[WIP/RFC/RFH] Implement compiler-rt libcalls in Julia #18927

vchuravy · 2016-10-14T12:53:41Z

This is an alternative proposal to #18734 (and still compatible with it)

The idea behind using compiler-rt was for me that it would allow us to use LLVM intrinsics and if those resolve to libcalls (as happens if there is no processor instruction) have a viable fallback.
This is still the case, but there can be an argument made for the case that we would want to be independent of compiler-rt. One of the major drawback of compiler-rt is that it does not support 128bit on 32bit targets, but we currently do.

This uses the current Julia implementations to form the basis of a pure Julia runtime library. (We are still missing a lot of functions).
These are resolved late (and only if required by LLVM). The benefit is that the Float16 implementation becomes more like the way we handle Float32 and Float64.
Also Float128 support is straight forward to implement on all platforms.

I am focusing right now on the conversion functions between floating point numbers and integers, but the list of compiler-rt functions is quite a big longer. (see https://github.com/llvm-mirror/llvm/blob/master/lib/CodeGen/TargetLoweringBase.cpp)

Todo:

Decided if we need to prefix our implementations with jl
Tests! Tests!
PR to BaseBenchmarks with benchmarks for all rtlib functions

@vtjnash Is that the correct usage of jl_extern_c

I would also like to ask for help for the implementation of what we currently are missing.

vtjnash · 2016-10-14T20:17:16Z

This looks cool. I think this is a correct usage of jl_extern_c, although I think there are some edge cases where it isn't consistent handed right now (specifically, it is a imperative call rather than declarative, so it doesn't deserialize correctly).

Instead of falling back onto compiler-rt we provide our own implementations. Using compiler-rt is non-trivial and it does not offer support for 128bit on 32bit platforms. This way we get the benefits of calling the LLVM intrinsics (possible native intstructions) and we are can be platform agnostic.

LLVM intrinsics either map to instructions or to functions in compiler-rt. Since we provide our own implementation we can just look them up in sys.so and resolve to the function there. On Darwin we have to use a unmangled version of the function name.

… see them

As of compiler-rt 3.9.0 the testsuite uses compareResultH to check for correctness. That function uses uint16_t under the hood instead of checking the full 32bit.

vchuravy · 2016-10-17T02:05:56Z

I have been thinking about the design of RTLIB and I was wondering whether it is a good idea for it to be part of sys.so or if we should have rtlib.so. The hierarchy would be Core -> RTLIB -> Base.

@vtjnash Do you have an idea where the test failure on Mac is coming from? It seems to be a mangling issue.

tkelman · 2016-10-17T14:12:08Z

base/rtlib.jl

+    ccall(:jl_extern_c, Void, (Any, Any, Any, Cstring),
+          f, rtype, argt, name)
+
+include("rtlib/fp_util.jl")


I think we're moving towards having submodules be in their own folders under base, including their top-level files - so rtlib/RTLIB.jl

This might be a bit tricky because I think relative include only becomes available a bit later.

julia/base/sysimg.jl

Line 228 in 2618d7d

INCLUDE_STATE = 2 # include = _include (from lines above)

I will see if I can move that to an earlier point.

tkelman · 2016-10-17T14:14:17Z

base/rtlib/fp_extend.jl

+#
+#                     The LLVM Compiler Infrastructure
+#
+# This file is dual licensed under the MIT and the University of Illinois Open


anything copied from LLVM should be acknowledged in the top-level LICENSE.md and added to the exception list in contrib/add_license_to_files.jl

tkelman · 2016-10-17T14:16:45Z

test/rtlib.jl

+    # Note:
+    # These tests are taken fromt the compiler-rt testsuite. Were as of 3.9.0
+    # the test are done with compareResultH (so with after casting to UInt16)
+    # Tests that are commented out fail as === Float32 comparisons.


use @test_broken then?

tkelman · 2016-10-17T14:16:54Z

test/rtlib.jl

+    # These tests are taken fromt the compiler-rt testsuite. Were as of 3.9.0
+    # the test are done with compareResultH (so with after casting to UInt16)
+    # Tests that are commented out fail as === Float32 comparisons.
+    # Some tests succedd with ≈ (and are consistent with Julia v0.5 convert)


tkelman · 2016-10-17T14:17:53Z

base/rtlib/fp_extend.jl

+#                     The LLVM Compiler Infrastructure
+#
+# This file is dual licensed under the MIT and the University of Illinois Open
+# Source Licenses. See LICENSE.TXT for details.


we don't have a LICENSE.TXT in this repo, so this is a bit confusing

In the earlier phases of building the sysimg the relative form of `include` is not yet available.

vchuravy · 2016-10-18T08:56:58Z

base/rtlib/RTLIB.jl

+          f, rtype, argt, name)
+
+# Check if relative include is available
+if isdefined(Base, :INCLUDE_STATE) && Base.INCLUDE_STATE == 1


@tkelman Is that an appropriate solution? We can only include files relative to ourselves once Base.INCLUDE_STATE == 2 (and I think this is the only way of testing if we are currently building the sysimg).

Oh, strange. Dunno if the organization is worth this complexity, but I'll let others give opinions.

staticfloat · 2017-01-01T01:59:20Z

I tried building this just for fun to test out one of my 32-bit Docker images. It broke during bootstrap, with the following message:

rtlib/RTLIB.jl
WARNING: Unable to load sysimage
LLVM ERROR: Program used external function '__gnu_h2f_ieee' which could not be resolved!
*** This error is usually fixed by running `make clean`. If the error persists, try `make cleanall`. ***
Makefile:228: recipe for target '/src/julia/usr/lib/julia/sys.o' failed
make[1]: *** [/src/julia/usr/lib/julia/sys.o] Error 1
Makefile:99: recipe for target 'julia-sysimg-release' failed
make: *** [julia-sysimg-release] Error 2

Note that this Docker image passes all tests on a vanilla build of master. Here're the steps I performed to test this: this should be all you need to test locally, assuming you have docker installed:

$ docker run -ti staticfloat/julia_workerbase_ubuntu16_04:x86
root@004e6d44457c# git clone -b vc/jlrt https://github.com/JuliaLang/julia /src/julia
root@004e6d44457c# cd /src/julia; make VERBOSE=1 -j8

vchuravy · 2017-01-01T15:12:04Z

Thanks for taking this for a spin (and setting up a docker image) the error you see is the added test case and it means that symbol lookup is not finding the Julia rtlib version of the function (probably name mangling related).

…

On Sun, 1 Jan 2017, 10:59 Elliot Saba, ***@***.***> wrote: I tried building this just for fun to test out one of my 32-bit Docker images. It broke during bootstrap, with the following message: rtlib/RTLIB.jl WARNING: Unable to load sysimage LLVM ERROR: Program used external function '__gnu_h2f_ieee' which could not be resolved! *** This error is usually fixed by running `make clean`. If the error persists, try `make cleanall`. *** Makefile:228: recipe for target '/src/julia/usr/lib/julia/sys.o' failed make[1]: *** [/src/julia/usr/lib/julia/sys.o] Error 1 Makefile:99: recipe for target 'julia-sysimg-release' failed make: *** [julia-sysimg-release] Error 2 Note that this Docker image passes all tests on a vanilla build of master. Here're the steps I performed to test this: this should be all you need to test locally, assuming you have docker installed: $ docker run -ti staticfloat/julia_workerbase_ubuntu16_04:x86 ***@***.***# cd /src # git clone -b vc/jlrt https://github.com/JuliaLang/julia # cd julia # make VERBOSE=1 -j8 — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#18927 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAI3aqn4ZUJ-lsA4_VuP53bHzLa0oXFUks5rNwh5gaJpZM4KW-d8> .

JeffBezanson · 2017-07-28T20:31:18Z

This is a cool idea but I'm concerned about re-implementing (and testing and maintaining) a big library that already exists elsewhere. Of course we prefer code to be in julia, but I'm not sure the benefit is worth it for a library like this.

vchuravy · 2017-07-29T07:55:56Z

I agree that this is an important concern (and one of the reason I haven't moved on this in a while) My main motivation is to support hardware Float128 on architectures that support it and software Float128 everywhere else. AFAIK there is no C library that supports Float128 on 32bit x86 (certainly compiler-rt doesn't) and the actual library surface we would have to port is quite small and naturally extends to any IEEE754 float.

…

On Fri, 28 Jul 2017, 22:31 Jeff Bezanson, ***@***.***> wrote: This is a cool idea but I'm concerned about re-implementing (and testing and maintaining) a big library that already exists elsewhere. Of course we prefer code to be in julia, but I'm not sure the benefit is worth it for a library like this. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#18927 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAI3asAFApZM6C-6xSn6KoFhtQPN9Pgjks5sSkUYgaJpZM4KW-d8> .

tkelman · 2017-07-29T08:02:02Z

doesn't libquadmath? though we have some alignment and ABI issues to fix before that works properly everywhere IIRC (ref https://github.com/simonbyrne/Quadmath.jl)

vchuravy · 2017-07-29T11:54:48Z

Ah yes I didn't look at GPL licensed packages to prevent license incumbent.

vchuravy · 2018-03-09T00:31:43Z

superseded by #26381

vchuravy added 11 commits October 16, 2016 11:10

lower Float16 to IEEEhalf

7157e28

resolve libcalls to sys.so

f62b59b

LLVM intrinsics either map to instructions or to functions in compiler-rt. Since we provide our own implementation we can just look them up in sys.so and resolve to the function there. On Darwin we have to use a unmangled version of the function name.

refactor float.jl and use intrinsics for Float16

9654b7f

Try registering the function outside the module so that everybody can…

51458ec

… see them

put prototypes in place for Float128 support

2c5068e

add convert alias to RTLIB

eafab63

add first round of tests

d5bd807

Compiler-rt adapt test extendhfsf2

475cd37

As of compiler-rt 3.9.0 the testsuite uses compareResultH to check for correctness. That function uses uint16_t under the hood instead of checking the full 32bit.

port fp_extend from compiler_rt

7a2b598

port fp_trunc from compiler-rt

2618d7d

vchuravy force-pushed the vc/jlrt branch 2 times, most recently from e261329 to 2618d7d Compare October 16, 2016 04:12

vchuravy mentioned this pull request Oct 17, 2016

Basic support for builtins from compiler-rt #18734

Closed

5 tasks

tkelman reviewed Oct 17, 2016

View reviewed changes

vchuravy added 2 commits October 18, 2016 10:45

first pass for fixint

70fd6b2

move rtlib.jl into its own subdirectory

282e1dc

In the earlier phases of building the sysimg the relative form of `include` is not yet available.

vchuravy commented Oct 18, 2016

View reviewed changes

vchuravy mentioned this pull request Oct 19, 2016

hardware independent isinf/isnan/isfinite #19008

Closed

vchuravy mentioned this pull request Oct 28, 2016

[RFC/WIP] Intrinsics for NTuple{N, VecElement} #18470

Closed

3 tasks

vchuravy mentioned this pull request Nov 4, 2016

RTLIB: A runtime-library for Julia JuliaLang/Juleps#11

Merged

tkelman mentioned this pull request Jan 1, 2017

temporarily disable LLVM assertions on travis #19803

Closed

Keno mentioned this pull request Jul 28, 2017

[0.6] LLVM intrinsics are looked up in the wrong place #23009

Closed

vchuravy mentioned this pull request Mar 9, 2018

Switch Float16 LLVM representation from i16 to half #26381

Closed

vchuravy closed this Mar 9, 2018

maleadt mentioned this pull request Sep 10, 2020

Switch Float16 to LLVM's half #37510

Merged

DilumAluthge deleted the vc/jlrt branch March 25, 2021 22:10

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP/RFC/RFH] Implement compiler-rt libcalls in Julia #18927

[WIP/RFC/RFH] Implement compiler-rt libcalls in Julia #18927

vchuravy commented Oct 14, 2016 •

edited

Loading

vtjnash commented Oct 14, 2016

vchuravy commented Oct 17, 2016

tkelman Oct 17, 2016

vchuravy Oct 18, 2016

tkelman Oct 17, 2016

tkelman Oct 17, 2016

tkelman Oct 17, 2016

tkelman Oct 17, 2016 •

edited

Loading

vchuravy Oct 18, 2016

tkelman Oct 18, 2016

staticfloat commented Jan 1, 2017 •

edited

Loading

vchuravy commented Jan 1, 2017 via email

JeffBezanson commented Jul 28, 2017

vchuravy commented Jul 29, 2017 via email

tkelman commented Jul 29, 2017 •

edited

Loading

vchuravy commented Jul 29, 2017

vchuravy commented Mar 9, 2018

[WIP/RFC/RFH] Implement compiler-rt libcalls in Julia #18927

[WIP/RFC/RFH] Implement compiler-rt libcalls in Julia #18927

Conversation

vchuravy commented Oct 14, 2016 • edited Loading

Todo:

vtjnash commented Oct 14, 2016

vchuravy commented Oct 17, 2016

tkelman Oct 17, 2016

Choose a reason for hiding this comment

vchuravy Oct 18, 2016

Choose a reason for hiding this comment

tkelman Oct 17, 2016

Choose a reason for hiding this comment

tkelman Oct 17, 2016

Choose a reason for hiding this comment

tkelman Oct 17, 2016

Choose a reason for hiding this comment

tkelman Oct 17, 2016 • edited Loading

Choose a reason for hiding this comment

vchuravy Oct 18, 2016

Choose a reason for hiding this comment

tkelman Oct 18, 2016

Choose a reason for hiding this comment

staticfloat commented Jan 1, 2017 • edited Loading

vchuravy commented Jan 1, 2017 via email

JeffBezanson commented Jul 28, 2017

vchuravy commented Jul 29, 2017 via email

tkelman commented Jul 29, 2017 • edited Loading

vchuravy commented Jul 29, 2017

vchuravy commented Mar 9, 2018

vchuravy commented Oct 14, 2016 •

edited

Loading

tkelman Oct 17, 2016 •

edited

Loading

staticfloat commented Jan 1, 2017 •

edited

Loading

tkelman commented Jul 29, 2017 •

edited

Loading