[Don't merge] docs: description of new mangling scheme #5428
Conversation
|
Some results (of a prototype implementation) on stdlib, benchmarks and a large framework: |
|
@rjmccall, @slavapestov (and all others): can you please take a look at this? |
|
|
||
| nominal-type-kind ::= 'C' // class | ||
| nominal-type-kind ::= 'O' // enum | ||
| nominal-type-kind ::= 'V' // struct |
There was a problem hiding this comment.
If we're going to redo the mangling, we should do away with differentiating nominal type kinds, IMO.
| type ::= 'XPM' metatype-repr type // existential metatype with representation | ||
| type ::= context identifier 'a' // Type alias (DWARF only) | ||
| type ::= type type 'c' THROWS-ANNOTATION? // function type | ||
| type ::= type type 'cu' THROWS-ANNOTATION? // uncurried function type |
There was a problem hiding this comment.
Function types should have a variable number of input arguments, instead of a single tuple argument. "Variadic" and "inout" should be argument modifiers on function types instead of separate type encodings.
| values indicates a single generic parameter at the outermost depth:: | ||
|
|
||
| q_q_cru // <T_0_0> T_0_0 -> T_0_0 | ||
| q_qd_0_cr_0_u // <T_0_0><T_1_0, T_1_1> T_0_0 -> T_1_1 |
There was a problem hiding this comment.
Having a single character encoding for t_0_0 (x in the old scheme) was a pretty significant size win, since a large number of generics have a single generic parameter. Is there room to do that in this scheme?
| encoded string itself. For example, the identifier ``vergüenza`` is mangled | ||
| to ``X12vergenza_JFa``. (The encoding in standard Punycode would be | ||
| to ``0012vergenza_JFa``. (The encoding in standard Punycode would be | ||
| ``vergenza-95a``) |
There was a problem hiding this comment.
There's a bug with this mangling scheme for Unicode we need to fix: https://bugs.swift.org/browse/SR-660 If the first ASCII character in an identifier is a digit, as in é22, then we'll end up butting the 22 against the run length of the encoded string. We could fix that by just inserting a _, or encoding UTF-8 directly (though Punycode should in theory be more efficient for single-script identifiers).
|
This is great! I dumped a laundry list of other things-to-fix, though we could do them incrementally on top of this. |
|
Just a heads-up that this will break compatibility with the swift remote mirrors library, so we should bump the reflection metadata version number. Alternatively we could keep the old mangling format around until we're ready to do an ABI break there. |
|
@eeckstein Since you are changing the prefix from _T => _S, we are probably going to have to change strip as well. |
eeckstein
force-pushed
the
new-mangling
branch
from
76da89a to
0f5f1c2
Compare
|
I accidentally close this PR |
|
Sorry, somehow I managed to close this PR accidentally. The new PR is #5433 |
[wasm] Revert unnecessary dso_local change
This PR is a proposal for a new mangling scheme. It's still draft and probably the grammar contains several bugs. But it should give an idea how it could look like.
These are the main changes:
This is the biggest change.
It will help to get more common prefixes in the mangled names to optimize the trie in the mach-o object files.
The length of the mangled names will mostly stay the same but the order of 'operands' inside the mangling is more or less reversed.
This change also required to use different 'operator' characters in some cases.
Similar to the S-substitutions, but finer grained. See section 'Identifiers'.
Reduces the size of mangled names in general.
A more efficient way to mangle multiple S-substitutions.
Reduces the size of mangled names with lots of substitutions, e.g. specialized functions.
(on John's request)
Because it's basically a completely new mangling scheme.