| 2cebb24bfc3247d3e9be138a3350106737455918 |
|
22-Apr-2015 |
Mathieu Chartier <mathieuc@google.com> |
Replace NULL with nullptr
Also fixed some lines that were too long, and a few other minor
details.
Change-Id: I6efba5fb6e03eb5d0a300fddb2a75bf8e2f175cb
|
| 1095793154d2ff33323ba9edaa4f83373bdb6c8e |
|
24-Mar-2015 |
Goran Jakovljevic <Goran.Jakovljevic@imgtec.com> |
[MIPS] Refactoring code for quick compiler
Code from compiler/dex/quick/mips64 is merged with code
in mips folder.
Change-Id: I785983c21549141306484647da86a0bb4815daaa
|
| f6737f7ed741b15cfd60c2530dab69f897540735 |
|
23-Mar-2015 |
Vladimir Marko <vmarko@google.com> |
Quick: Clean up Mir2Lir codegen.
Clean up WrapPointer()/UnwrapPointer() and OpPcRelLoad().
Change-Id: I1a91f01e1e779599c77f3f6efcac2a6ad34629cf
|
| 027f0ff64c2512b9a5f1f54f3fea1bec481eb0f5 |
|
28-Feb-2015 |
Douglas Leung <douglas.leung@imgtec.com> |
ART: Add Mips32r6 backend support
Add Mips32r6 compiler support.
Don't use deprecated Mips32r2 instructions if running in Mips32r6
mode.
Change-Id: I54e689aa8c026ccb75c4af515aa2794f471c9f67
|
| a2e18ed9397f21c96eae4a26df9ca35a6a97341d |
|
27-Jan-2015 |
Andreas Gampe <agampe@google.com> |
ART: Fix inlining of Mir2Lir
Missed -inl includes.
Change-Id: I39e6d603c7f5d36693aca3816653594488bff63f
|
| 0b9203e7996ee1856f620f95d95d8a273c43a3df |
|
23-Jan-2015 |
Andreas Gampe <agampe@google.com> |
ART: Some Quick cleanup
Make several fields const in CompilationUnit. May benefit some Mir2Lir
code that repeats tests, and in general immutability is good.
Remove compiler_internals.h and refactor some other headers to reduce
overly broad imports (and thus forced recompiles on changes).
Change-Id: I898405907c68923581373b5981d8a85d2e5d185a
|
| 8ebdc2bdbbae5dd014bce8d438f0eca02bad9ff9 |
|
14-Jan-2015 |
Andreas Gampe <agampe@google.com> |
ART: Fix indentation in Mips backend
Fix the indentation to be standard.
Change-Id: I39a16716be3429dfef6df0a585e24423b46363a2
|
| 8d36591d93920e7b7830c3ffee3759b561f5339e |
|
13-Jan-2015 |
Andreas Gampe <agampe@google.com> |
ART: Use jalr instead of jr for Mips
Use the jalr instruction instead of jr in stubs and compiled code.
Change-Id: Idacc5167a5bb0113dc2e7716e4767e5ed07b5e0b
|
| d582fa4ea62083a7598dded5b82dc2198b3daac7 |
|
06-Nov-2014 |
Ian Rogers <irogers@google.com> |
Instruction set features for ARM64, MIPS and X86.
Also, refactor how feature strings are handled so they are additive or
subtractive.
Make MIPS have features for FPU 32-bit and MIPS v2. Use in the quick compiler
rather than #ifdefs that wouldn't have worked in cross-compilation.
Add SIMD features for x86/x86-64 proposed in:
https://android-review.googlesource.com/#/c/112370/
Bug: 18056890
Change-Id: Ic88ff84a714926bd277beb74a430c5c7d5ed7666
|
| 277ccbd200ea43590dfc06a93ae184a765327ad0 |
|
04-Nov-2014 |
Andreas Gampe <agampe@google.com> |
ART: More warnings
Enable -Wno-conversion-null, -Wredundant-decls and -Wshadow in general,
and -Wunused-but-set-parameter for GCC builds.
Change-Id: I81bbdd762213444673c65d85edae594a523836e5
|
| 6a3c1fcb4ba42ad4d5d142c17a3712a6ddd3866f |
|
31-Oct-2014 |
Ian Rogers <irogers@google.com> |
Remove -Wno-unused-parameter and -Wno-sign-promo from base cflags.
Fix associated errors about unused paramenters and implict sign conversions.
For sign conversion this was largely in the area of enums, so add ostream
operators for the effected enums and fix tools/generate-operator-out.py.
Tidy arena allocation code and arena allocated data types, rather than fixing
new and delete operators.
