JIT should write bytes more efficiently, allow machine code deallocation, allow growing code buffer

[llvmbot]

Bugzilla Link	469
Resolution	FIXED
Resolved on	Feb 22, 2010 12:46
Version	1.0
OS	All
Reporter	LLVM Bugzilla Contributor

Extended Description

I'm using LLVM as a JIT engine, from a simple C++ driver (interpreter, actually).
It creates lots of code, and I recently clocked it at ~10,000 cycles per byte
of code generated. This is just a place-holder for now (sorry for the noise) but
I'll add concrete figures/profiling data as I get around to doing proper tests.
I also intend to hack up a simple benchmark that I can port across LLVM, Parrot
and libjit, just to see how they compare (in terms of MB/sec of x86 emitted for
some simple test case)

I'd appreciate it if anyone in the know would brain-dump any odd JIT
performance-related TODOs here. I'm aware of other bugzilla enteries, but I'll
kick off yet another entry here with a comment from Chris the other day:

we were just talking today about how the MachineCodeEmitter interface
needs to be changed to not use a virtual method dispatch per byte

Will paypal $1 to whoever fixes this. How's that for incentive! ;)

10,000 clocks is a lot you guys... ;)

[llvmbot]

assigned to @lattner

[lattner]

The recent changes to the register allocator related stuff have improved this situation fairly substantially.
Can you make this bug more specific so we have some hope of ever closing it?

-Chris

[lattner]

This bug has no chance of ever being closed, so I'm changing it to something specific. :)

In particular, the JIT code emitter interface should:

1. Not require one virtual method call for every byte (on x86) or word (on RISCs) emitted.
2. The JIT emitter should support deallocating memory for previously compiled functions.
3. The JIT emitter should not preallocate an ungrowable 16M chunk of memory for function bodies.

I am scheduled to start working on this the first week or two of May.

-Chris

[lattner]

mine

[lattner]

#​1 is done.

[lattner]

Finally, at long last, this patch implements ExecutionEngine::freeMachineCodeForFunction.
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20060508/034740.html

-Chris