On 6/4/25 11:43, Vineet Gupta wrote:
> On 6/4/25 11:34, Paul-Antoine Arras wrote:
>> Hi Vineet,
>>
>> On 04/06/2025 20:29, Vineet Gupta wrote:
>>> Hi Paul,
>>>
>>> On 5/30/25 03:04, Paul-Antoine Arras wrote:
>>>> This pattern enables the combine pass (or late-combine, depending on the
>>>> case)
>>>> to merge a vec_duplicate into a plus-mult or minus-mult RTL instruction.
>>>>
>>>> Before this patch, we have two instructions, e.g.:
>>>> vfmv.v.f v6,fa0
>>>> vfmadd.vv v9,v6,v7
>>>>
>>>> After, we get only one:
>>>> vfmadd.vf v9,fa0,v7
>>> This patch seems to cause a functional regression. I'm seeing SPEC2017 526
>>> blender output mismatch failures on QEMU icounts test harness
>>>
>>> running compare cmd 1
>>> cpu2017-1.1.7/bin/specperl cpu2017-1.1.7/bin/harness/specdiff -m -l 10
>>> --reltol 0.05
>>>
>>> cpu2017-1.1.7/benchspec/CPU/526.blender_r/data/refrate/output/imagevalidate_sh3_no_char_0849.out
>>> imagevalidate_sh3_no_char_0849.out
>>>
>>> exited with status 1
>> I have not tried to reproduce it yet but it likely comes from a sign
>> mistake in the handling of vfmsub. Does the attached patch fix it for you?
> Thx, I can try it.
Yeah I can confirm it fixes it.
Thx,
-Vineet
