Hello,
I am working on a pass that finds identical basic blocks and abstracts
them into functions, in order to reduce code size.
To decide, whether particular basic block group is worth factoring
out, I want to estimate code size of some function that uses one of
basic blocks in question 'as is', and after factoring the basic block
into a new function.
Can you please suggest a proper way to do this?
Gregory
On Fri, Jun 4, 2010 at 2:46 AM, Gregory Petrosyan <gregory.petrosyan at gmail.com> wrote:> Hello, > > I am working on a pass that finds identical basic blocks and abstracts > them into functions, in order to reduce code size. > > To decide, whether particular basic block group is worth factoring > out, I want to estimate code size of some function that uses one of > basic blocks in question 'as is', and after factoring the basic block > into a new function. > > Can you please suggest a proper way to do this?You could just count the number of Instructions as an extremely rough estimate. llvm/Analysis/InlineCost.h has the functions used by the inliner and other similar transforms to estimate the cost of a basic block or function, which are somewhat more accurate. -Eli
On Fri, Jun 4, 2010 at 3:47 PM, Eli Friedman <eli.friedman at gmail.com> wrote:> On Fri, Jun 4, 2010 at 2:46 AM, Gregory Petrosyan > <gregory.petrosyan at gmail.com> wrote: >> Hello, >> >> I am working on a pass that finds identical basic blocks and abstracts >> them into functions, in order to reduce code size. >> >> To decide, whether particular basic block group is worth factoring >> out, I want to estimate code size of some function that uses one of >> basic blocks in question 'as is', and after factoring the basic block >> into a new function. >> >> Can you please suggest a proper way to do this? > > You could just count the number of Instructions as an extremely rough > estimate. llvm/Analysis/InlineCost.h has the functions used by the > inliner and other similar transforms to estimate the cost of a basic > block or function, which are somewhat more accurate.Thanks Eli, The thing that worries me in this approach is that it is really rough, and to make it work I need to introduce and tune some target-specific coefficients. I thought that if there is a way to just generate real code and measure it's size, it would be simpler as well as more accurate. Gregory