fork/Graphite
1
0
Fork 0
mirror of https://github.com/GraphiteEditor/Graphite.git synced 2026-09-30 05:48:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

Align RFC summary, soundness, and rationale with the stack model

Browse Source
This commit is contained in:
Dennis Kobert
2026-08-05 23:02:38 +02:00
parent 18225d0c12
commit fa762379e3
1 changed files with 8 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
node-graph/rfcs/attribute-model.md
+8 -8
View File
@@ -7,9 +7,9 @@ declare their attribute reads and writes in their signatures, and the
compiler resolves every access to a byte offset during wiring, so there is compiler resolves every access to a byte offset during wiring, so there is
no name lookup at runtime. Storage and batch results are per-attribute no name lookup at runtime. Storage and batch results are per-attribute
columns. The contiguous record only exists as a per-lane view, assembled columns. The contiguous record only exists as a per-lane view, assembled
into buffers the compiler assigns. All of the machinery that could into activation frames on a per-thread record stack. All of the machinery
corrupt a layout is generated code, so getting it wrong is a type error or that could corrupt a layout is generated code, so getting it wrong is a
a graph compile error rather than undefined behavior. type error or a graph compile error rather than undefined behavior.
# Motivation # Motivation
@@ -484,9 +484,9 @@ make semantic mistakes (evaluating an input it did not need to) but
cannot misalign an offset. Kernels see contexts only as an opaque cannot misalign an offset. Kernels see contexts only as an opaque
`impl Ctx + ...` they cannot construct, and lifetimes keep them from `impl Ctx + ...` they cannot construct, and lifetimes keep them from
stashing handles in node state. The one remaining discipline lives stashing handles in node state. The one remaining discipline lives
inside generated code (a result buffer must not be borrowed across a inside generated code (a borrow of a released frame must not survive the
sibling evaluation it could alias), and debug assertions guard it at next claim), and debug assertions guard it at wiring boundaries,
wiring boundaries, following the existing precedent for TypeId checks. following the existing precedent for TypeId checks.
# Drawbacks # Drawbacks
@@ -515,8 +515,8 @@ wiring boundaries, following the existing precedent for TypeId checks.
per value, and no compile-time name checking. Interning the keys per value, and no compile-time name checking. Interning the keys
improves the constant (about 1.9ns per access vs. 0.43 for a resolved improves the constant (about 1.9ns per access vs. 0.43 for a resolved
offset) but keeps a per-access search and rules out the structural offset) but keeps a per-access search and rules out the structural
optimizations that need static layouts: bypass, uniform columns, slot optimizations that need static layouts: bypass, uniform columns,
coalescing. stack-allocated activation frames.
- Attributes as separate graph edges, one channel per attribute: bypass - Attributes as separate graph edges, one channel per attribute: bypass
and per-channel caching become graph structure. We prototyped and and per-channel caching become graph structure. We prototyped and
measured this. Without caching at fan-outs, every channel re-evaluates measured this. Without caching at fan-outs, every channel re-evaluates