occheung/artiq
1
0
Fork 0
forked from M-Labs/artiq
Code Pull Requests Activity
3,210 Commits 14 Branches 35 Tags 23 MiB
7be3ffe79f
Commit Graph

490 Commits

Author SHA1 Message Date
whitequark
9cc9e8b276 embedding: s/kernel_constant_attributes/kernel_invariants/g 
Requested in #359.
 2016-04-06 22:38:31 +00:00
whitequark
8b1c5e3a26 llvm_ir_generator: add fast-math flags to fcmp. 
This is allowed in 3.8.
 2016-04-05 00:48:41 +00:00
whitequark
ad83b092fa llvm_ir_generator: change !{→unconditionally_}dereferenceable. 
Since LLVM 3.8, !dereferenceable is weaker, so we introduce
!unconditionally_dereferenceable (http://reviews.llvm.org/D18738)
to regain its functionality.
 2016-04-05 00:48:41 +00:00
whitequark
72b952e850 llvm_ir_generator: fix DICompileUnit.language. 2016-04-05 00:48:41 +00:00
whitequark
a57aabb3ea compiler: purge generated functions from backtraces. 2016-04-02 18:29:36 +00:00
whitequark
a1e98a4437 llvm_ir_generator: add TBAA metadata for @now. 2016-04-02 18:19:14 +00:00
whitequark
b1f371e578 llvm_ir_generator: update debug info emission for LLVM 3.8. 2016-04-02 18:05:40 +00:00
whitequark
df8018f0be compiler: emit verbose assembly via ARTIQ_DUMP_ASM. 2016-04-01 09:34:51 +00:00
whitequark
d9dd7df3b2 compiler: add analysis passes from TargetMachine. 
This doesn't have any effect right now, but is the right thing to do.
 2016-04-01 09:34:36 +00:00
whitequark
10108e6d64 compiler: mark loaded pointers as !dereferenceable. 
Also, lower the bound for test_pulse_rate_dds, since we generate
better code for it now.
 2016-03-31 12:04:38 +00:00
whitequark
2c04979727 compiler: update for LLVM 3.7. 2016-03-31 11:23:47 +00:00
whitequark
b8bd344e54 compiler: use correct data layout. 2016-03-31 11:22:29 +00:00
whitequark
8a908a7656 llvm_ir_generator: do not use 'coldcc' calling convention. 
First, this calling convention doesn't actually exist in OR1K
and trying to use it in Asserts build causes an UNREACHABLE.

Second, I tried to introduce it and it does not appear to produce
any measurable benefit: not only OR1K has a ton of CSRs but also
it is quite hard, if not realistically impossible, to produce
the kind of register pressure that would be relieved by sparing
a few more CSRs for our exception raising function calls, since
temporaries don't have to be preserved before a noreturn call
and spilling over ten registers across an exceptional edge
is not something that the code we care about would do.

Third, it produces measurable drawbacks: it inflates code size
of check:* functions by adding spills. Of course, this could be
alleviated by making __artiq_raise coldcc as well, but what's
the point anyway?
 2016-03-29 15:19:46 +00:00
whitequark
586022023b llvm_ir_generator: mark loads as non-null where applicable. 2016-03-29 13:34:37 +00:00
whitequark
f6b391bb4e compiler: fix ARTIQ_DUMP_ELF. 2016-03-29 13:03:49 +00:00
whitequark
a9489246a4 llvm_ir_generator: don't mark non-constant attribute loads as invariant. 
Oops.
 2016-03-29 13:03:49 +00:00
whitequark
3ed852e077 Commit missing parts of 1d8b0d46. 2016-03-28 21:44:31 +00:00
whitequark
1038f1321f compiler: allow specifying per-function "fast-math" flags. 
Fixes #351.
 2016-03-28 21:44:08 +00:00
whitequark
6f5332f892 compiler: allow flagging syscalls, providing information to optimizer. 
This also fixes a crash in test_cache introduced in 1d8b0d46.
 2016-03-28 19:56:56 +00:00
whitequark
1d8b0d46bc compiler: mark FFI functions as ModRef=Ref using TBAA metadata. 
Fascinatingly, the fact that you can mark call instructions with
!tbaa metadata is completely undocumented. Regardless, it is true:
a !tbaa metadata for an "immutable" type will cause
AliasAnalysis::getModRefBehavior to return OnlyReadsMemory for that
call site.

Don't bother marking loads with TBAA yet since we already place
!load.invariant on them (which is as good as the TBAA "immutable"
flag) and after that we're limited by lack of !nonnull anyway.

