[llvm-commits] [llvm] r95466 - in /llvm/trunk: bindings/ocaml/llvm/llvm.ml bindings/ocaml/llvm/llvm.mli docs/LangRef.html include/llvm-c/Core.h include/llvm/Attributes.h include/llvm/Transforms/IPO/InlinerPass.h lib/AsmParser/LLLexer.cpp lib/AsmParser/LLParser.cpp lib/AsmParser/LLToken.h lib/Target/CppBackend/CPPBackend.cpp lib/Transforms/IPO/Inliner.cpp lib/VMCore/Attributes.cpp utils/llvm.grm utils/vim/llvm.vim

[Jakob Stoklund Olesen]

On Feb 5, 2010, at 10:45 PM, Nick Lewycky wrote:

> Jakob, it sounds like this is actually experimental. Assuming you don't make any progress between now and the time the branch is created, should this patch be backed out of the 2.7 release? Realize that if we don't, we're stuck with a bit allocated to the InlineHint function attr in the bitcode until 3.x.

There are two separate issues here:

1. Should the C99 and C++ inline keyword affect the optimizer?
2. If so, how?

The consensus is that the first answer is yes, the inline keyword should affect the optimizer. Without profiling data it is very difficult to make the right size/performance tradeoff, and the inline keyword can help.

The experimental part is how the inline hint should be used. My initial attempt essentially switches the inliner to -O3 mode for functions with the inlinehint attribute. We'll se how that goes.

I think that we do need this bit in the bitcode files. The experimental part is the details of what it does.

/jakob