Don't codegen items that are statically unreachable

[jyn514]

Currently, all items in a crate are codegened, and only stripped later in a linker pass:

rust/compiler/rustc_monomorphize/src/partitioning/mod.rs

Lines 201 to 205 in 6c4bda6

	// Next we try to make as many symbols "internal" as possible, so LLVM has
	// more freedom to optimize.
	if !tcx.sess.link_dead_code() {
	let _prof_timer = tcx.prof.generic_activity("cgu_partitioning_internalize_symbols");
	partitioner.internalize_symbols(cx, &mut post_inlining);

rust/compiler/rustc_codegen_ssa/src/back/link.rs

Lines 2179 to 2190 in a24a020

	// Try to strip as much out of the generated object by removing unused
	// sections if possible. See more comments in linker.rs
	if !sess.link_dead_code() {
	// If PGO is enabled sometimes gc_sections will remove the profile data section
	// as it appears to be unused. This can then cause the PGO profile file to lose
	// some functions. If we are generating a profile we shouldn't strip those metadata
	// sections to ensure we have all the data for PGO.
	let keep_metadata =
	crate_type == CrateType::Dylib || sess.opts.cg.profile_generate.enabled();
	if crate_type != CrateType::Executable || !sess.opts.unstable_opts.export_executable_symbols
	{
	cmd.gc_sections(keep_metadata);

This unnecessarily bloats the amount of time and memory it takes to compile a crate. We already have a dead_code pass; we should use its results to avoid running codegen on unreachable items.

This is a prerequisite for #103356, but I expect it will be a large boost to compile times even if that never lands.

[scottmcm]

A potential first step would be to handle if const { monomorphized-predicate }, #85836.

[jyn514]

We already do this. #104860 (comment)