[LLVMdev] Postponing more passes in LTO

[Krzysztof Parzyszek]

On 12/18/2014 6:10 AM, Sean Silva wrote:
> [...]

Pass ordering is a design decision, based on assumptions as to what each 
of the optimizations is going to do, what information it needs, what 
properties of the code will be created, preserved or eliminated, and so on.
What Daniel showed are the performance results from reordering passes, 
with keeping their current implementations. This serves more as an 
indicator of the potential impact of such change if it was made today. 
You're right that relying only on the statistics may lead to incorrect 
conclusions, but I don't think that there is any meaningful information 
that could be extracted through a more detailed analysis of the provided 
data. The question is really whether changing the pass order could allow 
the optimizations to be more aggressive or accurate (e.g. through 
exposing them to a more complete view of the program), and whether the 
benefits could be expected to exceed the costs (such as increased 
compilation time). Whether the current implementations of particular 
optimizations would immediately benefit from such a change is not really 
that important in the long term.

-Krzysztof

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, 
hosted by The Linux Foundation