| FazBrowse GitHub Viewer | Trending | | Home |
| Tools: [Download Repo ZIP] [Original HTTPS Page] |
Sorry, something went wrong.
|
According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅ |
Sorry, something went wrong.
#22133) Fixes mypyc/mypyc#1231 A reference to a local variable compiles to the variable's register, and the register is only read when the operation that uses the value runs. If a later operand reassigns the variable with an assignment expression, the earlier operand sees the new value, so `pair(i, (i := 1))` passed `(1, 1)` instead of `(0, 1)`. #21249 fixed this for tuple displays by copying every register item, but calls, method receivers, operators, list/set/dict displays, subscripts and comprehension iterables were still affected. The last one could crash: `[x for x in t if (t := (9,))]` read past the end of the new tuple. mypyc already collects the locals that assignment expressions reassign within the current top-level expression (`reassigned_in_expr`, added in #21702 to restrict borrowing). Now, when one of those locals is read, its value is copied to a temporary register. Expressions without such an assignment generate the same IR as before. A callee reassigned by its own arguments, as in `f(1, (f := g))`, was also evaluated after the arguments, so `g` was called. Such a callee is now evaluated first. Reading a possibly unbound variable before the assignment expression now raises `UnboundLocalError`, as in Python, instead of using the new value. This doesn't change `x += (x := 1)` or `a[(i := 1)] = i`, where the read happens in statement code outside the expression.
| Back | FazBrowse Home | New Git URL |
Final instance attributes of native classes can't be rebound after initialization, so we can borrow them more aggressively than regular attributes, as long as the base object is kept alive for the duration of the borrowing. This allows skipping some incref/decref operations.
Many workloads spend a significant fraction of CPU on incref/decref, and these operations are quite a bit more expensive on free-threaded builds, so the potential performance impact is significant especially on free-threaded builds.
For example, the attribute x can only be safely borrowed if it's Final, since otherwise foo could assign to the attribute and free the old value:
There are some subtleties in the implementation. Here are the main things that required extra care:
I used coding agent assist, especially for tests, but created the implementation is short, individually reviewed increments.