Re: [Tarantool-patches] [PATCH luajit] Fix TNEW load forwarding with instable types.

[Sergey Kaplun via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

Hi, Maxim!
Thanks for the review!

On 15.05.23, Maxim Kokryashkin wrote:
> 
> Hi, Sergey!
> Thanks for the patch!
> LGTM, except for several nits below.
>  
> > 
> >>From: Mike Pall <mike>
> >>
> >>(cherry picked from commit d4c0c6e17ef7edf1f2893bc807746b80612e63e9)
> >>
> >>During loop unrolling IR instructions are coped, substituted and
> >>re-emitted for the FOLD/CSE/etc. pipeline. Assume, we have the following
> >>loop-carried dependency: we create a new table or load some stored one
> >>depends on iteration number. In case when we copy the emitted ALOAD IR
> >Typo: s/depends on iteration/depending on the iteration/
> >Typo: s/In case when/If/
> >>instruction (from from the stored table), we copy it as is, including
> >Typo: s/from from/from/
> >>loading type. But the ALOAD from the new table created (on the previous
> >Typo: s/loading/load/
> >>iteration) should have nil type, so the assertion in `fwd_ahload()` is
> >>failed.
> >>
> >>This patch replaces the assertion to `return 0` to avoid the assertion
> >Typo: s/assertion to/assertion with/
> >>failure and stop forwarding in such situations.
> >>
> >>But the emitted type-guarded ALOAD will lead to the persistent failure
> >>of the assertion guard on the trace. Hence, another one fix should be
> >Typo: s/another one/another/
> >>added. Also, TDUP IR is affected, too.
> >>See also the issue mentioned in the test.

Thanks for the comments, fixed, branch is force-pushed.
The new commit message is the following:

| Fix TNEW load forwarding with instable types.
|
| (cherry picked from commit d4c0c6e17ef7edf1f2893bc807746b80612e63e9)
|
| During loop unrolling IR instructions are copied, substituted and
| re-emitted for the FOLD/CSE/etc. pipeline. Assume, we have the following
| loop-carried dependency: we create a new table or load some stored one
| depends on the iteration number. If we copy the emitted ALOAD IR
| instruction (from the stored table), we copy it as is, including load
| type. But the ALOAD from the new table created (on the previous
| iteration) should have nil type, so the assertion in `fwd_ahload()` is
| failed.
|
| This patch replaces the assertion with `return 0` to avoid the assertion
| failure and stop forwarding in such situations.
|
| But the emitted type-guarded ALOAD will lead to the persistent failure
| of the assertion guard on the trace. Hence, another fix should be
| added. Also, TDUP IR is affected, too.
| See also the issue mentioned in the test.
|
| Sergey Kaplun:
| * added the description and the test for the problem
|
| Part of tarantool/tarantool#8516

> >>
> >>Sergey Kaplun:
> >>* added the description and the test for the problem
> >>
> >>Part of tarantool/tarantool#8516
> >>---
> >>
> >>Branch:  https://github.com/tarantool/luajit/tree/skaplun/lj-994-instable-types-during-loop-unroll
> >>PR:  https://github.com/tarantool/tarantool/pull/8642
> >>Related issues:
> >>*  https://github.com/tarantool/tarantool/issues/8516
> >>*  https://github.com/LuaJIT/LuaJIT/issues/994
> >>
> >>I don't mention the 994 intentionally to avoid Mike bothering. Also, it
> >>isn't the origin of this commit. Quite the opposite: as a result of this
> >>backporting the following issue was created.
> >>
> >>I prefer to backport the patches (this one and the prospective for 994)
> >>separately. So, after this patch is backported, it doesn't add any
> >>critical bugs (always failing guard just creates a new side trace), but
> >>helps to avoid conflicts for future backporting (sadly remembered
> >>"Improve assertions.").
> >>
> >> src/lj_opt_mem.c | 3 +-
> >> ...instable-types-during-loop-unroll.test.lua | 53 +++++++++++++++++++
> >> 2 files changed, 55 insertions(+), 1 deletion(-)
> >> create mode 100644 test/tarantool-tests/lj-994-instable-types-during-loop-unroll.test.lua
> >>
> >>diff --git a/src/lj_opt_mem.c b/src/lj_opt_mem.c
> >>index cc177d39..c8265b4f 100644
> >>--- a/src/lj_opt_mem.c
> >>+++ b/src/lj_opt_mem.c

<snipped>

> >>diff --git a/test/tarantool-tests/lj-994-instable-types-during-loop-unroll.test.lua b/test/tarantool-tests/lj-994-instable-types-during-loop-unroll.test.lua
> >>new file mode 100644
> >>index 00000000..435f6e0e
> >>--- /dev/null
> >>+++ b/test/tarantool-tests/lj-994-instable-types-during-loop-unroll.test.lua

<snipped>

> >>+ -- Use non-constant key to avoid LJ_TRERR_GFAIL and undoing the
> >>+ -- loop.
> >Typo: s/Use/Use a/

Fixed.

> >>+ result = t[key]
> >>+ -- Swap table loaded by SLOAD to the created via TNEW.
> >>+ slot = _ % 2 ~= 0 and stored_tab or {}
> >>+end
> >>+test:is(result, nil, 'TNEW load forwarding was successful')
> >Nit: not sure about that assertion. AFAICS, we want to check whether we
> >have failed on assertion and not the validity of `result` variable. Feel free
> >to ignore.

I want to verify that behaviour isn't changed by JIT, so a leave this
check as is.
Ignoring.

> >>+
> >>+-- TDUP.
> >>+--[[
> >>+-- FIXME: Disable test case for now. Enable, with another one
> >Typo: s/another one/another/

Fixed, thanks!

> >>+-- backported commit with a fix for the aforementioned issue

<snipped>

> >>+test:is(result, true, 'TDUP load forwarding was successful')
> >>+]]
> >Nit: Not sure if it is good to add this test here. Maybe it is better
> >to create a ticket in our repo and move that chunk there. Feel free
> >to ignore.

Me, too, but I'll prefer to place these tests in one test file, since it
is the general issue for TDUP|TNEW IR-s. Also, reference the single
ticket.
Ignoring.

> >>+
> >>+os.exit(test:check() and 0 or 1)
> >>--
> >>2.34.1
> >--
> >Best regards,
> >Maxim Kokryashkin
> > 

-- 
Best regards,
Sergey Kaplun