Remove dead code.
Change-Id: I5b433e722d2f75baacfacae4d32aef4a828bfe1b
|
| fc787ecd91127b2c8458afd94e5148e2ae51a1f5 |
|
10-Oct-2014 |
Ian Rogers <irogers@google.com> |
Enable -Wimplicit-fallthrough.
Falling through switch cases on a clang build must now annotate the fallthrough
with the FALLTHROUGH_INTENDED macro.
Bug: 17731372
Change-Id: I836451cd5f96b01d1ababdbf9eef677fe8fa8324
|
| 5aa6e04061ced68cca8111af1e9c19781b8a9c5d |
|
14-Jun-2014 |
Ian Rogers <irogers@google.com> |
Tidy x86 assembler.
Use helper functions to compute when the kind has a SIB, a ModRM and RegReg
form.
Change-Id: I86a5cb944eec62451c63281265e6974cd7a08e07
|
| 660188264dee3c8f3510e2e24c11816c6b60f197 |
|
06-May-2014 |
Andreas Gampe <agampe@google.com> |
ART: Use utils.h::RoundUp instead of explicit bit-fiddling
Change-Id: I249a2cfeb044d3699d02e13d42b8e72518571640
|
| 091cc408e9dc87e60fb64c61e186bea568fc3d3a |
|
31-Mar-2014 |
buzbee <buzbee@google.com> |
Quick compiler: allocate doubles as doubles
Significant refactoring of register handling to unify usage across
all targets & 32/64 backends.
Reworked RegStorage encoding to allow expanded use of
x86 xmm registers; removed vector registers as a separate
register type. Reworked RegisterInfo to describe aliased
physical registers. Eliminated quite a bit of target-specific code
and generalized common code.
Use of RegStorage instead of int for registers now propagated down
to the NewLIRx() level. In future CLs, the NewLIRx() routines will
be replaced with versions that are explicit about what kind of
operand they expect (RegStorage, displacement, etc.). The goal
is to eventually use RegStorage all the way to the assembly phase.
TBD: MIPS needs verification.
TBD: Re-enable liveness tracking.
Change-Id: I388c006d5fa9b3ea72db4e37a19ce257f2a15964
|
| 9da5c1013215176f2a4dbe7a804be899e12d5f68 |
|
28-Mar-2014 |
buzbee <buzbee@google.com> |
Quick compiler, MIPS resource cleanup
MIPS architecture includes internal registers HI and LO.
Similar to condition codes in other architectures, these internal
resouces must be accounted for during instruction scheduling.
Previously, the Quick backend for MIPS dealt with them by defining
rHI and rLO pseudo registers - treating them as actual registers for
def/use masks. This CL changes the handling of these resources to
be in line with how condition codes are used elsewhere - leaving
register definitions to be used for registers.
Change-Id: Idcd77f3107b0c9b081ad05b1aab663fb9f41492d
|
| 2700f7e1edbcd2518f4978e4cd0e05a4149f91b6 |
|
07-Mar-2014 |
buzbee <buzbee@google.com> |
Continuing register cleanup
Ready for review.
Continue the process of using RegStorage rather than
ints to hold register value in the top layers of codegen.
Given the huge number of changes in this CL, I've attempted
to minimize the number of actual logic changes. With this
CL, the use of ints for registers has largely been eliminated
except in the lowest utility levels. "Wide" utility routines
have been updated to take a single RegStorage rather than
a pair of ints representing low and high registers.
Upcoming CLs will be smaller and more targeted. My expectations:
o Allocate float double registers as a single double rather than
a pair of float single registers.
o Refactor to push code which assumes long and double Dalvik
values are held in a pair of register to the target dependent
layer.
o Clean-up of the xxx_mir.h files to reduce the amount of #defines
for registers. May also do a register renumbering to bring all
of our targets' register naming more consistent. Possibly
introduce a target-independent float/non-float test at the
RegStorage level.
Change-Id: I646de7392bdec94595dd2c6f76e0f1c4331096ff
|
| 1da1e2fceb0030b4b76b43510b1710a9613e0c2e |
|
15-Nov-2013 |
buzbee <buzbee@google.com> |
More compile-time tuning
Another round of compile-time tuning, this time yeilding in the
vicinity of 3% total reduction in compile time (which means about
double that for the Quick Compile portion).
Primary improvements are skipping the basic block combine optimization
pass when using Quick (because we already have big blocks), combining
the null check elimination and type inference passes, and limiting
expensive local value number analysis to only those blocks which
might benefit from it.
