[Tarantool-patches] [PATCH luajit] Fix bytecode register allocation for comparisons.
Sergey Kaplun
skaplun at tarantool.org
Mon Aug 16 19:40:56 MSK 2021
Igor,
Thanks for the review!
See the new comment message above:
===================================================================
Fix bytecode register allocation for comparisons.
(cherry picked from commit 2f3f07882fb4ad9c64967d7088461b1ca0a25d3a)
When LuaJIT is built with LJ_FR2 (e.g. with GC64 mode enabled),
information about frame takes two slots -- the first takes the TValue
with the function to be called, the second takes the framelink. The JIT
recording machinery does pretty the same -- the function IR_KGC is
loaded in the first slot, and the second is set to TREF_FRAME value.
This value should be rewritten after return from a callee. This slot is
cleared either by return values or manually (set to zero), when there
are no values to return. The latter case is done by the next bytecode
with RA dst mode. This obliges that the destination of RA takes the next
slot after TREF_FRAME. Hence, this an earlier instruction must use the
smallest possible destination register (see `lj_record_ins()` for the
details).
Bytecode emitter swaps operands for ISGT and ISGE comparisons. As a
result, the aforementioned rule for registers allocations may be
violated. When it happens for a chunk being recorded, the slot with
TREF_FRAME is not rewritten (but the next empty slot after TREF_FRAME
is). This leads to JIT slots inconsistency and assertion failure in
`rec_check_slots()` during recording of the next bytecode instruction.
This patch fixes bytecode register allocation by changing the VM
register allocation order in case of ISGT and ISGE bytecodes.
Sergey Kaplun:
* added the description and the test for the problem
Resolves tarantool/tarantool#6227
Part of tarantool/tarantool#5629
===================================================================
On 16.08.21, Igor Munkin wrote:
> Sergey,
>
> Thanks for the explanation! Please consider the new comments below.
>
> On 01.08.21, Sergey Kaplun wrote:
> > Hi, Igor!
> >
> > Thanks for the review!
> >
> > On 01.08.21, Igor Munkin wrote:
> > > Sergey,
> > >
> > > Thanks for the patch! Please consider the comments below. I didn't check
> > > the test yet, since I don't get the JIT peculiarities from your commit
> > > message and comments. Please provide a clearer description and I'll
> > > proceed with the review of the test case then.
> > >
> > > On 19.07.21, Sergey Kaplun wrote:
> > > > From: Mike Pall <mike>
> > > >
> > > > (cherry picked from commit 2f3f07882fb4ad9c64967d7088461b1ca0a25d3a)
> > > >
> > > > When LuaJIT is build with LJ_FR2 (GC64), information about frame takes
> > > > two slots -- the first takes the TValue with the function to call, the
> > > > second takes the additional frame information. The recording JIT
> > >
> > > Minor: The second slot is the framelink in LuaJIT terms.
> >
> > Yes, because it takes the additional frame information. How do you want
> > to modify this line?
>
> Just say that the second slot takes the framelink: this is lapidary.
Fixed.
>
> >
> > >
> > > > machinery works pretty the same -- the function IR_KGC is loaded in the
> > > > first slot, and the second is set to TREF_FRAME value. This value
> > > > should be rewritten after return from a callee. It is done either by the
> > > > return values either this slot is cleared (set to zero) manually with
> > > > the next bytecode with RA dst mode with the assumption, that the dst RA
> > > > takes the next slot after TREF_FRAME, i.e. an earlier instruction uses
> > > > the smallest possible destination register (see `lj_record_ins()` for
> > > > the details).
> > >
> > > The main point lies in the monstrous 5-line sentence. I've read several
> > > times, but still don't get it. Could you please reword it in a not such
> > > complex sentence?
> >
> > The first option is rewrite this slot by return values from the
> > function.
>
> And this is not the case, right? I mean, this approach works fine even
> without the patch, doesn't it?
Yep, exactly.
>
> >
> > The second option is clearing slot (i.e. set to zero) manually, when
> > there is no values to return. It is done by the next bytecode having RA
> > dst mode. This obliges that the destination of RA takes the next slot
> > after TREF_FRAME. For this an earlier instruction must use the smallest
> > possible destination register (see `lj_record_ins()` for the details).
