Arg strides are the reason JIT happens lazily in Program::eval() today
instead of proactively in Builder::done() or Program's constructor. It
also just really doesn't make sense to delay this information... it's
not like you can change it up sanely between calls to eval().
The argument index now comes implicitly from the order of calling arg().
This may seem logically independent, but it prevents a weird situation
where you could use the same argument index twice with different
strides... not sure what that would mean.
Change-Id: I0f5d46e94a1ca112a72675c5492f17c0dd825ce0
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/227390
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Basically the same deal as aarch64:
- a bunch of instructions to rewrite control
flow to be two loops, body and tail
- a bunch of instructions to support scalar
loads and stores in the tail
We can now remove the JIT::mask field.
I've removed the SkUNREACHABLE I'd put in for the ARM code... as
written the interpreter is still reachable by the loser if two threads
race to JIT the program. Medium term I plan to move JIT compilation to
a more proactive time, eliminating the need for the lock and letting the
interpreter become truly unreachable.
I had a little bit of a false start with what instructions to use for
scalar load8 and store8, first starting with instructions that loaded
via GP registers, then remembering vpinsrb and vpextrb can take a memory
argument, loading into xmm directly. I've left the first instructions I
used in the file, still implemented but only used from the unit tests.
They're pretty common and will probably be useful some day.
Change-Id: I471b13026af4b1c6e861a53159f9df5f0285447c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/227178
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
I had been setting the REX R bit to select high registers,
but you actually set the B bit. Don't know how I got that
wrong before... the leading byte should be 49 not 4c.
$ cat test.s
foo:
addq $7, %r8
$ clang -c test.s && objdump -d test.o
0000000000000000 <foo>:
0: 49 83 c0 07 add $0x7,%r8
Change-Id: I039e1c4f4ea20523a1e2cc9bcf5f6d9321a6223b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/227177
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
This turned out to be quite an easy transformation
with yesterday's work already done. No codegen changes.
Change-Id: Ife19ab7731514c54cfed963a6d2e9b1ec2246997
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/227137
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Support forward references in Label.
In addition to tracking the current Label offset (used for
backward references essentially just the same as before this CL)
we also store a list of instructions that refer to each Label.
When a Label moves, each instruction gets a new displacement.
To make this a little easier, remove the 8-bit jump form on x86...
this way all x86 displacements are 32-bit and and all ARM 19-bit.
For now only cbz() supports this, just to start somewhere.
More to do but it's worth an early design review.
Change-Id: I23d2bcd7742965ab694ae4828f53409cb9fc807f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/226937
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
This adds a bunch of instructions we'll need to handle the N < 4 tail
within the JIT code on ARM.
- ldrb/strb are 1-byte load and stores
- sub subtracts without setting flags
- cmp just sets flags (actually just subs with an xzr destination)
- add b and b.lt, just like b.ne
- cbz and cbnz... we only need cbz but I accidentally did cbnz first
Once I add support for forward jumps, we'll be able to use these
instructions to restructure the loop to
entry:
hoisted setup
loop:
if N < 4, jump tail (cmp N,#4; b.lt tail)
... handle 4 values ...
jump loop (b loop)
tail:
if N == 0, jump end (cbz N, end)
... handle 1 value ...
jump tail (b tail)
end:
ret
Change-Id: I62d2d190f670f758197a25d99dfde13362189993
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/226828
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Noticed we were only dumping the final register
programs for the integer code. Might as well also
track the value programs.
Change-Id: I417c5c655b632691557bbbb136dcbd3f3167af9a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/225324
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
This is test-only code only used by SkVMTest.cpp,
so it can live there. This cuts the dependency
of SkVM on SkStream and co.
Change-Id: I7695e527b2d16e4485f8c5f4cd39bb8300e9221d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/225321
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Some small refactoring to common up redundant opcode building.
Oddly, I think I've got better codegen than what Clang would do here.