Following this CL, the actual compile phase consumes roughly 60%
of the total dex2oat time on the host, and 55% on the target (Note,
I'm subtracting out the Deduping time here, which the timing logger
normally counts against the compiler).
A sample breakdown of the compilation time follows (this taken on
PlusOne.apk w/ a Nexus 4):
39.00% -> MIR2LIR: 1374.90 (Note: includes local optimization & scheduling)
10.25% -> MIROpt:SSATransform: 361.31
8.45% -> BuildMIRGraph: 297.80
7.55% -> Assemble: 266.16
6.87% -> MIROpt:NCE_TypeInference: 242.22
5.56% -> Dedupe: 196.15
3.45% -> MIROpt:BBOpt: 121.53
3.20% -> RegisterAllocation: 112.69
3.00% -> PcMappingTable: 105.65
2.90% -> GcMap: 102.22
2.68% -> Launchpads: 94.50
1.16% -> MIROpt:InitRegLoc: 40.94
1.16% -> Cleanup: 40.93
1.10% -> MIROpt:CodeLayout: 38.80
0.97% -> MIROpt:ConstantProp: 34.35
0.96% -> MIROpt:UseCount: 33.75
0.86% -> MIROpt:CheckFilters: 30.28
0.44% -> SpecialMIR2LIR: 15.53
0.44% -> MIROpt:BBCombine: 15.41
(cherry pick of 9e8e234af4430abe8d144414e272cd72d215b5f3)
Change-Id: I86c665fa7e88b75eb75629a99fd292ff8c449969
|
| a61f49539a59b610e557b5513695295639496750 |
|
23-Aug-2013 |
buzbee <buzbee@google.com> |
Add timing logger to Quick compiler
Current Quick compiler breakdown for compiling the boot class path:
MIR2LIR: 29.674%
MIROpt:SSATransform: 17.656%
MIROpt:BBOpt: 11.508%
BuildMIRGraph: 7.815%
Assemble: 6.898%
MIROpt:ConstantProp: 5.151%
Cleanup: 4.916%
MIROpt:NullCheckElimination: 4.085%
RegisterAllocation: 3.972%
GcMap: 2.359%
Launchpads: 2.147%
PcMappingTable: 2.145%
MIROpt:CodeLayout: 0.697%
LiteralData: 0.654%
SpecialMIR2LIR: 0.323%
Change-Id: I9f77e825faf79e6f6b214bb42edcc4b36f55d291
|
| 0d82948094d9a198e01aa95f64012bdedd5b6fc9 |
|
12-Oct-2013 |
buzbee <buzbee@google.com> |
64-bit prep
Preparation for 64-bit roll.
o Eliminated storing pointers in 32-bit int slots in LIR.
o General size reductions of common structures to reduce impact
of doubled pointer sizes:
- BasicBlock struct was 72 bytes, now is 48.
- MIR struct was 72 bytes, now is 64.
- RegLocation was 12 bytes, now is 8.
o Generally replaced uses of BasicBlock* pointers with 16-bit Ids.
o Replaced several doubly-linked lists with singly-linked to save
one stored pointer per node.
o We had quite a few uses of uintptr_t's that were a holdover from
the JIT (which used pointers to mapped dex & actual code cache
addresses rather than trace-relative offsets). Replaced those with
uint32_t's.
o Clean up handling of embedded data for switch tables and array data.
o Miscellaneous cleanup.
I anticipate one or two additional CLs to reduce the size of MIR and LIR
structs.
Change-Id: I58e426d3f8e5efe64c1146b2823453da99451230
|
| 409fe94ad529d9334587be80b9f6a3d166805508 |
|
11-Oct-2013 |
buzbee <buzbee@google.com> |
Quick assembler fix
This CL re-instates the select pattern optimization disabled by
CL 374310, and fixes the underlying problem: improper handling of
the kPseudoBarrier LIR opcode. The bug was introduced in the
recent assembler restructuring. In short, LIR pseudo opcodes (which
have values < 0), should always have size 0 - and thus cause no
bits to be emitted during assembly. In this case, bad logic caused
us to set the size of a kPseudoBarrier opcode via lookup through the
EncodingMap.
Because all pseudo ops are < 0, this meant we did an array underflow
load, picking up whatever garbage was located before the EncodingMap.
This explains why this error showed up recently - we'd previuosly just
gotten a lucky layout.
This CL corrects the faulty logic, and adds DCHECKs to uses of
the EncodingMap to ensure that we don't try to access w/ a
pseudo op. Additionally, the existing is_pseudo_op() macro is
replaced with IsPseudoLirOp(), named similar to the existing
IsPseudoMirOp().
