codegen restructuring #25984

vtjnash · 2018-02-09T22:19:42Z

This creates the function jl_create_native for outputting a llvm::Module from a list (jl_array_t) of functions (jl_method_instance_t) which is independent from the JIT (as needed by JuliaGPU/CUDAnative.jl#122). It's now used by the build system also, so that it'll be tested too. This (mostly) eliminates the shadow_module, and does a slightly more aggressive job at ensuring functions are getting precompiled.

There's still some more work to do to make that work even better, but that can be done later. I'm also now printing out errors that happen as a result of codegen bugs, rather than propagating them to the user, which makes a few of the tests (for codegen breakage) are a bit noisier now – those will need to be fixed now.

fixes #26767
fixes #26714

src/jitlayers.h

maleadt · 2018-02-10T11:40:09Z

Cool! How should I access the call targets from CUDAnative in order to build a module with all functions defined?

vtjnash · 2018-02-12T18:34:25Z

How should I access the call targets from CUDAnative
Value::getContext

Feel free to add exports from aotlayers.cpp to expose these with C wrappers. I know jl_create_native needs some extra arguments (specifically, cgparams) too.

maleadt · 2018-02-13T12:42:45Z

Feel free to add exports from aotlayers.cpp to expose these with C wrappers. I know jl_create_native needs some extra arguments (specifically, cgparams) too.

OK. Could you describe the behavior/indented use of jl_create_native, and the fields of jl_native_code_desc_t then? Does jl_create_native compile all necessary functions (ie. no trampolines), without using a cache? And should I still bother with a module_setup hook, or just re-configure the datalayout/triple as jl_dump_native does?

eg. does this look sane jn/codegen-norecursion...tb/codegen-norecursion

vtjnash · 2018-02-13T19:09:43Z

And should I still bother with a module_setup

I'm thinking of moving the creation of the Module into the caller, to sidestep this question entirely.

maleadt · 2018-02-15T13:23:31Z

What about my other questions?

Also, I was using julia_type_to_llvm in generated functions in combination with LLVM.jl. This function now takes a codegen context, which a) isn't available from C/Julia, and b) doesn't exist yet at the time the LLVM.jl-based generated function runs. I guess that b) will become even worse once there isn't a global LLVM context anymore (which I assume is the direction you're going in).
Any comments on how to do this style of LLVM-based custom codegen now?

vtjnash · 2018-02-15T21:54:36Z

I was using julia_type_to_llvm in generated functions in combination with LLVM.jl. This function now takes a codegen context

That other method is not exported – I didn't change the signature of julia_type_to_llvm (yet 😛 – it might gain a LLVMContextRef eventually).

will become even worse once there isn't a global LLVM context anymore (which I assume is the direction you're going in). Any comments on how to do this style of LLVM-based custom codegen now?

Yes, probably true. But I don't really expect jl_create_native will be exactly right for your use case long-term anyways, but will just provides a sample usage. Eventually, I expect you'll just provide accessors from this file to be able to do ccalls to do everything that it can do.

maleadt · 2018-02-16T08:51:56Z

That other method is not exported – I didn't change the signature of julia_type_to_llvm (yet 😛 – it might gain a LLVMContextRef eventually).

~~Ah OK, I didn't see the second definition.~~ You renamed it to jl_type_to_llvm!
An explicit LLVM context should be fine, assuming we can maintain a global one in CUDAnative.

Also, renaming julia_gpu_accumulate!_146195 to julia_gpu_accumulate%NOT_146195 seems pretty... weird.

But this passes CUDAnative tests so 👍 from me (with the additions from jn/codegen-norecursion...tb/codegen-norecursion, that is).

iblislin · 2018-05-10T10:39:25Z

This FreeBSD CI build already ran more than 9 hrs: https://freebsdci.julialang.org/#/builders/1/builds/9298
I manually killed it.

vtjnash · 2018-05-10T17:36:24Z

@nanosoldier runbenchmarks(ALL, vs = ":master")

nanosoldier · 2018-05-10T23:03:24Z

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan

Ref JuliaLang/julia#25984

maleadt · 2020-02-27T11:33:55Z

src/codegen.cpp

 extern JITEventListener *CreateJuliaJITEventListener();

 // for image reloading
 bool imaging_mode = false;

+// shared llvm state
+static LLVMContext &jl_LLVMContext = *(new LLVMContext());


Suggested change

static LLVMContext &jl_LLVMContext = *(new LLVMContext());

JL_DLLEXPORT LLVMContext &jl_LLVMContext = *(new LLVMContext());

One of the purpose of this change is to eliminate this state object. (There's a function accessor for it in the aotcompiler file)

