fix(native-backend): decline SEH funclets, refresh stale corpora, drop setjmp-era attrs (#7302)

changelog.d/7310-native-seh-decline-and-corpus-refresh.md

@@ -0,0 +1,35 @@
+### In-process LLVM backend: decline SEH funclets, refresh the stale corpora (#7302 / #7301)
+
+Three follow-ups to the EH migration, all found by auditing what the native
+backend actually covers.
+
+**1. `--target windows` + try/catch was a hard compile error.** Perry lowers
+exceptions to SEH funclets on windows-msvc (`catchswitch`/`catchpad`/
+`catchret`), which the in-process reader cannot construct — inkwell 0.9
+exposes no `build_catch_switch`/`build_catch_pad`/`build_catch_ret` (only an
+opcode enum for *reading* them), so real support needs raw `llvm-sys` FFI.
+Reproduced: `PERRY_LLVM_INPROCESS=native … --target windows` on any
+try/catch (or async) program failed with `unknown instruction catchswitch`.
+Such modules now decline to the textual path — costing only the in-process
+speedup — instead of failing the build. The decline is narrow (windows
+triple AND a personality present), so the macOS EH path added in #7307 is
+untouched, and it is *not* the blanket personality bail #7307 removed.
+
+**2. The tracked corpora were stale.** `spike_text.ll` and `batch_kernel.ll`
+were captured before #7305 and still contained `setjmp` calls and the
+`#0`/`#1` attribute groups — so the reader's primary gate was validating IR
+Perry no longer emits, while missing the forms it does. Regenerated from
+current codegen: the spike corpus now carries 38 `invoke` edges and zero
+setjmp, and both round-trip through native construction and the LLVM
+verifier.
+
+**3. The setjmp-era attribute handling is deleted.** With the corpora
+refreshed, nothing emits or contains `#0` (returns_twice) or `#1` (noinline),
+so the reader's branches for them are gone — the losing mode stops compiling
+rather than lingering as an untested branch (CLAUDE.md kill policy). A corpus
+that still carries one now fails loudly as stale input, which is the point.
+
+Verified locally against LLVM 22.1.4: liveness, behavior parity and
+**byte-identical object verdicts** on the spike (21,393 bytes), the 3-unit
+batch kernel (56,296 bytes) and the try/catch program (46,745 bytes);
+569 `perry-codegen` tests green.

crates/perry-codegen/src/codegen/mod.rs

@@ -2654,6 +2654,12 @@ fn try_native_units(
     target: Option<&str>,
     module_prefix: &str,
 ) -> Option<Result<Vec<u8>>> {
+    // SEH funclets are the one EH shape the in-process reader cannot
+    // construct (see LlModule::needs_eh_funclets). Decline to the textual
+    // path rather than failing the compile.
+    if llmod.needs_eh_funclets() {
+        return None;
+    }
     match crate::native_emit::native_mode() {
         crate::native_emit::NativeMode::Off => None,
         crate::native_emit::NativeMode::Native => Some(
@@ -2686,6 +2692,12 @@ fn try_native_construction(
     target: Option<&str>,
     module_prefix: &str,
 ) -> Option<Result<Vec<u8>>> {
+    // SEH funclets are the one EH shape the in-process reader cannot
+    // construct (see LlModule::needs_eh_funclets). Decline to the textual
+    // path rather than failing the compile.
+    if llmod.needs_eh_funclets() {
+        return None;
+    }
     match crate::native_emit::native_mode() {
         crate::native_emit::NativeMode::Off => None,
         crate::native_emit::NativeMode::Native => Some(crate::native_emit::compile_module_native(