Also, add TBAA analysis passes in our pipeline to actually engage it.
 2016-03-28 02:52:15 +00:00
whitequark
418f0a5234 compiler: mark loads of kernel constant attributes as load invariant. 
Also, enable LICM, since it can take advantage of this.
 2016-03-28 00:44:32 +00:00
whitequark
ca7463a054 compiler: do not write back kernel constant attributes. 
Fixes #322.
 2016-03-28 00:05:57 +00:00
whitequark
f4e6b18690 compiler: implement kernel constant attributes. 
Part of #322.
 2016-03-27 16:10:07 +00:00
whitequark
f81930ffd6 compiler: run IPSCCP. 
This doesn't do much, only frees some registers.
 2016-03-27 02:00:05 +00:00
whitequark
7213984330 compiler: raise inliner threshold to the equivalent of -O3. 2016-03-27 01:08:13 +00:00
whitequark
e75ad3d1aa compiler: extract runtime checks into separate cold functions. 
This reduces register pressure as well as function size, which
favorably affects the inliner.
 2016-03-27 01:02:15 +00:00
whitequark
20ad7629ac llvm_ir_generator: generate code more amenable to LLVM's GlobalOpt. 
This exposes almost all embedded methods to inlining, with massive
gains.
 2016-03-26 21:35:48 +00:00
whitequark
186a564ba8 compiler: make quoted functions independent of outer environment. 2016-03-26 20:46:45 +00:00
whitequark
f5c720c3ee compiler: tune the LLVM optimizer pipeline (fixes #315). 2016-03-26 13:20:23 +00:00
whitequark
3ee9834197 compiler: significantly increase readability of LLVM and ARTIQ IRs. 2016-03-26 12:21:02 +00:00
whitequark
8d0566661a compiler: quote functions directly instead of going through a local. 2016-03-25 22:35:55 +00:00
whitequark
39d23793a4 transforms.llvm_ir_generator: extract class function attributes. 
This should give LLVM more visibility.
 2016-03-25 19:01:39 +00:00
whitequark
39599d4508 compiler: get rid of the GetConstructor opcode. 2016-03-25 19:01:39 +00:00
whitequark
f4ab507ab2 Bring back target print function. 2016-03-18 04:42:06 +00:00
whitequark
ac5061c205 validators.escape: track region of arguments. 
Fixes #232.
 2016-03-18 03:08:14 +00:00
whitequark
9492464ed9 Commit missing parts of 78fa5be. 2016-03-18 02:15:36 +00:00
whitequark
78fa5becea compiler: refuse to embed a function from another core device. 
Fixes #332.
 2016-03-18 02:01:14 +00:00
whitequark
82ab21dbd5 transforms.inferencer: better message for BinOpT unification. 
Fixes #311.
 2016-03-18 01:22:01 +00:00
whitequark
73bfbe51db compiler: reject lambdas used as kernel functions (fixes #313). 2016-03-03 08:33:28 +00:00
whitequark
7e16da4a77 transforms.llvm_ir_generator: ignore assignments of None (fixes #309). 2016-03-01 12:26:42 +00:00
whitequark
dc70029b91 transforms.asttyped_rewriter: set loc for ForT (fixes #302). 2016-03-01 05:22:12 +00:00
whitequark
8bbffab8c8 Fix tests. 2016-02-27 13:40:37 +00:00
whitequark
e421b22953 types.TypePrinter: don't waste screen space on empty attribute lists. 2016-02-27 13:29:47 +00:00
whitequark
82a8e819ac transforms.llvm_ir_generator: use private linkage instead of internal. 
This reduces the size of symbol tables (internal adds an STB_LOCAL
symbol, but private doesn't).
 2016-02-25 20:15:40 +00:00
whitequark
f838b8be49 compiler.embedding: cache attribute types (fixes #276). 2016-02-25 19:56:45 +00:00
whitequark
d899d7307e compiler.types: TDelay is always unifiable with self. 2016-02-25 19:56:12 +00:00
whitequark
919a49b6bc compiler: quell excessively detailed diagnostics. 2016-02-25 19:43:52 +00:00
whitequark
a1dd909bc4 Take alignment into account during attribute writeback (fixes #293). 2016-02-25 01:44:05 +00:00
whitequark
914bc9f360 artiq_run: allow running LLVM IR/bitcode files, even with trivial RPCs. 2016-02-24 22:50:45 +00:00
whitequark
da31d29897 compiler: fix ARTIQ_DUMP_*. 2016-02-24 21:43:46 +00:00