Parsing performance

[soutaro]

This is related to the RBS file parsing performance degradation.

The RBS file parsing in rbs-4.0 (dev) is ~5x slower than rbs-3.9, based on the core/**/*.rbs files, especially for smaller files.

https://docs.google.com/spreadsheets/d/1dQBGIC1_zWco6c5OHH5VunP8ZuZLzwaMDB-U31OgUAc/edit?gid=0#gid=0

I implemented minor changes to improve the parsing performance, reusing empty array and hash objects. The changes improved the parsing performance slightly, but it still clearly slower than rbs-3.9.

[soutaro]

• Try another allocator API, not memmap with linked list of heap (try reverting Use one massive slice of virtual memory per arena Shopify/rbs#38)
• Try introduce batch loading API -- to reuse the allocators, reduce C-Ruby swtiching
• Reuse hash object for keyword arguments

[amomchilov]

Another thought:

How much time do we spend computing assertions? Should we disable them in "release" mode?

Currently, rbs_assert() is just a function that's never disabled:

rbs/src/util/rbs_assert.c

Lines 8 to 19 in 4baf465

	void rbs_assert(bool condition, const char *fmt, ...) {
	if (condition) {
	return;
	}
	va_list args;
	va_start(args, fmt);
	vfprintf(stderr, fmt, args);
	va_end(args);
	fprintf(stderr, "\n");
	exit(EXIT_FAILURE);
	}

Perhaps we should rename the function to rbs_assert_impl(), and create an #define rbs_assert() macro that calls it only if we're in debug mode.

For comparison, C's standard library's assert.h, makes assert() a no-op if you #define NDEBUG

[amomchilov]

Another thought: our extconf.rb doesn't specify an optimization level. We disable optimizations (-O0) if DEBUG is defined:

rbs/ext/rbs_extension/extconf.rb

Line 21 in 4baf465

append_cflags ['-O0', '-g'] if ENV['DEBUG']

But we never specify -O3 if we're in a "release" mode