Camil Demetrescu edited osr-llvm.tex  over 8 years ago

Commit id: 68056446d941cc60f5aa6ca0a5837bbbe6d7ed5a

deletions | additions      

       

\fi  \paragraph{OSR Instrumentation in IR.}  We use deferred To defer the  compilation by instrumenting of the continuation function until the comparator is known at run time, we instrument  {\tt isord} with an open OSR point  at the beginning of the loop body, as shown in \myfigure\ref{fig:isordfrom}. Portions added to the original code by OSR istrumentation are highlighted in grey\footnote{Virtual register names and labels in the LLVM-produced IR code have been refactored to make the code more readable.}. %The figure illustrates how the original {\tt isord} code is instrumented by \tinyvm, highlighting in grey the added portions.   A new basic block is placed at the beginning of the loop body, which increments a hotness counter {\tt p.osr} and jumps to an OSR-firing block if the counter reaches the threshold (1000 iterations in this example). The OSR block contains a tail call to the target generation stub, which receives as parameters the four live variables at the OSR point ({\tt v}, {\tt n}, {\tt i}, {\tt c}). Notice that maintaining the SSA form requires adjusting $\phi$-nodes. The stub (see \myfigure[...]) calls a code generator that: 1) builds an optimized version of {\tt isord} by inlining the comparator (which is known when the OSR is fired), and 2) uses it to create the continuation function {\tt isordto} shown in \myfigure\ref{fig:isordascto}. The stub terminates with a tail call to {\tt isordto}. To generate the continuation function from the optimized version created by the inliner, we need to replace the function entry point, remove dead code, replace live variables with the function parameters, and fix $\phi$-nodes accordingly. Additions resulting from the IR instrumentation are in grey, while removals are struck-through.