Eliminate the set_tuple_element instruction #10144
Conversation
bjorng
added
team:VM
CT Test Results 4 files 457 suites 54m 23s ⏱️ Results for commit a144913. ♻️ This comment has been updated with latest results. To speed up review, make sure that you have read Contributing to Erlang/OTP and that all checks pass. See the TESTING and DEVELOPMENT HowTo guides for details about how to run test locally. Artifacts// Erlang/OTP Github Action Bot |
|
Do the sequences of |
No. Optimization of |
| @@ -1102,6 +1102,11 @@ void BeamModuleAssembler::emit_update_record_in_place( | |||
| void BeamModuleAssembler::emit_set_tuple_element(const ArgSource &Element, | |||
| const ArgRegister &Tuple, | |||
| const ArgWord &Offset) { | |||
| /* TODO: As of Erlang/OTP 29, this instruction is no longer | |||
| * emitted by the compiler. It can be removed when the runtime | |||
| * system no longer supports loading code compiled by Erlang/OTP | |||
| @@ -204,8 +204,6 @@ validate_0([{function, Name, Arity, Entry, Code} | Fs], Module, Level, Ft) -> | |||
| hf=0, | |||
| %% List of hot catch/try tags | |||
| ct=[], | |||
| %% Previous instruction was setelement/3. | |||
| setelem=false, | |||
There was a problem hiding this comment.
We can do a drive-by removal of puts_left too, it's not used anywhere.
The `set_tuple_element` instruction was used before OTP 26 to optimize chains of
`setelement/3` calls. For example:
foo(S0) when tuple_size(S0) =:= 8 ->
S1 = setelement(8, S0, a),
S2 = setelement(7, S1, b),
setelement(5, S2, c).
The compiler would keep the first call to `setelement/3` and replace
the other two with `set_tuple_element`, which would destructively
update the tuple created by the `setelement/3` call.
Starting in OTP 26, this was changed so that all three `setelement/3`
calls are replaced with a single `update_record` instruction.
The `set_tuple_element` instruction was only used when the size of the
tuple being updated was not known at compile time, as in the following
example:
bar(S0) ->
S1 = setelement(8, S0, a),
S2 = setelement(7, S1, b),
setelement(5, S2, c).
The only difference compared to the previous example is that the type
of `S0` is not known. When the size of the tuple cannot be determined
at compile time, the compiler cannot use `update_record` and must fall
back to `setelement/3` and `set_tuple_element`. However, the way this
fallback was implemented caused extra useless instructions to be
generated (see erlang#10125 for examples).
This commit eliminates the use of `set_tuple_element` in the fallback
path, allowing us to remove some particularly messy code from
`beam_validator`.
Resolves erlang#10125
bjorng
force-pushed
the
bjorn/compiler/eliminate-set_tuple_element/GH-10125/OTP-19751
branch
from
596432e to
a144913
Compare
bjorng
deleted the
bjorn/compiler/eliminate-set_tuple_element/GH-10125/OTP-19751
branch
The
set_tuple_elementinstruction was used before OTP 26 to optimize chains ofsetelement/3calls. For example:The compiler would keep the first call to
setelement/3and replace the other two withset_tuple_element, which would destructively update the tuple created by thesetelement/3call.Starting in OTP 26, this was changed so that all three
setelement/3calls are replaced with a singleupdate_recordinstruction.The
set_tuple_elementinstruction was only used when the size of the tuple being updated was not known at compile time, as in the following example:The only difference compared to the previous example is that the type of
S0is not known. When the size of the tuple cannot be determined at compile time, the compiler cannot useupdate_recordand must fall back tosetelement/3andset_tuple_element. However, the way this fallback was implemented caused extra useless instructions to be generated (see #10125 for examples).This commit eliminates the use of
set_tuple_elementin the fallback path, allowing us to remove some particularly messy code frombeam_validator.Resolves #10125