[PATCH] PR jit/66812: Candidate fix for for the code generation issue, v1

[Dibyendu Majumdar]

On 10 July 2015 at 08:38, Dibyendu Majumdar <mobile@majumdar.org.uk> wrote:
> On 10 July 2015 at 01:45, David Malcolm <dmalcolm@redhat.com> wrote:
>>> however performance degraded similar to when
>>> -fno-strict-aliasing is used.
>>
>> Bother.  What kind of numbers are we talking about?
>>
>
> I will post some numbers tonight.
>

I will not compare performance with LLVM here as I have a workaround
for the code that causes segmentation fault in libgccjit - so the
benchmarks are probably impacted due to this. Here is a comparison of
the benchmarks with / without the fix.


1. fornum_test1.lua

    without fix: 2e-6
    with fix: 2.053

2. fornum_test2.ravi

    without fix: 1.04
    with fix: 3.14

3. mandel1.ravi

    without fix: 3.52
    with fix: 5.94

4. fannkuchen.ravi

    without fix: 11.29
    with fix: 16.98

5. matmul1.ravi 1000

    without fix: 5.13
    with fix: 11.28


> I am beginning to think that using unions was not a good idea - I
> should do it the way I implement the LLVM version - i.e. use a struct
> and emulate the union functionality. I think unions are just hard for
> the optimizer to reason about.
>
> Unfortunately this means a bit of rework - so I won't know if the
> struct approach is better until after I have done the changes and
> compared results.
>