Skip to content

Attempt to merge Scion into main - #215

Open
rootjalex wants to merge 763 commits into
mainfrom
ajr/pldi26-merge-pr
Open

rootjalex wants to merge 763 commits into
mainfrom
ajr/pldi26-merge-pr

Conversation

@rootjalex

@rootjalex rootjalex commented Aug 31, 2026 •

Copy link
Copy Markdown
Owner

Attempt to merge in all features/some sample apps, without all the benchmarking cruft and stuff.

A set is iterable but owns its storage and has no static length, so the C++
backend's stack allocation should default-construct it rather than assert on
a size it cannot have.
The C++ runtime provides set<T>; CUDA has no equivalent.
layout-bad-path had no flags line, so it never reached verify-layouts and
recorded a clean compile. build-inner-group was registered nowhere.
The rt harness references apps/rt/layouts, but the merge never brought the
directory over. Splits that left tag values uncovered now use a wildcard arm,
and counts materialized from a bitfield carry the cast the tree type needs.
Scion disabled both passes rather than teach them pointer groups, pointer-encoded
child references, and bitfield split widths. Two collision schedules had real
layout errors hidden behind that.
alignas can only raise alignment, so a layout asking for less than a wide vector
member's natural alignment failed to compile.
Lost when the Scion merge dropped apps/wos; ported to the current layout syntax.
Its golden was left orphaned in the tree.
Scion deleted the source and left the golden behind.
Rewrite::visit(Access) called mutate on its own node, which dispatched straight
back into it. A parent. access reached that path, so any implicit hierarchy
crashed the compiler. Also resolve a variant field carried by a root argument,
which validate_layout already accepts, and let a variant that stores nothing
have an empty count arm.
Scion deleted apps/wos and the cl-cw-bvh8-idx layouts. Ported to the current
layout syntax; the fcpw schedule still needs child-bound sorting.
A distinct configuration from align16, not a duplicate of it.
Removes results collection and the Embree and FCPW comparisons, and strips FCL
from the collision driver. The apps still build and run on Bonsai alone.
The build regenerates these; the checked-in copies predate the vector and
runtime-path changes.
It offers a median split as well as the surface-area heuristic.
Only the x86_64 override was re-blessed, so CI kept the old prologue.
Token::size() is an exhaustive switch with no default, so an IN token fell off
the end of it. UBSan traps that in a Debug build, aborting every test that uses
an interval bound.
std::source_location::column() differs between Apple Clang and GCC, so every
golden recording it was host-specific.
The caret already marks the column, and a column taken from the host compiler
made the diagnostic machine-specific.
Neither takes a value, but both skipped one.
Codegen tests now pass --target instead of inheriting the host's, so one golden
is right on every machine and the per-arch override files are gone.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants