[RFC/WIP] Leverage compiler-rt for builtins support of llvm intrinsics #17344

vchuravy · 2016-07-08T22:15:27Z

Note:

This feature will remain unavailable until we compile LLVM with cmake and bump our llvm version to ~~3.8~~ 3.9

Reason

llvm will only build compiler-rt with cmake and turns building it with autotools explicitly off. We currently build llvm as a shlib and that only works in combination with cmake from version 3.8 upwards.

Compiler-rt (builtins) provides low-level support for llvm intrinsics that don't map to an appropriate cpu instructions.

builtins - a simple library that provides an implementation of the low-level target-specific hooks required by code generation and other runtime components. For example, when compiling for a 32-bit target, converting a double to a 64-bit unsigned integer is compiling into a runtime call to the "__fixunsdfdi" function. The builtins library provides optimized implementations of this and other low-level routines, either in target-independent C form, or as a heavily-optimized assembly.
builtins provides full support for the libgcc interfaces on supported targets and high performance hand tuned implementations of commonly used functions like __floatundidf in assembly that are dramatically faster than the libgcc implementations. It should be very easy to bring builtins to support a new target by adding the new routines needed by that target.

It provides fallbacks for llvm intrinsics on platfroms, where they don't map to an appropriate instruction. As an example we could use the cpu instructions to convert Float16 -> Float32 on any platform that supports f16c.

Caveats

Compiler-rt is build and distributed as a static library so we have to convert it into a shared library.
Tested under Linux x86_64 and the build system will need adjustments for other targets.
compiler-rt build system is quite special (see rusts build-system https://github.com/rust-lang/rust/blob/master/mk/rt.mk)
needs cmake and llvm 3.8

Todo

Buildsystem support for Windows/ARM/PPC/Mac
~~ Enable building with LLVM_USE_CMAKE = 0~~
Support USE_SYSTEM_LLVM = 1
Build per default

maleadt · 2016-07-08T22:18:43Z

Some timings on Float16 conversions, on current master:

julia> @benchmark Float16(1.0f0)
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     4.304 ns (0.00% GC)
  median time:      4.312 ns (0.00% GC)
  mean time:        4.342 ns (0.00% GC)
  maximum time:     7.633 ns (0.00% GC)

julia> @benchmark Float16(1.0)
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     4.306 ns (0.00% GC)
  median time:      4.313 ns (0.00% GC)
  mean time:        4.345 ns (0.00% GC)
  maximum time:     8.068 ns (0.00% GC)

julia> @benchmark Float32(Float16(1.0))
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     6.799 ns (0.00% GC)
  median time:      6.808 ns (0.00% GC)
  mean time:        7.015 ns (0.00% GC)
  maximum time:     12.528 ns (0.00% GC)

julia> @benchmark Float64(Float16(1.0))
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     6.787 ns (0.00% GC)
  median time:      6.792 ns (0.00% GC)
  mean time:        6.803 ns (0.00% GC)
  maximum time:     10.169 ns (0.00% GC)

julia> @benchmark Float16(1.0) + 1
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     6.799 ns (0.00% GC)
  median time:      6.808 ns (0.00% GC)
  mean time:        7.152 ns (0.00% GC)
  maximum time:     9.103 ns (0.00% GC)

And with this PR:

julia> @benchmark Float16(1.0f0)
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     1.368 ns (0.00% GC)
  median time:      1.371 ns (0.00% GC)
  mean time:        1.373 ns (0.00% GC)
  maximum time:     3.727 ns (0.00% GC)

julia> @benchmark Float16(1.0)
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     1.370 ns (0.00% GC)
  median time:      1.373 ns (0.00% GC)
  mean time:        1.375 ns (0.00% GC)
  maximum time:     4.262 ns (0.00% GC)

julia> @benchmark Float32(Float16(1.0))
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     1.370 ns (0.00% GC)
  median time:      1.373 ns (0.00% GC)
  mean time:        1.375 ns (0.00% GC)
  maximum time:     2.876 ns (0.00% GC)

julia> @benchmark Float64(Float16(1.0))
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     1.369 ns (0.00% GC)
  median time:      1.372 ns (0.00% GC)
  mean time:        1.374 ns (0.00% GC)
  maximum time:     3.667 ns (0.00% GC)

julia> @benchmark Float16(1.0) + 1
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     1000
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     1.367 ns (0.00% GC)
  median time:      1.371 ns (0.00% GC)
  mean time:        1.374 ns (0.00% GC)
  maximum time:     4.347 ns (0.00% GC)

yuyichao · 2016-07-08T22:22:17Z

base/float.jl

@@ -14,7 +14,7 @@ const NaN = NaN64
 ## conversions to floating-point ##
 convert(::Type{Float16}, x::Integer) = convert(Float16, convert(Float32,x))
 for t in (Int8,Int16,Int32,Int64,Int128,UInt8,UInt16,UInt32,UInt64,UInt128)
-    @eval promote_rule(::Type{Float16}, ::Type{$t}) = Float32
+    @eval promote_rule(::Type{Float16}, ::Type{$t}) = Float16


This should probably be a different PR

Yes I should have noted that this is #17297

vchuravy · 2016-07-09T19:52:08Z

@tkelman thanks for the review and I hope I addressed most of your comments.