Change-Id: I46761a0275a923d85b545664cadf052e1ab120dc
|
| b48819db07f9a0992a72173380c24249d7fc648a |
|
15-Sep-2013 |
buzbee <buzbee@google.com> |
Compile-time tuning: assembly phase
Not as much compile-time gain from reworking the assembly phase as I'd
hoped, but still worthwhile. Should see ~2% improvement thanks to
the assembly rework. On the other hand, expect some huge gains for some
application thanks to better detection of large machine-generated init
methods. Thinkfree shows a 25% improvement.
The major assembly change was to establish thread the LIR nodes that
require fixup into a fixup chain. Only those are processed during the
final assembly pass(es). This doesn't help for methods which only
require a single pass to assemble, but does speed up the larger methods
which required multiple assembly passes.
Also replaced the block_map_ basic block lookup table (which contained
space for a BasicBlock* for each dex instruction unit) with a block id
map - cutting its space requirements by half in a 32-bit pointer
environment.
Changes:
o Reduce size of LIR struct by 12.5% (one of the big memory users)
o Repurpose the use/def portion of the LIR after optimization complete.
o Encode instruction bits to LIR
o Thread LIR nodes requiring pc fixup
o Change follow-on assembly passes to only consider fixup LIRs
o Switch on pc-rel fixup kind
o Fast-path for small methods - single pass assembly
o Avoid using cb[n]z for null checks (almost always exceed displacement)
o Improve detection of large initialization methods.
o Rework def/use flag setup.
o Remove a sequential search from FindBlock using lookup table of 16-bit
block ids rather than full block pointers.
o Eliminate pcRelFixup and use fixup kind instead.
o Add check for 16-bit overflow on dex offset.
Change-Id: I4c6615f83fed46f84629ad6cfe4237205a9562b4
|
| 252254b130067cd7a5071865e793966871ae0246 |
|
09-Sep-2013 |
buzbee <buzbee@google.com> |
More Quick compile-time tuning: labels & branches
This CL represents a roughly 3.5% performance improvement for the
compile phase of dex2oat. Move of the gain comes from avoiding
the generation of dex boundary LIR labels unless a debug listing
is requested. The other significant change is moving from a basic block
ending branch model of "always generate a fall-through branch, and then
delete it if we can" to a "only generate a fall-through branch if we need
it" model.
The data motivating these changes follow. Note that two area of
potentially attractive gain remain: restructing the assembler model and
reworking the register handling utilities. These will be addressed
in subsequent CLs.
--- data follows
The Quick compiler's assembler has shown up on profile reports a bit
more than seems reasonable. We've tried a few quick fixes to apparently
hot portions of the code, but without much gain. So, I've been looking at
the assembly process at a somewhat higher level. There look to be several
potentially good opportunities.
First, an analysis of the makeup of the LIR graph showed a surprisingly
high proportion of LIR pseudo ops. Using the boot classpath as a basis,
we get:
32.8% of all LIR nodes are pseudo ops.
10.4% are LIR instructions which require pc-relative fixups.
11.8% are LIR instructions that have been nop'd by the various
optimization passes.
Looking only at the LIR pseudo ops, we get:
kPseudoDalvikByteCodeBoundary 43.46%
kPseudoNormalBlockLabel 21.14%
kPseudoSafepointPC 20.20%
kPseudoThrowTarget 6.94%
kPseudoTarget 3.03%
kPseudoSuspendTarget 1.95%
kPseudoMethodExit 1.26%
kPseudoMethodEntry 1.26%
kPseudoExportedPC 0.37%
kPseudoCaseLabel 0.30%
kPseudoBarrier 0.07%
kPseudoIntrinsicRetry 0.02%
Total LIR count: 10167292
The standout here is the Dalvik opcode boundary marker. This is just a
label inserted at the beginning of the codegen for each Dalvik bytecode.
If we're also doing a verbose listing, this is also where we hang the
pretty-print disassembly string. However, this label was also
being used as a convenient way to find the target of switch case
statements (and, I think at one point was used in the Mir->GBC conversion
process).
This CL moves the use of kPseudoDalvikByteCodeBoundary labels to only
verbose listing runs, and replaces the codegen uses of the label with
the kPseudoNormalBlockLabel attached to the basic block that contains the
switch case target. Great savings here - 14.3% reduction in the number of
LIR nodes needed. After this CL, our LIR pseudo proportions drop to 21.6%
of all LIR. That's still a lot, but much better. Possible further
improvements via combining normal labels with kPseudoSafepointPC labels
where appropriate, and also perhaps reduce memory usage by using a
short-hand form for labels rather than a full LIR node. Also, many
of the basic block labels are no longer branch targets by the time
we get to assembly - cheaper to delete, or just ingore?
