[mypyc] feat: add ascii format fast path

.github/workflows/build_wheels.yml

@@ -2,6 +2,14 @@ name: Trigger wheel build
 
 on:
   push:
+    paths:
+      - '**/*.py'
+      - '**/*.pyi'
+      - 'pyproject.toml'
+      - 'setup.py'
+      - 'tox.ini'
+      - '.pre-commit-config.yaml'
+      - '.github/workflows/build_wheels.yml'
     branches: [main, master, 'release*']
     tags: ['*']
 

.gitignore

@@ -60,3 +60,4 @@ test_capi
 test_capi
 /mypyc/lib-rt/build/
 /mypyc/lib-rt/*.so
+.worktrees/

AGENTS.md

@@ -0,0 +1,18 @@
+# Agent Requirements
+
+All agents must follow these rules:
+
+1) Fully test your changes before submitting a PR (run the full suite or all relevant tests).
+2) PR titles must be descriptive and follow Conventional Commits-style prefixes:
+   - Common: `feat:`, `fix:`, `chore:`, `refactor:`, `docs:`, `test:`, `perf:`
+   - Support titles: `fix(docs):`, `fix(benchmarks):`, `fix(cicd):`
+3) Commit messages must follow the same Conventional Commits-style prefixes and include a short functional description plus a user-facing value proposition.
+4) PR descriptions must include Summary, Rationale, and Details sections.
+5) Run relevant Python tests for changes (pytest/unittest or the repo's configured runner).
+6) Follow formatting/linting configured in pyproject.toml, setup.cfg, tox.ini, or ruff.toml.
+7) Update dependency lockfiles when adding or removing Python dependencies.
+8) If the repo uses mypyc, verify tests run against compiled extensions (not interpreted Python) and note how you confirmed.
+9) Keep base image tags pinned.
+10) If the branch you're assigned to work on is from a remote (ie origin/master or upstream/awesome-feature) you must ensure you fetch and pull from the remote before you begin your work.
+
+Reference: https://www.conventionalcommits.org/en/v1.0.0/

mypyc/irbuild/format_str_tokenizer.py

@@ -26,7 +26,7 @@
 from mypyc.irbuild.constant_fold import constant_fold_expr
 from mypyc.primitives.bytes_ops import bytes_build_op
 from mypyc.primitives.int_ops import int_to_str_op
-from mypyc.primitives.str_ops import str_build_op, str_op
+from mypyc.primitives.str_ops import ascii_op, str_build_op, str_op
 
 
 @unique
@@ -42,6 +42,7 @@ class FormatOp(Enum):
 
     STR = "s"
     INT = "d"
+    ASCII = "a"
     BYTES = "b"
 
 
@@ -53,14 +54,25 @@ def generate_format_ops(specifiers: list[ConversionSpecifier]) -> list[FormatOp]
     format_ops = []
     for spec in specifiers:
         # TODO: Match specifiers instead of using whole_seq
-        if spec.whole_seq == "%s" or spec.whole_seq == "{:{}}":
+        # Conversion flags for str.format/f-strings (e.g. {!a}); only if no format spec.
+        if spec.conversion and not spec.format_spec:
+            if spec.conversion == "!a":
+                format_op = FormatOp.ASCII
+            else:
+                return None
+        # printf-style tokens and special f-string lowering patterns.
+        elif spec.whole_seq == "%s" or spec.whole_seq == "{:{}}":
             format_op = FormatOp.STR
         elif spec.whole_seq == "%d":
             format_op = FormatOp.INT
+        elif spec.whole_seq == "%a":
+            format_op = FormatOp.ASCII
         elif spec.whole_seq == "%b":
             format_op = FormatOp.BYTES
+        # Any other non-empty spec means we can't optimize; fall back to runtime formatting.
         elif spec.whole_seq:
             return None
+        # Empty spec ("{}") defaults to str().
         else:
             format_op = FormatOp.STR
         format_ops.append(format_op)
@@ -152,6 +164,11 @@ def convert_format_expr_to_str(
                 var_str = builder.primitive_op(int_to_str_op, [builder.accept(x)], line)
             else:
                 var_str = builder.primitive_op(str_op, [builder.accept(x)], line)
+        elif format_op == FormatOp.ASCII:
+            if (folded := constant_fold_expr(builder, x)) is not None:
+                var_str = builder.load_literal_value(ascii(folded))
+            else:
+                var_str = builder.primitive_op(ascii_op, [builder.accept(x)], line)
         elif format_op == FormatOp.INT:
             if isinstance(folded := constant_fold_expr(builder, x), int):
                 var_str = builder.load_literal_value(str(folded))

mypyc/primitives/str_ops.py

@@ -51,6 +51,15 @@
     error_kind=ERR_MAGIC,
 )
 
+# ascii(obj)
+ascii_op = function_op(
+    name="builtins.ascii",
+    arg_types=[object_rprimitive],
+    return_type=str_rprimitive,
+    c_function_name="PyObject_ASCII",
+    error_kind=ERR_MAGIC,
+)
+
 # translate isinstance(obj, str)
 isinstance_str = function_op(
     name="builtins.isinstance",

mypyc/test-data/irbuild-constant-fold.test

@@ -192,9 +192,14 @@ L0:
 from typing import Any, Final
 
 FMT: Final = "{} {}"
+FMT_A: Final = "{!a}"
 
 def f() -> str:
     return FMT.format(400 + 20, "roll" + "up")
+def g() -> str:
+    return FMT_A.format("\u00e9")
+def g2() -> str:
+    return FMT_A.format("\u2603")
 [out]
 def f():
     r0, r1, r2, r3 :: str
@@ -204,6 +209,18 @@ L0:
     r2 = ' '
     r3 = CPyStr_Build(3, r0, r2, r1)
     return r3
+def g():
+    r0, r1 :: str
+L0:
+    r0 = "'\\xe9'"
+    r1 = CPyStr_Build(1, r0)
+    return r1
+def g2():
+    r0, r1 :: str
+L0:
+    r0 = "'\\u2603'"
+    r1 = CPyStr_Build(1, r0)
+    return r1
 
 [case testIntConstantFoldingFinal]
 from typing import Final

mypyc/test-data/irbuild-str.test

@@ -291,14 +291,15 @@ def f(var: Union[str, NewStr], num: int) -> None:
     s2 = "I am %d years old." % num
     s3 = "Hi! I'm %s. I am %d years old." % (var, num)
     s4 = "Float: %f" % num
+    s5 = "Ascii: %a" % var
 [typing fixtures/typing-full.pyi]
 [out]
 def f(var, num):
     var :: str
     num :: int
     r0, r1, r2, s1, r3, r4, r5, r6, s2, r7, r8, r9, r10, r11, s3, r12 :: str
     r13, r14 :: object
-    r15, s4 :: str
+    r15, s4, r16, r17, r18, s5 :: str
 L0:
     r0 = "Hi! I'm "
     r1 = '.'
@@ -320,6 +321,10 @@ L0:
     r14 = PyNumber_Remainder(r12, r13)
     r15 = cast(str, r14)
     s4 = r15
+    r16 = PyObject_ASCII(var)
+    r17 = 'Ascii: '
+    r18 = CPyStr_Build(2, r17, r16)
+    s5 = r18
     return 1
 
 [case testDecode]