Path: utzoo!attcan!ncrcan!scocan!ron
From: ron@sco.COM (Ron Irvine)
Newsgroups: comp.arch
Subject: Re: Inlining subroutines at link time
Message-ID: <1990Jul6.115849.4556@sco.COM>
Date: 6 Jul 90 15:58:49 GMT
Reply-To: ron@scocan.UUCP (Ron Irvine)
Organization: SCO Canada, Inc. (formerly HCR Corporation)
Lines: 71

The CDC post-loader (called the afterburner) was developed to
optimize the executable.  It can inline functions, reallocate
registers across function calls and turn far calls into near
calls. It can be a big win on the CYBER since subroutine
calls are expensive.

In general a post-loader can be a powerful tool in "optimizing"
the executable, it can:

Inline/Outline:
	- inline functions, faster execution larger code size.
	- outline code (make into a function), this can reduce
		code size and be profitable on some machines
		(smaller code may fit into cache).
		
Do global register allocation:
	- this can be of great benifit for processors with
		global registers.

Assign registers across function calls:
	- a post loader has full program scope (including libraries)
		so it can create register usage information and
		eliminate some register save/restores across function
		calls.

Use short addresses:
	- data can be grouped to allow short memory references.
		This could be 16 bit offsets instead of 32 bit.
	- or fast access through a data register + small
		offset.
	- far calls can become near calls

Change instructions:
	- Instruction sequences can be tuned for a specific
		machine (without a recompile).
	- Patch around hardware/software bugs without changing
		the compiler.
	- Emulate new hardware/software.
	- Yet another place to reposition load/store and delay
		branch code.

Relocate data and functions:
	- This can be done to improve cache hits based on
		static or dynamic infomation (with feeback
		from a program run).
	- Data can be grouped so that often used data resides
		in a common page in memory.
	- functions can be gouped so that paging can be minimized
		and startup time reduced.
	- get rid of function and data that is newer used.

Performance monitor:
	- Add code to record the performance of a program.  This
		can be done to the "final" executable without
		the need to recompile and link with special
		libraries.

Misc:
	- add code to memory references to find a specific bug
	- modify the scope of symbols in a object file (for
		example this can allow more than one yacc/lex
		parser in a program)
	- allow a function (or data) to be replaced with a new one.
	- rearrange a production binary so that the function
		positions (addresses) are unique for each copy
		(and thus can encode a serial number for identification).

See, "Postloading for Fun and Profit", S.C.Johnson, UNENIX, Winter 90.


Ron Irvine, uunet!scocan!ron