One point that I would like to discuss is if we want to put the shared library version of compiler-rt to go into build_private_libdir, since we create it purely to be able for Julia to resolve symbols to it at runtime and there is no other consumer (clang and rust link to the static version) and we need to distribute it ourselves.

tkelman · 2016-07-10T15:23:32Z

Why would we put it in build_private_libdir? That's not really on the search path, from what I can tell the only thing we put there is the system image.

vchuravy · 2016-07-10T16:30:17Z

The place I would think is appropriate is lib/julia to emphasise that we
are the only consumer of that shared library.

On Sun, 10 Jul 2016, 11:23 Tony Kelman, notifications@github.com wrote:

Why would we put it in build_private_libdir? That's not really on the
search path, from what I can tell the only thing we put there is the system
image.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#17344 (comment),
or mute the thread
https://github.com/notifications/unsubscribe/AAI3avfFAMs216vAF5OIfLKbleJXMgyzks5qUQ52gaJpZM4JIcMK
.

tkelman · 2016-07-10T16:37:20Z

Shared libraries will be put in lib/julia in the install tree on unix.

This allows us the more freely use llvm intrinsics and have fallbacks in place for systems where they don't map to instructions. Currently this requires building with a Make.user containing ``` override BUILD_COMPILER-RT = 1 override LLVM_USE_CMAKE = 1 override USE_LLVM_SHLIB = 0 ``` TODO: * Allow for a system installation * Figure out what is needed to make this work without LLVM_USE_CMAKE

LLVM intrinsics either map to instructions or to functions in compiler-rt. If we can't find a symbol look into a shared version of compiler-rt and resolve the functions there.

We need a cmake build that also works with SHLIB=1 for compiler-rt

vchuravy · 2016-07-11T22:18:03Z

I will leave this branch up for now, but I currently don't see a way of doing this properly (and not lose shlib support), without switching to 3.9

tkelman · 2016-07-11T22:23:39Z

deps/Versions.make

@@ -1,4 +1,4 @@
-LLVM_VER = 3.7.1
+LLVM_VER = 3.8.0


we'll probably do 3.8.1 some time not too long after branching for 0.6-dev, but it'll need CI preparation

edit: and checksum updates

vchuravy · 2016-09-02T21:16:37Z

For the build-system we could try to use the Makefile from https://github.com/ReservedField/arm-compiler-rt/ to make your lifes easier (not using cmake)

vchuravy · 2016-10-06T06:54:05Z

#18734 has the basic support for compiler-rt and I plan to add PR for full Float16 support after that got merged

yuyichao reviewed Jul 8, 2016
View reviewed changes

vchuravy added 12 commits July 12, 2016 06:02

teach llvm.mk to create compiler-rt

a3c651d

link sys.so against compiler-rt

a34bb7f

Load a shared version of compiler-rt in jitlayers

b504f99

LLVM intrinsics either map to instructions or to functions in compiler-rt. If we can't find a symbol look into a shared version of compiler-rt and resolve the functions there.

rename COMPILER-RT to COMPILER_RT

b96355d

fix patch for compiler-rt

d8d9508

use build_shlibdir

086c6f3

use llvm intrinsics to convert from and to Float16

e5f8f17

cleanup makefile for compiler-rt

86da367

enable compiler_rt to be build for USE_SYSTEM_LLVM

956b4fc

use CMAKE for llvm from 3.7 upwards

999577f

Bump llvm to 3.8

8bf61ad

We need a cmake build that also works with SHLIB=1 for compiler-rt

tkelman reviewed Jul 11, 2016
View reviewed changes

vchuravy mentioned this pull request Jul 15, 2016

integrate Float16 into stdlib along side Float32, Float64 #17440

Closed

vtjnash mentioned this pull request Jul 20, 2016

LLVM checked arithmetic for 128-bit ints is broken #4905

Closed

This was referenced Sep 23, 2016

Move Float16 functions away from float16.jl to places where other flo… #18287

Merged

Basic support for builtins from compiler-rt #18734

Closed

vchuravy closed this Oct 6, 2016

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[RFC/WIP] Leverage compiler-rt for builtins support of llvm intrinsics #17344

[RFC/WIP] Leverage compiler-rt for builtins support of llvm intrinsics #17344

vchuravy commented Jul 8, 2016 •

edited

Loading

maleadt commented Jul 8, 2016

yuyichao Jul 8, 2016

vchuravy Jul 8, 2016

vchuravy commented Jul 9, 2016

tkelman commented Jul 10, 2016

vchuravy commented Jul 10, 2016

tkelman commented Jul 10, 2016

vchuravy commented Jul 11, 2016

tkelman Jul 11, 2016 •

edited

Loading

vchuravy commented Sep 2, 2016

vchuravy commented Oct 6, 2016

[RFC/WIP] Leverage compiler-rt for builtins support of llvm intrinsics #17344

[RFC/WIP] Leverage compiler-rt for builtins support of llvm intrinsics #17344

Conversation

vchuravy commented Jul 8, 2016 • edited Loading

Note:

Reason

Caveats

Todo

maleadt commented Jul 8, 2016

yuyichao Jul 8, 2016

Choose a reason for hiding this comment

vchuravy Jul 8, 2016

Choose a reason for hiding this comment

vchuravy commented Jul 9, 2016

tkelman commented Jul 10, 2016

vchuravy commented Jul 10, 2016

tkelman commented Jul 10, 2016

vchuravy commented Jul 11, 2016

tkelman Jul 11, 2016 • edited Loading

Choose a reason for hiding this comment

vchuravy commented Sep 2, 2016

vchuravy commented Oct 6, 2016

vchuravy commented Jul 8, 2016 •

edited

Loading

tkelman Jul 11, 2016 •

edited

Loading