[AutoDiff] Mangle derivative functions and linear maps #35259

rxwei · 2021-01-05T02:56:53Z

Mangle derivative functions and linear maps.

Mangle::ASTMangler::mangleAutoDiffDerivativeFunction() and Mangle::ASTMangler::mangleAutoDiffLinearMap() accept original function declarations and return a mangled name for a derivative function or linear map. This is called during SILGen and TBDGen.
Mangle::DifferentiationMangler handles differentiation function mangling in the differentiation transform. This part is necessary because we need to perform demangling on the original function and remangle it as part of a differentiation function mangling tree in order to get the correct substitutions in the mangled derivative generic signature.

A mangled differentiation function name includes:

The original function.
The differentiation function kind.
The parameter indices for differentiation.
The result indices for differentiation.
The derivative generic signature.

global ::= global generic-signature? 'TJ' AUTODIFF-FUNCTION-KIND INDEX-SUBSET 'p' INDEX-SUBSET 'r' // autodiff function

AUTODIFF-FUNCTION-KIND ::= 'f'        // JVP (forward-mode derivative)
AUTODIFF-FUNCTION-KIND ::= 'r'        // VJP (reverse-mode derivative)
AUTODIFF-FUNCTION-KIND ::= 'd'        // differential
AUTODIFF-FUNCTION-KIND ::= 'p'        // pullback

INDEX-SUBSET ::= ('S' | 'U')+

Example:

> swift-demangle --compact s13test_mangling4foo21xq_x_t16_Differentiation14DifferentiableR_AA1P13TangentVectorRp_r0_lFAdERzAdER_AafGRpzAafHRQr0_lTJrSpSr
reverse-mode derivative of test_mangling.foo2<A, B where B: _Differentiation.Differentiable, B.TangentVector: test_mangling.P>(x: A) -> B with respect to parameters {0} and results {0} with <A, B where A: _Differentiation.Differentiable, B: _Differentiation.Differentiable, A.TangentVector: test_mangling.P, B.TangentVector: test_mangling.P>

Also:

Make AD-generated linear map structures' fields have private access level so that TBDGen won't generate symbols for their accessors. (SR-14014)

Resolves SR-13507.

dan-zheng

Nice!

rxwei · 2021-01-05T13:01:47Z

@swift-ci please test

rxwei · 2021-01-05T13:21:07Z

@swift-ci please smoke test

swift-ci · 2021-01-05T14:07:05Z

Build failed
Swift Test OS X Platform
Git Sha - d0ec73681f881d17420bb07eefaaffe58c29d138

rxwei · 2021-01-05T14:18:50Z

@swift-ci please test

swift-ci · 2021-01-05T14:41:39Z

Build failed
Swift Test OS X Platform
Git Sha - 3b65e15826745991db49c8a8063feba5b8193907

swift-ci · 2021-01-05T15:01:25Z

Build failed
Swift Test Linux Platform
Git Sha - 3b65e15826745991db49c8a8063feba5b8193907

rxwei · 2021-01-05T18:34:58Z

@swift-ci please smoke test

rxwei · 2021-01-05T18:35:07Z

@swift-ci please test

swift-ci · 2021-01-05T18:54:37Z

Build failed
Swift Test OS X Platform
Git Sha - 3b65e15826745991db49c8a8063feba5b8193907

swift-ci · 2021-01-05T20:19:40Z

Build failed
Swift Test Linux Platform
Git Sha - 3b65e15826745991db49c8a8063feba5b8193907

rxwei · 2021-01-05T23:58:19Z

@swift-ci please test

swift-ci · 2021-01-06T01:00:52Z

Build failed
Swift Test Linux Platform
Git Sha - 3b65e15826745991db49c8a8063feba5b8193907

rxwei · 2021-01-06T01:29:51Z

@swift-ci please test linux

swift-ci · 2021-01-06T02:23:20Z

Build failed
Swift Test OS X Platform
Git Sha - 3b65e15826745991db49c8a8063feba5b8193907

swift-ci · 2021-01-06T02:33:25Z

Build failed
Swift Test Linux Platform
Git Sha - 3b65e15826745991db49c8a8063feba5b8193907

rxwei · 2021-01-06T07:47:52Z

@swift-ci please test

eeckstein

Basically LGTM. But I have a few comments.

