> On 28 Aug 2025, at 14:36, Jeff Law via Gcc <gcc@gcc.gnu.org> wrote:
>
>
>
> On 8/28/25 5:42 AM, Richard Biener via Gcc wrote:
>> On Thu, Aug 28, 2025 at 1:12 PM Rainer Orth via Gcc <gcc@gcc.gnu.org> wrote:
>>>
>>> Hi Sam,
>>>
>>>> When a test fails with 'excess errors', there's often only one actual
>>>> error (an excess "(error|warning|note):") and it'd be nice to not have
>>>> to dig in the .log files to fish that out.
>>>
>>> I think such a move would be a bad mistake. Consider ICEs where you
>>> have something like
>>>
>>> FAIL: gcc.c-torture/compile/pr35318.c -O0 (test for excess errors)
>>> Excess errors:
>>> /vol/gcc/src/hg/master/local/gcc/testsuite/gcc.c-torture/compile/pr35318.c:9:1:
>>> error: unrecognizable insn:
>>> (insn 13 25 26 2 (parallel [
>>> (set (reg:DF 10 %o2 [orig:112 x ] [112])
>>> (asm_operands/v:DF ("") ("=r,r") 0 [
>>> (reg:SI 11 %o3 [orig:112 x+4 ] [112])
>>> (mem/c:DF (plus:SI (reg/f:SI 30 %fp)
>>> (const_int -24 [0xffffffffffffffe8])) [3
>>> %sfp+-24 S8 A64])
>>> [and many more lines...]
>>>
>>> This would clutter the output beyond recognition, especially if this is
>>> a torture test which is run at several optimization options.
>>>
>>> Excess errors, like all others, always require further investigation.
>>> In my experience, digging the full error messages from the .log files is
>>> usually the smallest part of that. You often even have to rerun the
>>> compilation manually to also get the parts that are filtered out by the
>>> prune procs.
>> I find the classification this would provide useful, like I like the
>> (internal compiler error) classification we already have. I would
>> of course not duplicate all of th eabove message but only
>> '(unrecognizable insn)' in the above case.
> It's not the only consideration, but keep in mind that such output is not
> stable and will cause some headaches with scripting that compares two summary
> files.
This is something that I think we want to tackle anyway ..
>
> Even with the bit of instability to the line number in the ICE message, I do
> find the ICE classification useful as well.
… the classification is useful, but the false positive “new fail / old fail
went away” pairs are a real nuiscance .. hopefully we can have some
brainstorming @cauldron about ways to deal with this (e.g. fuzzy matching or
some smart way to discard the twinkling line numbers).
Iain
>
> Jeff
