interp: limit each value to its own bytes in toLLVMValue

[0pcom]

toLLVMValue gave each struct field and array element the rest of the buffer from its offset, and the zero check at the top scanned all of it. Serializing a large global that an init function has written to was quadratic in its size.

github.com/mattn/go-runewidth has a [2][0x110000]byte width table that its init writes the start of. Programs that import it, for example through charmbracelet/x/ansi, spent about 26 minutes in interp. With this change it takes under a second and the output is unchanged.

The new test uses a global of the same shape. It passes in 0.2s and did not finish in 3 minutes without the change. This is similar to #5368, which fixed a rescan of the same kind in memoryView.store.

[dgryski]

Seems reasonable but I wouldn't mind a second look from someone who understands interp more.

[jakebailey]

The regression test covers the array path, but not the changed struct field slicing. This could go in interp/testdata/store.ll, to cover padded fields and a nested struct:

@struct.value = global { i8, i64, { i32, i8 } } zeroinitializer

define internal void @struct.init(ptr %context) unnamed_addr {
entry:
  %b = getelementptr { i8, i64, { i32, i8 } }, ptr @struct.value, i32 0, i32 1
  store i64 42, ptr %b
  %e = getelementptr { i8, i64, { i32, i8 } }, ptr @struct.value, i32 0, i32 2, i32 1
  store i8 7, ptr %e
  ret void
}

With call void @struct.init(ptr undef) added to runtime.initAll, the expected initializer in store.out.ll is:

@struct.value = global { i8, i64, { i32, i8 } } { i8 0, i64 42, { i32, i8 } { i32 0, i8 7 } }

[0pcom]

Thanks for the IR. I added the struct test to store.ll and store.out.ll as you gave it, with local_unnamed_addr on the output global to match the other entries in that file. It now covers the padded field and the nested struct that the array test could not reach. I did not build or run it locally because this machine has no LLVM build for the tree, so CI is the check.

[0pcom]

I have now built and run it. On LLVM 23 (on top of #5782) go test ./interp passes, including the new struct case in store.ll and TestInterpLargeArray (0.13s).

The struct case also passes on the code before this change, which is expected. The old path was slow but gave the same output, so the test guards the output of the new slicing for padded and nested fields.