OK, so I can't be assuming a global state anymore starting with this PR? That'll require some more work on my side then 🙂

src/aotcompile.cpp

Ref JuliaLang/julia#25984

also provides support for using a different code_native format, as a fallback, later we'll want to make this more configurable there are now several primary interfaces to native code: - codegen: mostly internal, support for translating IR to LLVM - jitlayers: manages runtime codegen results and executable memory - aotcompile: support for managing external code output - disasm: pretty-printer for code objects - debuginfo: tracking for unwind info also removes the global type caches and move all codegen pass handling to aotcompile.cpp

maleadt · 2020-03-02T12:38:30Z

A non-global LLVM context (but jl_get_llvm_context to get one from the AOT compiler output) won't currently work for the GPU compiler, because we implement generated functions with LLVM.jl by returning a llvm::Function* that relies on the Julia context. Do you have a short-term plan for that? If not, maybe we should keep the context (which is still global anyway) exported until we have a solution.

Ref JuliaLang/julia#25984

also provides support for using a different code_native format, as a fallback, later we'll want to make this more configurable there are now several primary interfaces to native code: - codegen: mostly internal, support for translating IR to LLVM - jitlayers: manages runtime codegen results and executable memory - aotcompile: support for managing external code output - disasm: pretty-printer for code objects - debuginfo: tracking for unwind info also removes the global type caches and move all codegen pass handling to aotcompile.cpp

Other bad cases are probably less common, but still exist. The code to fix those is included in #25984. Fixes #34680

also provides support for using a different code_native format, as a fallback, later we'll want to make this more configurable there are now several primary interfaces to native code: - codegen: mostly internal, support for translating IR to LLVM - jitlayers: manages runtime codegen results and executable memory - aotcompile: support for managing external code output - disasm: pretty-printer for code objects - debuginfo: tracking for unwind info also removes the global type caches and move all codegen pass handling to aotcompile.cpp

Gnimuc · 2020-04-15T15:40:21Z

src/codegen.cpp

-
-
-extern "C" JL_DLLEXPORT
-void *jl_get_llvmf_decl(jl_method_instance_t *mi, size_t world, bool getwrapper, const jl_cgparams_t params)


Is there an alternative function to use instead of this one? It was used in Cxx.jl: https://github.com/JuliaInterop/Cxx.jl/blob/256485c04c76feee515e83f091ba0d6c8f3286ee/src/cxxstr.jl#L107

ararslan added the compiler:codegen Generation of LLVM IR and native code label Feb 9, 2018

ararslan requested review from Keno and JeffBezanson February 9, 2018 22:41

maleadt reviewed Feb 10, 2018

View reviewed changes

src/jitlayers.h Outdated Show resolved Hide resolved

vtjnash force-pushed the jn/codegen-norecursion branch 2 times, most recently from 029e6fb to 16f53fe Compare February 12, 2018 22:11

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Feb 13, 2018

WIP: use new codegen from JuliaLang/julia#25984

f781c93

maleadt mentioned this pull request Feb 13, 2018

Use new non-recursive codegen JuliaGPU/CUDAnative.jl#162

Merged

vtjnash force-pushed the jn/codegen-norecursion branch from 52ea79e to 3e4a182 Compare February 13, 2018 19:03

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Feb 15, 2018

WIP: use new codegen from JuliaLang/julia#25984

a79b36e

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Feb 16, 2018

Use the new non-recursive codegen from JuliaLang/julia#25984.

04ce564

vtjnash force-pushed the jn/codegen-norecursion branch 3 times, most recently from 43f5ab9 to 4da4982 Compare March 2, 2018 21:40

vtjnash force-pushed the jn/codegen-norecursion branch from 4da4982 to 4ac634d Compare March 18, 2018 03:50

vtjnash force-pushed the jn/codegen-norecursion branch from 4ac634d to 50be707 Compare April 13, 2018 17:08

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request May 7, 2018

Use the new non-recursive codegen from JuliaLang/julia#25984.

8b7e9d4

vtjnash force-pushed the jn/codegen-norecursion branch from 50be707 to 333da81 Compare May 9, 2018 23:57

vtjnash force-pushed the jn/codegen-norecursion branch from 333da81 to d28eaf3 Compare May 10, 2018 16:38

vtjnash force-pushed the jn/codegen-norecursion branch from d28eaf3 to 3c90511 Compare May 12, 2018 02:36

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Feb 27, 2020

Adapt to nonrecursive codegen. [ci skip]

171692e

Ref JuliaLang/julia#25984

maleadt reviewed Feb 27, 2020

View reviewed changes

src/aotcompile.cpp Outdated Show resolved Hide resolved

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Feb 27, 2020

Adapt to nonrecursive codegen. [ci skip]

0ecfa88

Ref JuliaLang/julia#25984

vtjnash force-pushed the jn/codegen-norecursion branch from 4fdefdf to 973aefe Compare February 27, 2020 17:10

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Mar 2, 2020

Adapt to nonrecursive codegen. [ci skip]

289489c

Ref JuliaLang/julia#25984

vtjnash merged commit 3b53f54 into master Mar 3, 2020

vtjnash deleted the jn/codegen-norecursion branch March 3, 2020 22:03

This was referenced Mar 4, 2020

Deal with changed runtime function. maleadt/LLVM.jl#171

Merged

Reexport the LLVM context. #34991

Merged

jl_create_native generates bad code for *(Float16,Float16) #34993

Closed

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Mar 4, 2020

Adapt to nonrecursive codegen.