Clang doesn't generate uxtl-based code to unpack 8-bit to 32-bit,
instead preferring to load each byte one at a time and insert them one
at a time.
Me:
ldr s0, [x0]
uxtl v0.8h, v0.8b
uxtl v0.4s, v0.8h
Clang:
ldrb w8, [x0]
ldrb w9, [x0, #1]
ldrb w10, [x0, #2]
ldrb w11, [x0, #3]
fmov s0, w8
mov v0.s[1], w9
mov v0.s[2], w10
mov v0.s[3], w11
Change-Id: I0fdf5c6cdcde6a4eb9290936284fd3ffcb2159f6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/224821
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
So far this is just as easy as I had hoped.
Change-Id: I5f69a900b32d9bf70156b55e334233d7376b820f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/223340
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Instead of allocating into a std::vector, we do one quick first pass to
measure how much memory we need to allocate, mmap enough pages for that,
then another real writing pass.
This cuts a microsecond or so off the profile. There's another
microsecond left to cut if we could eliminate that first measuring pass,
but I'm no longer sure it's easy to come up with a good upper limit on
the program size now that I'm thinking about the data part of the
program as well.
vpshufb is our current max instruction at 9 bytes of code, but that also
implies another 32 bytes of control data. I'm not sure I feel very
clever allocating 41 * |instructions| bytes to be conservatively safe...
it seems like ridiculous overkill.
Ultimately I found it easier to just measure twice, cut once.
Change-Id: I16ccdafbc789711837b41b3d5a557808798eb1b4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/223305
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
I was just reading the ARM docs and realized that their BIC ("BIt
Clear") is the same as SSE's ANDN ("AND Not") instruction. It's kind of
a neat little tool to have laying around... comes up more than you'd
think, and it's sometimes the clearest way to express what you're doing,
as in the changed program here where the comment is "mask away the low
bits". That's a bit_clear with a mask for what you want to clear away!
And the real reason to write this up is that I want to have a CL to
point to that shows how to add an instruction top to bottom.
Change-Id: I99690ed9c1009427b3986955e7ae6264de4d215c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/223120
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Reviewed-by: Mike Reed <reed@google.com>
For now, disable the vpmovusdb AVX-512 instruction, using the compound
AVX2 fallback instead. I need to learn how to encode EVEX prefixes
before we can use that, and it's not very important.
That's everything! We're fully in control now, and should be able to
run this on any x86-64 Linux or Mac. And we can relax some of the
defined(SKVM_JIT) guards so that, e.g., we can unit test Assembler even
on all platforms.
Stifle some warnings about ~bool by ~(int)bool.
Would like to enable when is_mac too but can't seem to get past
(bogus?) thread annotation on the bots. My local Mac is fine. :/
Change-Id: If00bdd97ebd9684ed109933e2fa70c5e6f6ea339
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/222631
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
This shows off a little how easy backwards-only labels are.
The rip == rbp + Mod::Indirect convention isn't something
you'd be able to guess without just looking at the docs.
I'm not actually sure if you can only use rbp or also r13,
but LLVM seems to always do the equivalent of rbp... might
just be that high bit in VEX is ignored: they're registers
5 and 13, 8 apart, only distinguished by that bit.
Convenienly RIP addressing is always 32-bit, so there's
no benefit to spending time checking whether the offset
fits in a byte, though most of our offsets would.
Change-Id: I01b7fb1500667e1bf98490d5144459f92e1b375d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/222857
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
By putting data first in descending alignment then code, we never need
any alignment padding.
This also makes all jumps and ip-relative data loads backward, so
they're really easy to assemble. No need for any sort of deferred
where-does-this-label-mean logic; the label can just be a simple byte
offset established before you need to use it.
Nothing new switched off of Xbyak in this CL, but the rearrangement
makes the rest a lot easier.
The one downside I've found so far is that the disassembly of the
first instruction can get confused into data or other instructions,
e.g.
