when
>>> working on the RTL representation. The usual "workaround" (ugh)
>>> is to use make_tree (), so in the !reg_lhs path you'd do
>>>
>>> /* Get a new GENERIC representation for the RTL. That's necesary
>>> in case LHS is an SSA name. */
>>> lhs = make_tree (TREE_TYPE (lhs), tem);
>>
>> This resolved the issue.
>>
>> Another question,
>>
>> Previously, I used
>>
>> if (TREE_CODE (lhs) == SSA_NAME)
>> lhs = SSA_NAME_VAR (lhs);
>>
>> To resolve this issue. The purpose looks like the same as “make_tree”, just
>> get an generic tree for the LHS.
>>
>> Why you said using SSA_NAME_VAR is broken? Is it because SSA_NAME_VAR will
>> not always return a valid TREE?
>
> Because it's simply the wrong entity - I have no idea why that even
> worked. Ah, cfgexpand associates it with some DECL_RTL for the
> benefit of debug info. But it's still wrong.
>
>> I should use as following
>>
>>
>> If (TREE_CODE (lhs) == SSA_NAME) && SSA_NAME_VAR (lhs))
>> Lhs = SSA_NAME_VAR (lhs)
>>
>> ?
>
> No. A SSA_NAME_VAR can have multiple SSA_NAMEs (obviously) and
> they do not necessarily have to be allocated to the same variable
> partition - that is, there's no 1:1 relationship between SSA_NAME
> and stack slot or (pseudo) register. You want to initialize the
> storage associated with the SSA_NAME in the .DEFERRED_INIT call,
> not some other storage.
Okay, thanks for the explanation. Now I understand.
>
>>>
>>> alternatively you could maybe do
>>>
>>> if (DECL_P (lhs))
>>> {
>>> + rtx tem = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
>>> + reg_lhs = !MEM_P (tem);
>>> }
>>> else if (TREE_CODE (lhs) == SSA_NAME)
>>> reg_lhs = true;
>>> else
>>> reg_lhs = false;
>>>
>>> thus treat SSA names as register storage always (even if it will end
>>> up on the stack).
>>
>> My question here, for a complicate structure SSA_NAME, will expanding
>> through memset better than expand_asssignment?
>
> It depends. In the end I'd consider it a missed-optimization bug on
> the side that generates worse code - but I do expect cases will exist
> for both.
Okay. I agree.
> Clearly memset will be worse when dealing with register
> initialization (thus the !MEM_P check) and I expect memset to be OK
> for stack where member-wise init esp. with non-zero might turn up
> worse code.
So, for SSA_NAME, since they are all treated as reg_lhs, they will be expanded
through “expand_assignment”
as member-wise init, therefore might generate worse code for a SSA_NAME that is
in stack actually.
Qing
>
> Richard.
>
>> Qing
>>>
