Cranelift meeting notes from 2024-05-22.

cranelift/2024/cranelift-05-22.md

@@ -18,5 +18,74 @@
 
 ### Attendees
 
+- elliottt
+- fitzgen
+- abrown
+- jameysharp
+- alexcrichton
+- cfallin
+
 ### Notes
 
+- fitzgen: moving stack maps, safepoints, and knowledge of GC moving out of
+  core cranelift into frontend
+  - work started with elliottt
+  - current design needs reworking to support non-word-sized GC refs (e.g.
+    Wasmtime's 32-bit indices/offsets even on 64-bit hosts) -- one bit in
+    stackmap is one 64-bit word
+  - want to be able to do certain operations on GC refs, such as adding to
+    base, etc. Result is not a GC ref, but is derived from one. bitcasts are
+    footguns because they can be dedup'd, and liveranges of resulting IR are
+    not the same.
+  - goal: move all GC knowledge into frontend; insert safepoints during IR
+    finalization in frontend as explicit dataflow
+    - will still compute and use liveness; otherwise we'll keep all GC refs
+      alive all the time, major pessimization
+    - moving GC is interesting: new refs are outputs of every safepoint
+    - possible alternative: stackslot loads/stores non-side-effecting?
+      - cfallin: could track perfect aliasing of stackslots if we give them
+        names, then carry liveness through them
+        - fitzgen: fair, but new complexity
+        - jameysharp: if we extend alias analysis to have dynamically-indexed
+          alias regions, we could reuse that, not much more complexity?
+        - cfallin: still need dead-store elimination in midend, which would be new
+        - clarification: explicit loads and stores inserted, not part of
+          safepoint instruction
+        - also we'd need to remove unused stackslots
+        - cfallin: maybe better to do liveness in frontend; otherwise need more
+          generic algorithms (dead-store, unused stackslot) and more work in
+          midend over potentially many extraneous GC vals
+    - jameysharp: need a "volatile" memflag? non-trapping, but still opts we
+      can't do
+      - fitzgen: distinction may not be much in practice because there will
+        always be calls after stores and so every store is needed to be visible
+
+- fitzgen: interpreter for Wasmtime involving new backend to Cranelift
+  - using Cranelift to generate optimized bytecode that interpreter VM can run
+  - have a simple prototype working with some arithmetic
+  - will write an RFC
+  - abrown: replace CLIF interpreter?
+    - fitzgen: different purposes; CLIF interp is good oracle for fuzzing, this
+      is for performance
+
+- status
+  - jameysharp:
+    - inlining in ISLE rules, issue created with thoughts
+      - relevant for verification folks, but also worth thinking about for
+        codegen improvement in islec (more matches in same scope)
+    - multiplication, `__multi3` optimization (LLVM compiler-rt helper for wide multiplies)
+      - draft PR with rules that recognize lower half of its body and rewrites
+        into a single wider imul; still TBD matching upper half, interesting
+        discussion in PR
+    - some students looking at instruction scheduling in egraph, built a
+      prototype, implemented algorithm described a year ago in issue; will work
+      with them
+  - alexcrichton: no updates
+  - abrown: no updates
+  - elliottt: removed virtual SP adj and nominal SP from ABI code; SP always
+    stays at end of outgoing args region, only modified in prologues and
+    epilogues (nice!) and in tailcall sequence. Consistent view of frame,
+    matches s390x now.
+  - fitzgen: thinking about GC refs and safepoints; backend for interpreter
+    bytecode
+  - cfallin: no updates