Segmentation fault in editor analysis references after for...of typedtree change #7887
The editor analysis binary can segfault when running references against an existing project (using v12) after commit f97609381a56063edf37c6b3bdd352f77fe44189. In 1ec0a25be752e97e22fce0d827b08c5e8c69c8fc (Move rescript.json reading out of bsc into rewatch - #8365) references works.
That commit adds new typedtree constructors for for...of / for await...of:
Texp_for_ofTexp_for_await_of
but the CMT magic number was not changed. Because .cmt files are marshalled OCaml values, changing the typedtree shape without changing the magic number can make the analysis binary read incompatible/stale .cmt files as if they were valid. This can produce invalid runtime values and segfault before OCaml exception handling can catch it.
Reproduction
From rescript-compiler:
git checkout f97609381a56063edf37c6b3bdd352f77fe44189
dune build @install
rescript-lang.org repo master branch:
/home/pedro/Desktop/projects/rescript-compiler/_build/default/analysis/bin/main.exe references /home/pedro/Desktop/projects/rescript-lang.org/apps/docs/app/DocsRoot.res 3 9
[1] 2261129 segmentation fault (core dumped) references 3 9
It's a break because the compiler repo is on v13 and rescript-lang.org uses v12.2.0, but it's an important point for migration. It is necessary to remove all lib folders with cmt files.
Codex suggestions
Suspected cause
f97609381 changes the marshalled typedtree representation by adding new constructors, but Config.cmt_magic_number remains unchanged.
Since editor analysis reads .cmt files through Cmt_format.read_cmt, stale or ABI-incompatible .cmt files can pass the magic-number check and then crash during unmarshalling or typedtree traversal.
Possible fix
Bump the CMT magic number when changing typedtree constructors. Since this commit also changes parsetree constructors, AST magic may need to be reviewed as well.
Segmentation fault in editor analysis
referencesafterfor...oftypedtree change #7887The editor analysis binary can segfault when running
referencesagainst an existing project (using v12) after commitf97609381a56063edf37c6b3bdd352f77fe44189. In1ec0a25be752e97e22fce0d827b08c5e8c69c8fc(Move rescript.json reading out of bsc into rewatch - #8365) references works.That commit adds new typedtree constructors for
for...of/for await...of:Texp_for_ofTexp_for_await_ofbut the CMT magic number was not changed. Because
.cmtfiles are marshalled OCaml values, changing the typedtree shape without changing the magic number can make the analysis binary read incompatible/stale.cmtfiles as if they were valid. This can produce invalid runtime values and segfault before OCaml exception handling can catch it.Reproduction
From
rescript-compiler:rescript-lang.org repo master branch:
It's a break because the compiler repo is on v13 and rescript-lang.org uses v12.2.0, but it's an important point for migration. It is necessary to remove all
libfolders with cmt files.Codex suggestions
Suspected cause
f97609381changes the marshalled typedtree representation by adding new constructors, butConfig.cmt_magic_numberremains unchanged.Since editor analysis reads
.cmtfiles throughCmt_format.read_cmt, stale or ABI-incompatible.cmtfiles can pass the magic-number check and then crash during unmarshalling or typedtree traversal.Possible fix
Bump the CMT magic number when changing typedtree constructors. Since this commit also changes parsetree constructors, AST magic may need to be reviewed as well.