63: 01 ff add %edi,%edi
65: 00 ff add %bh,%bh
67: 00 00 add %al,(%rax)
69: ff 00 incl (%rax)
6b: ff c4 inc %esp
6d: e2 7d loop ec <skvm-jit-884702985+0xac>
6f: 18 05 eb ff ff ff sbb %al,-0x15(%rip) # 60 <skvm-jit-884702985+0x20>
75: c4 e2 7d 18 0d e6 ff ff ff vbroadcastss -0x1a(%rip),%ymm1 # 64 <skvm-jit-884702985+0x24>
7e: c4 e2 7d 18 15 e1 ff ff ff vbroadcastss -0x1f(%rip),%ymm2 # 68 <skvm-jit-884702985+0x28>
There are 3 vbroadcastss instructions here, each starting with c4 e2 7d
18, but the first has been disassembled as if its c4 were part of the
last data entry (0xff00ff00) as inc %esp.
Probably not a big deal for now, particularly since those vbroadcastss
are all outside the loop and never show up on a profile. If it gets too
confusing I think we can dump the programs starting from the beginning
of the code instead of from the data; we won't be able to inspect the
data, but everything should disassemble perfectly.
Change-Id: I0cc864359fd0740fc026070eaf2b6cb130783a57
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/222574
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
The encoding kind of all goes through the same paths,
as the three argument instructions, but like the nursery
rhyme when there are only two they kind of all roll over
and the op-extension hops into the bed.
vpermq is the first place we need to set the W bit
to indicate a 64-bit lane operation, so a little
minimal plumbing for that. It takes its arguments
a little differently too, passing dst where you'd
expect, the source where we'd pass y, and requiring
us to pass literal 0000 for the vvvv bits in VEX
(inverted as normal to literal 1111).
Change-Id: I91a4cd1b316eb908992631ce8b2cb3c62078e8c6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/222565
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
- 32x8 i32 add,sub,mul
- add I32_Naive bench/test builder to get better i32 mul coverage
- minor refactoring all over
Change-Id: I13cc19ff37a2da0bcff289ba51baac08f456d6c5
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/222485
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
This is a reland of a224fc1105
Changes since original:
- switch fJIT_K to less error prone fJITMask
- guard fJIT Assembler Program member with SKVM_JIT
Not really sure why the mips64el-Debug bot's compiler is crashing;
it does at least make sense to crash where it does... the file
includes SkOpts.h which includes SkVM.h.
If no reasonable code transformation can get it working again
I'll remove the bot. The -Release version is fine, and mips64el
is one of those things I'd happily flush if it blocks progress.
In this end I think all this SKVM_JIT and Xbyak stuff should
go away and make things simple again, hopefully too simple to
crash GCC. :|
Original change's description:
> extract Assembler so it can be tested
>
> And start documenting some structs we'll need
> to replace xbyak.
>
> Change-Id: I21c91642799a54e10af85afc8edbe12a9b4aa062
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/221644
> Reviewed-by: Herb Derby <herb@google.com>
> Commit-Queue: Mike Klein <mtklein@google.com>
Cq-Include-Trybots: skia.primary:Test-Debian9-Clang-GCE-CPU-AVX2-x86_64-Release-All-SK_CPU_LIMIT_SSE2,Build-Debian9-GCC-mips64el-Debug
Change-Id: I6d7c27bc758b23c164ee67067cdfacc291e289fc
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/221983
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
This reverts commit a224fc1105.
Reason for revert: breaking x86-64 bots without AVX2, e.g. Test-Debian9-Clang-GCE-CPU-AVX2-x86_64-Release-All-SK_CPU_LIMIT_SSE2
Original change's description:
> extract Assembler so it can be tested
>
> And start documenting some structs we'll need
> to replace xbyak.
