Re: [PATCH] PR gold/17704

[Sriraman Tallam <tmsriram at google dot com>]

On Sun, Jul 17, 2016 at 10:50 PM, Alan Modra <amodra@gmail.com> wrote:
> On Thu, Jul 14, 2016 at 11:48:39AM -0700, Sriraman Tallam wrote:
>> @@ -671,7 +667,31 @@ match_sections(unsigned int iteration_num,
>>                           this_secn_contents.c_str(),
>>                           this_secn_contents.length()) != 0)
>>                    continue;
>> -              (*kept_section_id)[i] = kept_section;
>> +
>> +           // Check section alignment here.
>> +              // The section with the larger alignment requirement
>> +              // should be kept. Do not fold if the larger is not
>> +              // divisible by the smaller, but that case will never
>> +              // happen if the alignments are powers of 2.
>> +              uint64_t align_i = section_addraligns[i];
>> +              uint64_t align_kept = section_addraligns[kept_section];
>> +              if (align_i <= align_kept)
>> +                {
>> +                  if (align_i != align_kept && align_i != 0
>> +                   && (align_kept % align_i) != 0)
>> +                      continue;
>> +                  (*kept_section_id)[i] = kept_section;
>> +                }
>> +              else
>> +                {
>> +                  if (align_kept != 0 && (align_i % align_kept) != 0)
>> +                      continue;
>> +                  (*kept_section_id)[kept_section] = i;
>> +                  it->second = i;
>> +                  full_section_contents[kept_section].swap(
>> +                      full_section_contents[i]);
>> +                }
>> +
>>                converged = false;
>>                break;
>>              }
>
> The gabi says of sh_addralign "Currently, only 0 and positive integral
> powers of two are allowed", so I think you can discount the
> possibility of alignment not a power of two.  It's not worth
> supporting here when other parts of gold do not, eg. see layout.cc
> calls to align_address.
>
> Please also fix indentation, only some of the above lines use tabs.

New patch attached with the changes.

PR gold/17704
* icf.cc (match_sections): Add new parameter section_addraligns.
Check section alignment and keep the section with the strictest
alignment.
(find_identical_sections): New local variable section_addraligns.
Store each section's alignment.
* testsuite/pr17704a_test.s: New file.
* testsuite/Makefile.am (pr17704a_test): New test.
* testsuite/Makefile.in: Regenerate.

Thanks
Sri

>
>> diff --git a/gold/testsuite/Makefile.am b/gold/testsuite/Makefile.am
>> index c8d3093..7b9b267 100644
>> --- a/gold/testsuite/Makefile.am
>> +++ b/gold/testsuite/Makefile.am
>> @@ -355,6 +355,12 @@ icf_sht_rel_addend_test: icf_sht_rel_addend_test_1.o icf_sht_rel_addend_test_2.o
>>  icf_sht_rel_addend_test.stdout: icf_sht_rel_addend_test
>>       $(TEST_NM) icf_sht_rel_addend_test > icf_sht_rel_addend_test.stdout
>>
>> +check_PROGRAMS += pr17704a_test
>> +pr17704a_test.o: pr17704a_test.s
>> +     $(TEST_AS) --64  -o $@ $<
>> +pr17704a_test: pr17704a_test.o gcctestdir/ld
>> +     gcctestdir/ld --icf=all -o $@ $<
>> +
>>  check_PROGRAMS += large_symbol_alignment
>>  large_symbol_alignment_SOURCES = large_symbol_alignment.cc
>>  large_symbol_alignment_DEPENDENCIES = gcctestdir/ld
>
> The new test contains x86 assembly, but is enabled for all native
> targets using gcc.  Please move to a more appropriate part of the
> makefile.
>
> --
> Alan Modra
> Australia Development Lab, IBM

Attachment: icf_section_alignment.txt
Description: Text document