34dc9b4

Ref JuliaLang/julia#25984

thomasfaingnaert mentioned this pull request Mar 4, 2020

julia_type_to_llvm is renamed maleadt/LLVM.jl#172

Closed

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Mar 5, 2020

Adapt to nonrecursive codegen.

daecaf0

Ref JuliaLang/julia#25984

tkf mentioned this pull request Mar 7, 2020

Fix: _dump_function_linfo not defined JuliaDebug/Cthulhu.jl#70

Merged

MasonProtter mentioned this pull request Mar 24, 2020

codegen-norecursion has merged tshort/StaticCompiler.jl#28

Closed

tshort mentioned this pull request Mar 24, 2020

StaticCompiler won't start on master tshort/StaticCompiler.jl#29

Closed

maleadt added a commit to JuliaGPU/CUDAnative.jl that referenced this pull request Mar 26, 2020

Adapt to nonrecursive codegen.

9d22860

Ref JuliaLang/julia#25984

KristofferC pushed a commit that referenced this pull request Apr 11, 2020

ccall: avoid some bad symbol mangling

3df21fb

Other bad cases are probably less common, but still exist. The code to fix those is included in #25984. Fixes #34680

Gnimuc reviewed Apr 15, 2020

View reviewed changes

KristofferC mentioned this pull request Apr 21, 2020

Outputting a @ccallable function to an object file broken on master #35014

Closed

maleadt mentioned this pull request Jun 16, 2020

remove more global state from codegen #36260

Merged

martinholters mentioned this pull request Aug 20, 2020

Fix llvmcall in uninferred code #31006

Closed

Gnimuc mentioned this pull request Oct 3, 2020

Works with Julia 1.3 or older JuliaInterop/Cxx.jl#484

Merged

maleadt mentioned this pull request Nov 17, 2020

Codegen nondeterminism based on inference cache (emit_invoke) #38468

Open

t-bltg mentioned this pull request Jun 5, 2021

Remove deprecated --output-jit-bc option #41100

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

codegen restructuring #25984

codegen restructuring #25984

vtjnash commented Feb 9, 2018 •

edited

Loading

maleadt commented Feb 10, 2018

vtjnash commented Feb 12, 2018

maleadt commented Feb 13, 2018 •

edited

Loading

vtjnash commented Feb 13, 2018

maleadt commented Feb 15, 2018

vtjnash commented Feb 15, 2018

maleadt commented Feb 16, 2018 •

edited

Loading

iblislin commented May 10, 2018

vtjnash commented May 10, 2018

nanosoldier commented May 10, 2018

maleadt Feb 27, 2020

vtjnash Feb 27, 2020

maleadt Feb 27, 2020 •

edited

Loading

maleadt commented Mar 2, 2020

Gnimuc Apr 15, 2020

	static LLVMContext &jl_LLVMContext = *(new LLVMContext());
	JL_DLLEXPORT LLVMContext &jl_LLVMContext = *(new LLVMContext());



		extern "C" JL_DLLEXPORT
		void jl_get_llvmf_decl(jl_method_instance_t mi, size_t world, bool getwrapper, const jl_cgparams_t params)

codegen restructuring #25984

codegen restructuring #25984

Conversation

vtjnash commented Feb 9, 2018 • edited Loading

maleadt commented Feb 10, 2018

vtjnash commented Feb 12, 2018

maleadt commented Feb 13, 2018 • edited Loading

vtjnash commented Feb 13, 2018

maleadt commented Feb 15, 2018

vtjnash commented Feb 15, 2018

maleadt commented Feb 16, 2018 • edited Loading

iblislin commented May 10, 2018

vtjnash commented May 10, 2018

nanosoldier commented May 10, 2018

maleadt Feb 27, 2020

Choose a reason for hiding this comment

vtjnash Feb 27, 2020

Choose a reason for hiding this comment

maleadt Feb 27, 2020 • edited Loading

Choose a reason for hiding this comment

maleadt commented Mar 2, 2020

Gnimuc Apr 15, 2020

Choose a reason for hiding this comment

vtjnash commented Feb 9, 2018 •

edited

Loading

maleadt commented Feb 13, 2018 •

edited

Loading

maleadt commented Feb 16, 2018 •

edited

Loading

maleadt Feb 27, 2020 •

edited

Loading