crates/perry-codegen/src/dialect/mod.rs

@@ -292,11 +292,12 @@ impl<'ctx, 'm> FnReader<'ctx, 'm> {
         let attr_str = h.attr_str;
         for a in attr_str.split_whitespace() {
             match a {
-                // Attribute-group references: the only group perry stamps on
-                // defines is #1 (noinline, the setjmp boundary). Group
-                // contents live in the skeleton for *declares*; for defines
-                // we apply the concrete attribute.
-                "#1" => add_enum_attr(ctx, func, "noinline"),
+                // #7302 deleted the setjmp-era groups (#0 returns_twice, #1
+                // noinline); perry no longer stamps an attribute-group ref
+                // on any define. The losing mode stops compiling rather
+                // than lingering as an untested branch (CLAUDE.md kill
+                // policy) — a corpus still carrying one is stale and must
+                // say so loudly.
                 "alwaysinline" | "inlinehint" | "noinline" => add_enum_attr(ctx, func, a),
                 other => bail!("unknown define attribute `{other}`"),
             }
@@ -724,9 +725,9 @@ impl<'ctx, 'm> FnReader<'ctx, 'm> {
         let callee = &after[..paren];
         let close = rmatch_paren(after, paren)?;
         let args_str = &after[paren + 1..close];
-        // Trailing callsite attribute-group ref. Only `#0` (returns_twice,
-        // on setjmp calls) exists in the dialect; the declare carries it too,
-        // so this is fidelity, not correctness.
+        // Trailing callsite attribute-group ref. #7302 removed the only one
+        // perry ever emitted (`#0`, returns_twice on setjmp calls), so any
+        // ref here now means stale input.
         let trailing_attr = after[close + 1..].trim();
 
         let mut args: Vec<BasicMetadataValueEnum> = Vec::new();
@@ -771,15 +772,6 @@ impl<'ctx, 'm> FnReader<'ctx, 'm> {
             .map_err(be)?;
         match trailing_attr {
             "" => {}
-            "#0" => {
-                let kind = inkwell::attributes::Attribute::get_named_enum_kind_id("returns_twice");
-                if kind != 0 {
-                    site.add_attribute(
-                        inkwell::attributes::AttributeLoc::Function,
-                        self.ctx.create_enum_attribute(kind, 0),
-                    );
-                }
-            }
             other => bail!("unknown callsite attribute `{other}`"),
         }
         match site.try_as_basic_value() {

crates/perry-codegen/src/module.rs

@@ -262,6 +262,22 @@ impl LlModule {
         ));
     }
 
+    /// SEH funclets (#7302): true when this module targets windows-msvc AND
+    /// contains try/catch, i.e. when its EH lowering is
+    /// `catchswitch`/`catchpad`/`catchret` rather than Itanium landing pads.
+    ///
+    /// The in-process LLVM reader can build `invoke`/`landingpad` but NOT
+    /// the funclet forms: inkwell 0.9 exposes no `build_catch_switch` /
+    /// `build_catch_pad` / `build_catch_ret` (only an opcode enum for
+    /// reading them), so constructing them needs raw `llvm-sys` FFI. Until
+    /// that lands, such modules take the textual path — declining costs
+    /// nothing but the in-process speedup, whereas letting the reader hit
+    /// the instruction is a hard compile error.
+    pub fn needs_eh_funclets(&self) -> bool {
+        self.target_triple.contains("-windows-")
+            && self.functions.iter().any(|f| f.personality.is_some())
+    }
+
     /// Invoke-EH (#7302): declare the personality routine referenced by
     /// every `define ... personality ptr @perry_eh_personality`. Declared
     /// varargs — the symbol is only ever *named* on define lines and in the

experiments/llvm-inprocess-spike/batch_kernel.ll

@@ -1275,8 +1275,8 @@ declare void @js_wait_for_event()
 declare i32 @js_event_loop_host_driven()
 declare void @js_unsettled_top_level_await_exit()
 declare void @js_throw(double)
-declare ptr @js_try_push()
-declare i32 @_setjmp(ptr) #0
+declare void @js_eh_try_push()
+declare i32 @perry_eh_personality(...)
 declare void @js_try_end()
 declare double @js_get_exception()
 declare void @js_clear_exception()
@@ -10646,9 +10646,6 @@ entry.0:
 }
 
 
-attributes #0 = { returns_twice }
-attributes #1 = { noinline }
-
 attributes #2 = { nounwind willreturn readnone }
 
 attributes #3 = { nounwind willreturn readonly }