Two general comments:

Can you add tests to test/Demangle/Inputs/manglings.txt?
It's good that you split such a large PR into multiple commits. But ideally each commit should be self-contained, i.e. each commit should contain a good commit message and the project should be compilable after each commit, and ideally the tests should pass. This makes it easier to understand "git log" and it enables "git bisect".
Your commit messages "Make compile" and "Fix all tests" suggest that this is not the case here.

docs/ABI/Mangling.rst

lib/SILOptimizer/Utils/DifferentiationMangler.cpp

include/swift/Demangling/Demangle.h

lib/SILOptimizer/Utils/DifferentiationMangler.cpp

lib/Demangling/Demangler.cpp

lib/Demangling/NodePrinter.cpp

lib/SILGen/SILGenThunk.cpp

- `Mangle::ASTMangler::mangleAutoDiffDerivativeFunction()` and `Mangle::ASTMangler::mangleAutoDiffLinearMap()` accept original function declarations and return a mangled name for a derivative function or linear map. This is called during SILGen and TBDGen. - `Mangle::DifferentiationMangler` handles differentiation function mangling in the differentiation transform. This part is necessary because we need to perform demangling on the original function and remangle it as part of a differentiation function mangling tree in order to get the correct substitutions in the mangled derivative generic signature. A mangled differentiation function name includes: - The original function. - The differentiation function kind. - The parameter indices for differentiation. - The result indices for differentiation. - The derivative generic signature.

rxwei · 2021-01-07T10:24:58Z

It's good that you split such a large PR into multiple commits. But ideally each commit should be self-contained, i.e. each commit should contain a good commit message and the project should be compilable after each commit, and ideally the tests should pass. This makes it easier to understand "git log" and it enables "git bisect".
Your commit messages "Make compile" and "Fix all tests" suggest that this is not the case here.

Sorry, the commit log in the PR may appear confusing, but I did not intend to merge these commits as is. My intention is to keep the original WIP commit by @dan-zheng to acknowledge his original contribution in the PR. When I merge I always use squash-and-merge, so only one commit will go into the tree. Nonetheless, I have squashed those commits.

rxwei · 2021-01-07T10:33:12Z

@swift-ci please test

swift-ci · 2021-01-07T12:05:53Z

Build failed
Swift Test Linux Platform
Git Sha - ffe6064

rxwei · 2021-01-07T12:14:53Z

@swift-ci please test linux

eeckstein

Thanks, LGTM!

rxwei requested review from jckarter, joshnewnham, dan-zheng, marcrasi and eeckstein January 5, 2021 02:56

dan-zheng approved these changes Jan 5, 2021

View reviewed changes

rxwei force-pushed the autodiff-mangling branch from 769d91c to d0ec736 Compare January 5, 2021 13:01

rxwei marked this pull request as ready for review January 5, 2021 13:01

rxwei force-pushed the autodiff-mangling branch from d0ec736 to 3b65e15 Compare January 5, 2021 13:20

rxwei requested a review from CodaFi January 5, 2021 18:37

rxwei force-pushed the autodiff-mangling branch from 3b65e15 to 6e1df19 Compare January 6, 2021 07:47

eeckstein requested changes Jan 6, 2021

View reviewed changes

rxwei force-pushed the autodiff-mangling branch 2 times, most recently from c1858a3 to 72b165a Compare January 7, 2021 10:12

rxwei force-pushed the autodiff-mangling branch from 72b165a to ffe6064 Compare January 7, 2021 10:21

eeckstein self-requested a review January 7, 2021 13:18

eeckstein approved these changes Jan 7, 2021

View reviewed changes

rxwei merged commit 5e5f484 into swiftlang:main Jan 7, 2021

rxwei deleted the autodiff-mangling branch January 7, 2021 16:46

This was referenced Jan 5, 2021

[SR-13507] Implement demangling/remangling for AutoDiff generated functions #53606

Closed

[SR-14014] [AD] Nonexistent symbols for AD linear map struct property accessors #56407

Closed

[AutoDiff] Mangle derivative functions and linear maps #35259

[AutoDiff] Mangle derivative functions and linear maps #35259

Uh oh!

Conversation

rxwei commented Jan 5, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dan-zheng left a comment

Choose a reason for hiding this comment

Uh oh!

rxwei commented Jan 5, 2021

Uh oh!

rxwei commented Jan 5, 2021

Uh oh!

swift-ci commented Jan 5, 2021

Uh oh!

rxwei commented Jan 5, 2021

Uh oh!

swift-ci commented Jan 5, 2021

Uh oh!

swift-ci commented Jan 5, 2021

Uh oh!

rxwei commented Jan 5, 2021

Uh oh!

rxwei commented Jan 5, 2021

Uh oh!

swift-ci commented Jan 5, 2021

Uh oh!

swift-ci commented Jan 5, 2021

Uh oh!

rxwei commented Jan 5, 2021

Uh oh!

swift-ci commented Jan 6, 2021

Uh oh!

rxwei commented Jan 6, 2021

Uh oh!

swift-ci commented Jan 6, 2021

Uh oh!

swift-ci commented Jan 6, 2021

Uh oh!

rxwei commented Jan 6, 2021

Uh oh!

eeckstein left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

rxwei commented Jan 7, 2021

Uh oh!

rxwei commented Jan 7, 2021

Uh oh!

swift-ci commented Jan 7, 2021

Uh oh!

rxwei commented Jan 7, 2021

Uh oh!

eeckstein left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

rxwei commented Jan 5, 2021 •

edited

Loading