nandi/jolt-native · aabacba

⑂ aabacba

Ship the object without the notes nobody reads

Debug info was 5.9M of the 14.5M the phone carries, and no one is on the
far end of it: the .so goes into the APK, and a crash comes back through
logcat as a bionic backtrace over the loaded segments, never a DWARF line
table. cargo's release profile already carried none — its object strips to
within 168 bytes of itself — so buck2 was the odd one out rather than the
generous one.

Fat LTO on the cdylib is the other half. It takes .text down a tenth and
halves the unwind tables, which are mostly duplicate landing pads once the
graph is one unit. opt-level stays at 2: 3 was measured beside it and came
out 134K larger.

Android libvidya, 14,496,256 to 8,804,920. The desktop pair comes along
through the shared profile — libvidya 12.6M to 10.8M, libjoltmoq 22.2M to
19.0M.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

nandi committed 2026-08-31T19:13:57-07:00 Browse files

aabacba parent: c71fc59

modified Cargo.toml +5 -0 Viewed

@@ -41,6 +41,11 @@ repository = "https://gitlab.com/nandithebull/jolt-native"
41	41	# worth looking at. A member cannot set this, so it is set here.
42	42	[profile.release]
43	43	opt-level = 2
	44	+# The same pair the buck2 cdylib sets in crates/jolt-vidya/BUCK, so the two
	45	+# builds of libvidya agree on more than opt-level. 1.5M off the desktop object;
	46	+# opt-level 3 was measured alongside and is larger, so it stays at 2.
	47	+lto = "fat"
	48	+codegen-units = 1
44	49
45	50	[workspace.dependencies]
46	51	jolt-abi = { path = "crates/jolt-abi" }

  # worth looking at. A member cannot set this, so it is set here.
  [profile.release]
  opt-level = 2
++# The same pair the buck2 cdylib sets in crates/jolt-vidya/BUCK, so the two
++# builds of libvidya agree on more than opt-level. 1.5M off the desktop object;
++# opt-level 3 was measured alongside and is larger, so it stays at 2.
++lto = "fat"
++codegen-units = 1
  [workspace.dependencies]
  jolt-abi = { path = "crates/jolt-abi" }

modified crates/jolt-vidya/BUCK +12 -0 Viewed

@@ -10,6 +10,18 @@ rust_library(
10	10	# and that name, not the filename, is what the linker writes into a
11	11	# dependent's DT_NEEDED. Every consumer looks for `libvidya`.
12	12	soname = "libvidya.so",
	13	+ # Cross-crate inlining across the whole graph, and one codegen unit so the
	14	+ # dedup is not fighting parallelism inside the crate. Measured on the
	15	+ # desktop object: 12.06M to 10.57M, of which 628K is .text and the rest is
	16	+ # the unwind tables halving — LTO alone folds most of the duplicate landing
	17	+ # pads away. Only the cdylib asks for it; the rlibs above stay as they are.
	18	+ #
	19	+ # opt-level stays at 2. 3 was measured too and came out 134K larger, so
	20	+ # the level in the toolchain is the one to keep.
	21	+ rustc_flags = [
	22	+ "-Clto=fat",
	23	+ "-Ccodegen-units=1",
	24	+ ],
13	25	visibility = ["PUBLIC"],
14	26	# Cargo renames the dependency so the cdylib itself can keep the `vidya`
15	27	# crate name; mirror that here.

      # and that name, not the filename, is what the linker writes into a
      # dependent's DT_NEEDED. Every consumer looks for `libvidya`.
      soname = "libvidya.so",
++    # Cross-crate inlining across the whole graph, and one codegen unit so the
++    # dedup is not fighting parallelism inside the crate. Measured on the
++    # desktop object: 12.06M to 10.57M, of which 628K is .text and the rest is
++    # the unwind tables halving — LTO alone folds most of the duplicate landing
++    # pads away. Only the cdylib asks for it; the rlibs above stay as they are.
++    #
++    # opt-level stays at 2. 3 was measured too and came out 134K *larger*, so
++    # the level in the toolchain is the one to keep.
++    rustc_flags = [
++        "-Clto=fat",
++        "-Ccodegen-units=1",
++    ],
      visibility = ["PUBLIC"],
      # Cargo renames the dependency so the cdylib itself can keep the `vidya`
      # crate name; mirror that here.

modified toolchains/BUCK +7 -0 Viewed

@@ -127,6 +127,13 @@ hermetic_rust_toolchain(
127	127	# without this the two objects are debug builds wearing release names.
128	128	# OPT_LEVEL=2 in the buildscript fixups is the same setting for the C.
129	129	"-Copt-level=2",
	130	+ # Debug info is 5.9M of the 14.5M Android object, and nothing reads it:
	131	+ # the .so ships in the APK, and a crash comes back through logcat as a
	132	+ # bionic backtrace against the loaded segments, not a DWARF line table.
	133	+ # cargo's release profile already carries none — its object strips to
	134	+ # within 168 bytes of itself — so this is buck2 matching that, not a
	135	+ # new decision about what is debuggable. Drop the flag to get it back.
	136	+ "-Cstrip=debuginfo",
130	137	"--cfg=jolt_cc=\"" + _ZIG_VERSION + "\"",
131	138	],
132	139	# Only triples there is an archive for can be named here: the desktop one

          # without this the two objects are debug builds wearing release names.
          # OPT_LEVEL=2 in the buildscript fixups is the same setting for the C.
          "-Copt-level=2",
++        # Debug info is 5.9M of the 14.5M Android object, and nothing reads it:
++        # the .so ships in the APK, and a crash comes back through logcat as a
++        # bionic backtrace against the loaded segments, not a DWARF line table.
++        # cargo's release profile already carries none — its object strips to
++        # within 168 bytes of itself — so this is buck2 matching that, not a
++        # new decision about what is debuggable. Drop the flag to get it back.
++        "-Cstrip=debuginfo",
          "--cfg=jolt_cc=\"" + _ZIG_VERSION + "\"",
      ],
      # Only triples there is an archive for can be named here: the desktop one