Here's the "after" LIR pseudo op breakdown:
kPseudoNormalBlockLabel 37.39%
kPseudoSafepointPC 35.72%
kPseudoThrowTarget 12.28%
kPseudoTarget 5.36%
kPseudoSuspendTarget 3.45%
kPseudoMethodEntry 2.24%
kPseudoMethodExit 2.22%
kPseudoExportedPC 0.65%
kPseudoCaseLabel 0.53%
kPseudoBarrier 0.12%
kPseudoIntrinsicRetry 0.04%
Total LIR count: 5748232
Not done in this CL, but it will be worth experimenting with actually
deleting LIR nodes from the graph when they are optimized away, rather
than just setting the NOP bit. Keeping them around is invaluable
during debugging - but when not debugging it may pay off if the cost of
node removal is less than the cost of traversing through dead nodes
in subsequent passes.
Next up (and partially in this CL - but mostly to be done in follow-on
CLs) is the overall assembly process. Inherited from the trace JIT,
the Quick compiler has a fairly simple-minded approach to instruction
assembly. First, a pass is made over the LIR list to assign offsets
to each instruction. Then, the assembly pass is made - which generates
the actual machine instruction bit patterns and pushes the instruction
data into the code_buffer. However, the code generator takes the "always
optimistic" approach to instruction selection and emits the shortest
instruction. If, during assembly, we find that a branch or load doesn't
reach, that short-form instruction is replaces with a longer sequence.
Of course, this invalidates the previously-computed offset calculations.
Assembly thus is an iterative process: compute offsets and then assemble
until we survive an assembly pass without invalidation. This seems
like a likely candidate for improvement. First, I analyzed the
number of retries required, and the reason for invalidation over the
boot classpath load.
The results: more than half of methods don't require a retry, and
very few require more than 1 extra pass:
5 or more: 6 of 96334
4 or more: 22 of 96334
3 or more: 140 of 96334
2 or more: 1794 of 96334 - 2%
1 or more: 40911 of 96334 - 40%
0 retries: 55423 of 96334 - 58%
The interesting group here is the one that requires 1 retry. Looking
at the reason, we see three typical reasons:
1. A cbnz/cbz doesn't reach (only 7 bits of offset)
2. A 16-bit Thumb1 unconditional branch doesn't reach.
3. An unconditional branch which branches to the next instruction
is encountered, and deleted.
The first 2 cases are the cost of the optimistic strategy - nothing
much to change there. However, the interesting case is #3 - dead
branch elimination. A further analysis of the single retry group showed
that 42% of the methods (16305) that required a single retry did so
*only* because of dead branch elimination. The big question here is
why so many dead branches survive to the assembly stage. We have
a dead branch elimination pass which is supposed to catch these - perhaps
it's not working correctly, should be moved later in the optimization
process, or perhaps run multiple times.
Other things to consider:
o Combine the offset generation pass with the assembly pass. Skip
pc-relative fixup assembly (other than assigning offset), but push
LIR* for them into work list. Following the main pass, zip through
the work list and assemble the pc-relative instructions (now that we
know the offsets). This would significantly cut back on traversal
costs.
o Store the assembled bits into both the code buffer and the LIR.
In the event we have to retry, only the pc-relative instructions
would need to be assembled, and we'd finish with a pass over the
LIR just to dumb the bits into the code buffer.
Change-Id: I50029d216fa14f273f02b6f1c8b6a0dde5a7d6a6
|
| 38f85e4892f6504971bde994fec81fd61780ac30 |
|
18-Jul-2013 |
Brian Carlstrom <bdc@google.com> |
Fix cpplint whitespace/operators issues
Change-Id: I730bd87b476bfa36e93b42e816ef358006b69ba5
|
| 2ce745c06271d5223d57dbf08117b20d5b60694a |
|
18-Jul-2013 |
Brian Carlstrom <bdc@google.com> |
Fix cpplint whitespace/braces issues
Change-Id: Ide80939faf8e8690d8842dde8133902ac725ed1a
|
| 7940e44f4517de5e2634a7e07d58d0fb26160513 |
|
12-Jul-2013 |
Brian Carlstrom <bdc@google.com> |
Create separate Android.mk for main build targets
The runtime, compiler, dex2oat, and oatdump now are in seperate trees
to prevent dependency creep. They can now be individually built
without rebuilding the rest of the art projects. dalvikvm and jdwpspy
were already this way. Builds in the art directory should behave as
before, building everything including tests.
Change-Id: Ic6b1151e5ed0f823c3dd301afd2b13eb2d8feb81
|