build: be more aggressive with linker optimizations; wrap mold

Signed-off-by: NotAShelf <raf@notashelf.dev>
Change-Id: I0e3132ab1499684eda715c3cee9b27a16a6a6964

.cargo/config.toml

@@ -1,5 +1,31 @@
-# Link with Mold, and without libc! We use nostartfiles to avoid the C runtime
+# Use a linker wrapper that invokes mold then strips junk sections with objcopy.
+# mold cannot discard .eh_frame/.dynstr/.comment via linker scripts, so we do
+# it as a post-link step.
 # See:
 #  <https://github.com/rui314/mold?tab=readme-ov-file#how-to-use>
 [target.'cfg(target_os = "linux")']
-rustflags = [ "-C", "link-arg=-fuse-ld=mold", "-C", "link-arg=-nostartfiles" ]
+linker = "scripts/ld-wrapper"
+rustflags = [
+  # No C runtime, we provide _start ourselves
+  "-C",
+  "link-arg=-nostartfiles",
+  # Fully static,  no dynamic linker, no .interp/.dynsym/.dynamic overhead
+  "-C",
+  "link-arg=-static",
+  # Static PIE is incompatible with -static :(
+  "-C",
+  "relocation-model=static",
+  # Suppress .eh_frame emission from our own codegen (does not cover compiler_builtins;
+  # those remnants are removed by the linker wrapper via objcopy post-link)
+  "-C",
+  "force-unwind-tables=no",
+  # Linker flags
+  "-C",
+  "link-arg=-Wl,--gc-sections",   # remove unreferenced input sections
+  "-C",
+  "link-arg=-Wl,--strip-all",     # strip all symbol table entries
+  "-C",
+  "link-arg=-Wl,--build-id=none", # omit the .note.gnu.build-id section
+  "-C",
+  "link-arg=-Wl,-z,norelro",      # disable RELRO (removes relro_padding)
+]

Cargo.toml

@@ -11,15 +11,15 @@ version      = "0.4.13"
 
 [workspace.dependencies]
 microfetch-alloc = { path = "./crates/alloc" }
-microfetch-asm = { path = "./crates/asm" }
+microfetch-asm   = { path = "./crates/asm" }
-microfetch-lib = { path = "./crates/lib" }
+microfetch-lib   = { path = "./crates/lib" }
 
 criterion                 = { default-features = false, features = [ "cargo_bench_support" ], version = "0.8.2" }
 criterion-cycles-per-byte = "0.8.0"
 
 [profile.dev]
 opt-level = 1
-panic = "abort"
+panic     = "abort"
 
 [profile.release]
 codegen-units = 1

microfetch/Cargo.toml

@@ -11,14 +11,14 @@ repository             = "https://github.com/notashelf/microfetch"
 publish                = false
 
 [dependencies]
-hotpath                   = { optional = true, version = "0.14.0" }
+hotpath                    = { optional = true, version = "0.14.0" }
 microfetch-alloc.workspace = true
-microfetch-lib.workspace = true
+microfetch-asm.workspace   = true
-microfetch-asm.workspace = true
+microfetch-lib.workspace   = true
 
 [features]
 hotpath       = [ "dep:hotpath" ]
 hotpath-alloc = [ "hotpath/hotpath-alloc" ]
 
 [lints]
-workspace = true
+workspace = true

scripts/ld-wrapper

@@ -0,0 +1,34 @@
+#!/usr/bin/env sh
+# Invoke mold, then strip junk sections from the output binary with objcopy.
+# This (more or less) removes sections that mold cannot discard itself, suck as:
+#  - .eh_frame / .eh_frame_hdr  - unwind tables from compiler_builtins
+#  - dynstr                     - mold emits this, even for fully static binaries
+#  - .comment                   - compiler version string
+#
+#  We forward everything to mold via -fuse-ld, then post-process the output in place.
+
+set -eu
+
+# Locate the output file
+OUTPUT=""
+prev=""
+for arg in "$@"; do
+	if [ "$prev" = "-o" ]; then
+		OUTPUT="$arg"
+		break
+	fi
+	prev="$arg"
+done
+
+# Invoke mold via the cc driver, forward all original arguments
+cc -fuse-ld=mold "$@"
+
+# Remove sections that mold cannot discard
+if [ -n "$OUTPUT" ] && [ -f "$OUTPUT" ]; then
+	objcopy \
+		--remove-section=.eh_frame \
+		--remove-section=.eh_frame_hdr \
+		--remove-section=.dynstr \
+		--remove-section=.comment \
+		"$OUTPUT"
+fi