>
> Here is the case, got it, thanks! So, I guess it's enough to adjust the
> commit message to be similar to the section above.
Updated.
>
> >
> > >
> > > >
> > > > Bytecode allocator swaps operands for ISGT and ISGE comparisons.
>
> I believe this should be called "bytecode emitter" or just "frontend".
Fixed.
>
> > > > When it happens, the aforementioned rule for registers allocations
> > > > may be violated. When it happens, and this chunk is recording, the slot
> > > > with TREF_FRAME is not rewritten (but the next empty slot after
> > > > TREF_FRAME is) during bytecode recording. This leads to JIT slots
> > > > inconsistency and assertion failure in `rec_check_slots()` during
> > > > recording the next bytecode instruction.
> > > >
> > > > This patch fixes bytecode register allocation by changing the register
> > > > allocation order in case of ISGT and ISGE bytecodes.
> > >
> > > It's better to use "virtual register" or even "VM register" to avoid
> > > ambiguous plain "register" usage.
> >
> > Changed to VM register.
> >
> > >
> > > >
> > > > Sergey Kaplun:
> > > > * added the description and the test for the problem
> > > >
> > > > Resolves tarantool/tarantool#6227
> > >
> > > Minor: Why #5629 is not mentioned?
> >
> > Added.
> > Branch is updated and force-pushed.
> >
> > >
> > > > ---
> > > >
> > > > Branch: https://github.com/tarantool/luajit/tree/skaplun/gh-6227-fix-bytecode-allocator-for-comp
> > > > Tarantool branch: https://github.com/tarantool/tarantool/tree/skaplun/gh-6227-fix-bytecode-allocator-for-comp
> > > > Issue: https://github.com/tarantool/tarantool/issues/6227
> > > >
> > > > src/lj_parse.c | 7 +++-
> > > > ...ytecode-allocator-for-comparisons.test.lua | 41 +++++++++++++++++++
> > > > 2 files changed, 46 insertions(+), 2 deletions(-)
> > > > create mode 100644 test/tarantool-tests/gh-6227-bytecode-allocator-for-comparisons.test.lua
> > > >
>
> <snipped>
>
> > > > diff --git a/test/tarantool-tests/gh-6227-bytecode-allocator-for-comparisons.test.lua b/test/tarantool-tests/gh-6227-bytecode-allocator-for-comparisons.test.lua
> > > > new file mode 100644
> > > > index 00000000..66f6885e
> > > > --- /dev/null
> > > > +++ b/test/tarantool-tests/gh-6227-bytecode-allocator-for-comparisons.test.lua
> > > > @@ -0,0 +1,41 @@
> > > > +local tap = require('tap')
> > > > +local test = tap.test('gh-6227-bytecode-allocator-for-comparisons')
> > > > +test:plan(1)
> > > > +
> > > > +-- Test file to demonstrate assertion failure during recording
> > > > +-- wrong allocated bytecode for comparisons.
> > > > +-- See also https://github.com/tarantool/tarantool/issues/6227.
> > > > +
> > > > +-- Need function with RET0 bytecode to avoid reset of
> > > > +-- the first JIT slot with frame info. Also need no assignments
> > > > +-- by the caller.
> > > > +local function empty() end
> > > > +
> > > > +local uv = 0
> > > > +
> > > > +-- This function needs to reset register enumerating.
> > > > +-- Also set `J->maxslot` to zero.
>
> Please add the reason, why J->maxslot is zero (it is initialized with
> nargs in <rec_call_setup>).
Fixed.
>
> > > > +-- The upvalue function to call is loaded to 0 slot.
> > > > +local function bump_frame()
> > > > + -- First call function with RET0 to set TREF_FRAME in the
> > > > + -- last slot.
> > > > + empty()
> > > > + -- Test ISGE or ISGT bytecode. These bytecodes swap their
> > > > + -- operands. Also, a constant is always loaded into the slot
> > > > + -- smaller than upvalue. So, if upvalue loads before KSHORT,
> > > > + -- then the difference between registers is more than 2 (2 is
> > > > + -- needed for LJ_FR2) and TREF_FRAME slot is not rewriting by
> > > > + -- the bytecode after call and return as expected. That leads
>
> If the constant is loaded into a slot prior to the one with an upvalue,
> then how upvalue can be loaded *before* KSHORT? How the difference
> becomes more than 2? I don't get this math.