>
> Change-Id: I21c91642799a54e10af85afc8edbe12a9b4aa062
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/221644
> Reviewed-by: Herb Derby <herb@google.com>
> Commit-Queue: Mike Klein <mtklein@google.com>
TBR=mtklein@google.com,herb@google.com
Change-Id: Ie90d57f66e4d45f94db4ab4f485155533faddae1
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/221655
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
And start documenting some structs we'll need
to replace xbyak.
Change-Id: I21c91642799a54e10af85afc8edbe12a9b4aa062
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/221644
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Any time we implement a Program::Instruction with multiple low-level
operations, we risk overwriting any arguments that alias the
destination.
This is why the _I32 tests are failing, mad_unorm8 where d == x. We
want (x*y+x)/256+z, but end up calculating (x*y+x*y)/256+z when x == d.
We could fix this by never allowing any arguments to alias any
destinations, but most instructions don't have this problem, and doing
that blindly would bloat the register count significantly.
We could fix this by knowing which Ops may be prone to aliasing in any
backend, but I find that somewhat error prone and also a little
abstraction- level-violatey. I would have thought, for instance, that
the mad_f32 Op might be vulnerable here, but it's actually not... in any
situation where there is aliasing, we actually lower it to a single
vfmadd instruction, never mul-then-add.
This sort of aliasing issue is going to keep coming back up again and
again, especially with 2-argument architectures like SSE. Luckily it's
trivially easy to fix by reserving a single tmp register to use as the
result of all but the final instructions.
The interpreter is safe because all its switch cases are single r(d) =
... statements. The right hand sides are evaluated before anything is
written back to a destination register slot. Had it been written a
little differently, it could have easily had this same aliasing issue.
Change-Id: I996392ef6af48268238ecae4a97d3bf3b4fba002
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/220600
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
The mask-only special case for extract is wrong...
it never looked it its input!
This not only makes things correct-er, but oddly it also
makes them faster by breaking inter-loop data dependencies.
Disable tests for _I32... they're actually still broken
because of a much more systemic flaw in how I've evaluated
programs. The _F32 and _I32_SWAR JIT code and all interpreted
code is just getting lucky. o_O
While here, update the I32_SWAR code to use the same math as I32,
(x*y+x)/256 for unorm8 mul. This just helps keep me sane.
Change-Id: I1acc09adb84c426fca4b2be5ca8c2d46d9678dd8
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/220577
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
At head we're redoing any n<8 tail from the start,
not continuing from (n/8)*8 like we'd want.
Change-Id: I1a3d24cdffc843bbe6f3e01a163b6e3a20fdd0ca
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/220556
Reviewed-by: Brian Osman <brianosman@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
I used to have a dump of the value program before it was
translated to registers, but it went away a while ago.
This restores it.
Change-Id: I9b8bfcb124843cad4b0dc44bdf0a03e95a0c83d8
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/219757
Commit-Queue: Mike Klein <mtklein@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Can sometimes be hard to know what's going on
on the bots without a little bit more debug help.
Change-Id: Ie556a8de88349170e9d9e44c16098223442838a2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/218316
Auto-Submit: Mike Klein <mtklein@google.com>
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Eliminate the duplicate functionality,
and better testing for the bench builders.
Change-Id: If20e52107738903f854aec431416e573d7a7d640
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/218041
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
- keep expectations in resources/
- overwrite automatically if needed
so we can see the diff in Git
Change-Id: I2486b127ebcc7f40332fd0462e38b1af04d3e32b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/218038
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Change-Id: I5aa5f789180b9caba70952cc60a2e9bbcf3b5a97
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/217983
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
With all the thinking around a stack-based interpreter,
I figured I'd sketch out some ideas for a register VM too.
I kind of have the hunch that this is the direction that
will actually let us replace large amounts of Skia's CPU
backend with an efficient interpreter or JIT.
Change-Id: Ia2b5ba4a3fc27556f5b6ba95cd1ace46d3217403
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/216665
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
