[LLVMdev] [cfe-dev] RFC: Codifying (but not formalizing) the optimization levels in LLVM and Clang
kparzysz at codeaurora.org
Mon Jan 21 11:26:14 PST 2013
On 1/21/2013 1:00 PM, Michael Ilseman wrote:
> It's different in that optimizations here are passes to be ran, and fast-math is a semantic specifier. Currently, clang has fast-math support via a command-line argument (-ffast-math), which sets fath-math flags on individual LLVM IR instructions. Those flags don't govern the running or scheduling of passes, but provide (relaxed) semantics for the optimizations themselves to obey.
Ok, so maybe fast-math was a bad example. Also, if the part of the
discussion about inlining was only meant to be limited to -O[...], then
I misunderstood it, and my concerns may not apply.
However, if we want to reflect things like optimization plans (i.e. more
than just -O, but things like "-Oparallel", for example) via
attributes, then we do run into the issue of how inlining will interact
with the attributes. Both, parallelization and vectorization may
require different sets of optimizations, and either of them may affect
the semantics. Will we be forced to have them happen before inlining?
What about the "section" attribute? Inlining a function from one
section into a function from another section may not be what the user
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation
More information about the llvm-dev