On Tue, Oct 11, 2016 at 11:26 AM, Martin Liška <mli...@suse.cz> wrote:
> On 10/07/2016 12:50 PM, Richard Biener wrote:
>> On Fri, Oct 7, 2016 at 10:39 AM, Martin Liška <mli...@suse.cz> wrote:
>>> I'm resending the patch, where I implemented all builtins mentions in
>>> subject
>>> in gimp-fold.c.
>>>
>>> Patch can bootstrap on ppc64le-redhat-linux and survives regression tests.
>>>
>>> Ready to be installed?
>>
>> + case BUILT_IN_STRNCASECMP:
>> + {
>> + r = strncmp (p1, p2, length);
>> + if (r == 0)
>> + known_result = true;
>>
>> length might be -1 here -- I think you need to guard against that (but you
>> can
>> handle BUILT_IN_STRCASECMP which you miss?). Likewise for the strncmp case.
>
> Fixed, I've added comment to STRCASECMP case.
>
>>
>> Also do we know whether the c_getstr () result is '\0'-terminated? AFAICS
>> these
>> constant foldings were not implemented in builtins.c, I see a STRCMP one in
>> fold-const-call.c though. I believe the STRING_CST string is not guaranteed
>> to
>> be '\0' terminated (STRING_CST comes with explicit length).
>
> You are absolutely right that we do not have always '\0'-terminated string
> constants.
> Thus I'll send a patch that would return a string from c_getstr just in case
> string[string_length] == 0 (separate email with patch will be sent).
Maybe add a second output to c_getstr to pass down the string length
in case it is known?
In this case you could use strN* () variants for constant folding.
"not found" would need
to be folded conditional on null termination to avoid folding
undefined behavior.
Richard.
>>
>> + tree temp = fold_build2_loc (loc, MEM_REF, cst_uchar_node, str1,
>> + off0);
>> + temp = gimple_build (&stmts, loc, NOP_EXPR, cst_uchar_node, temp);
>>
>> please don't use gimple_build here, there is nothing to simplify for it.
>> Using
>> a NOP_EXPR is also confusing (to match the API...). Just do
>> gimple_build_assign (make_ssa_name (...), ..) like other folders do.
>>
>> + replace_call_with_value (gsi, fold_convert_loc (loc, type, temp));
>>
>> and you'd want to replace the call with the MEM_REF stmt using
>> gsi_replace_with_seq_vops as you fail to set virtual operands properly
>> above (thus you'll get ICEs when this only matches during GIMPLE opts).
>>
>> + location_t loc = gimple_location (stmt);
>> + tree cst_uchar_node = build_type_variant (unsigned_char_type_node, 1, 0);
>> + tree cst_uchar_ptr_node
>> + = build_pointer_type_for_mode (cst_uchar_node, ptr_mode, true);
>> + tree off0 = build_int_cst (cst_uchar_ptr_node, 0);
>>
>> it would be nice to not do this tree building if nothign is folded.
>>
>> + case BUILT_IN_STRCMP:
>> + return gimple_fold_builtin_string_compare (gsi);
>> + case BUILT_IN_STRCASECMP:
>> + return gimple_fold_builtin_string_compare (gsi);
>> + case BUILT_IN_STRNCMP:
>> + return gimple_fold_builtin_string_compare (gsi);
>> + case BUILT_IN_STRNCASECMP:
>> + return gimple_fold_builtin_string_compare (gsi);
>>
>> please do
>>
>> + case BUILT_IN_STRCMP:
>> + case BUILT_IN_STRCASECMP:
>> ...
>> + return gimple_fold_builtin_string_compare (gsi);
>>
>> Thanks,
>> Richard.
>
> Sure, all notes will be fixed in an email which reply to this one.
>
> Martin
>
>>
>>> Martin
>
