nativecomp: optimize local CU calls at speed 2scratch/better-comp

* etc/NEWS: Document speed-2 local call optimization.
* etc/TODO: Remove completed item.
* lisp/emacs-lisp/comp.el
(comp--cu-local-func-c-name-v): New function.
(comp--call-optim-form-call): Enable named local direct calls
at speed 2.
(comp--function-trampoline-form): New helper.
(comp-trampoline-compile, comp-local-function-trampoline-compile):
Share trampoline generation.
* lisp/emacs-lisp/comp-run.el
(comp-local-function-trampoline--install-now): New function.
(comp-local-function-trampoline-install): Install local trampolines.
Defer trampoline compilation until after load.
* src/comp.c (emit_call, emit_ctxt_code, compile_function)
(load_comp_unit, unload_comp_unit): Add local function relocation
support.
(native_comp_local_function_p): New function.
(comp--install-local-function-trampoline): New subr.
(syms_of_comp): Register it and update trampoline docs.
* src/comp.h (Lisp_Native_Comp_Unit): Add local relocation slot.
(native_comp_local_function_p): Declare.
* src/data.c (Ffset): Install local trampolines for redefined
named local native functions.  Keep skipping anonymous lambdas.
* src/pdumper.c: Clear local relocation state.
* test/src/comp-tests.el
(comp-tests--run-in-sub-emacs): New helper.
(comp-tests--direct-call-redefinition-form): New helper.
(comp-tests-direct-call-redefinition-speed-split): New test.
(comp-tests-anonymous-lambda-recompile): New test.
(comp-tests-direct-call-with-lambdas): Use an explicit output file.

2 files changed, 6 insertions, 9 deletions

diff --git a/etc/NEWS b/etc/NEWS
index 6107ee153f5..8f67b3f870f 100644
--- a/etc/NEWS
+++ b/etc/NEWS
@@ -3885,6 +3885,12 @@ At native compilation speed 2 and above, declared function types are
 used for type propagation, value prediction, and optimization within
 function bodies.
 
+** Native compilation speed 2 now optimizes local calls.
+Named calls between native-compiled functions in the same compilation
+unit can now be optimized at speed 2 while still honoring later
+redefinition through a trampoline.  At speed 3, these calls keep the
+previous direct-call behavior.
+
 ** Nested backquotes are not supported any more in Pcase patterns.
 
 ---
diff --git a/etc/TODO b/etc/TODO
index b93236d3759..e577fec0e6e 100644
--- a/etc/TODO
+++ b/etc/TODO
@@ -993,15 +993,6 @@ It would make it easy to add (and remove) mappings like
 
 *** Performance
 
-**** Intra compilation unit call optimization
-
-We could have a mechanism similar to what we use for optimizing calls
-to primitive functions.  IE using a link table for each compilation
-unit (CU) such that calls from functions in a CU targeting functions
-in the same CU don't have to go through funcall.  If one of these
-functions is redefined, a trampoline is compiled and installed to
-restore the redirection through funcall.
-
 **** Better runtime function inlining
 
 Several functions could be open-coded in generated code once exposed to