[LLVMdev] Runtime optimization and other doubts

[Chris Lattner]

On Wed, 2 Nov 2005, Sandra Johnson wrote:
> $lli -load ../../../mypass.so test1.bc
> using JIT and it worked perfectly.
> Now I am trying to generate static native code generation and then execute the
> native code program as:
> $opt -load ../../../mypass.so -myoption test.bc > test1.bc
> $llc -load ../../../mypass.so test1.bc -o test1.s
> $gcc test1.s -o test1.native -lpthread
> but now I get the following error
> /tmp/cc8ilAEp.o(.text+0x251): In function `main':
> : undefined reference to `llvm_profile_init'

Try

gcc test1.s -o test1.native -lpthread ../../../mypass.so

Make sure that ../../.. is in your LD_LIBRARY_PATH when running your 
program.

> I am also interested in knowing more about profile driven runtime optimizations

There is one profile driven optimization (basic block layout), it does not 
run at runtime though, and is not enabled by default.

> I found that optimizations of FunctionPass or MachineFunctionPass when 
> used with JIT are executed only once for each function. Am I wright ?

Yes.

> Is there any pass which will be executed each time a function is called 
> or whenever it is needed i.e. during runtime ?

No.  You could use a function pass that inserts a function call into each 
function though, that would give you control every time each function is 
entered.

> Is it possible to inline/ clone the function after the machine code is 
> generated ?

No, not currently.  High-level optimizations like inlining are best 
performed on the LLVM IR level, not the machine code level.

-Chris

-- 
http://nondot.org/sabre/
http://llvm.org/