See updated comment below.
>
> Furthermore, what does stop you from using local variables?
They occupy new slots and make it harder to maintain, see the new
comment below.
>
> > > > + -- to recording slots inconsistency and assertion failure at
> > > > + -- `rec_check_slots()`.
> > > > + empty(1>uv)
> > > > +end
> > > > +
> > > > +jit.opt.start('hotloop=1')
>
> It's worth to mention, that such JIT engine tuning allows to compile
> <empty> function at first, and only later compile the loop below. As a
> result <empty> function is not inlined into the loop body, so the fix
> can be checked.
Two call to `empty()` function is necessary for 2 slot gap after call,
see the new comment below.
>
> > > > +
> > > > +for _ = 1,3 do
>
> Minor: Space is missing after the comma.
Fixed.
===================================================================
diff --git a/test/tarantool-tests/gh-6227-bytecode-allocator-for-comparisons.test.lua b/test/tarantool-tests/gh-6227-bytecode-allocator-for-comparisons.test.lua
index 66f6885e..9788923a 100644
--- a/test/tarantool-tests/gh-6227-bytecode-allocator-for-comparisons.test.lua
+++ b/test/tarantool-tests/gh-6227-bytecode-allocator-for-comparisons.test.lua
@@ -14,26 +14,39 @@ local function empty() end
local uv = 0
-- This function needs to reset register enumerating.
--- Also set `J->maxslot` to zero.
--- The upvalue function to call is loaded to 0 slot.
+-- `J->maxslot` is initialized with `nargs` (i.e. zero in this
+-- case) in `rec_call_setup()`.
local function bump_frame()
-- First call function with RET0 to set TREF_FRAME in the
-- last slot.
empty()
+ -- The old bytecode to be recorded looks like the following:
+ -- 0000 . FUNCF 4
+ -- 0001 . UGET 0 0 ; empty
+ -- 0002 . CALL 0 1 1
+ -- 0000 . . JFUNCF 1 1
+ -- 0001 . . RET0 0 1
+ -- 0002 . CALL 0 1 1
+ -- 0003 . UGET 0 0 ; empty
+ -- 0004 . UGET 3 1 ; uv
+ -- 0005 . KSHORT 2 1
+ -- 0006 . ISLT 3 2
-- Test ISGE or ISGT bytecode. These bytecodes swap their
- -- operands. Also, a constant is always loaded into the slot
- -- smaller than upvalue. So, if upvalue loads before KSHORT,
- -- then the difference between registers is more than 2 (2 is
- -- needed for LJ_FR2) and TREF_FRAME slot is not rewriting by
- -- the bytecode after call and return as expected. That leads
- -- to recording slots inconsistency and assertion failure at
- -- `rec_check_slots()`.
+ -- operands (consider ISLT above).
+ -- Two calls of `empty()` function in a row is necessary for 2
+ -- slot gap in LJ_FR2 mode.
+ -- Upvalue loads before KSHORT, so the difference between slot
+ -- for upvalue `empty` (function to be called) and slot for
+ -- upvalue `uv` is more than 2. Hence, TREF_FRAME slot is not
+ -- rewritten by the bytecode after return from `empty()`
+ -- function as expected. That leads to recording slots
+ -- inconsistency and assertion failure at `rec_check_slots()`.
empty(1>uv)
end
jit.opt.start('hotloop=1')
-for _ = 1,3 do
+for _ = 1, 3 do
bump_frame()
end
===================================================================
>
> > > > + bump_frame()
> > > > +end
> > > > +
> > > > +test:ok(true)
> > > > +os.exit(test:check() and 0 or 1)
> > > > --
> > > > 2.31.0
> > > >
> > >
> > > --
> > > Best regards,
> > > IM
> >
> > --
> > Best regards,
> > Sergey Kaplun
>
> --
> Best regards,
> IM
--
Best regards,
Sergey Kaplun
More information about the Tarantool-patches
mailing list