RFR: 8240363: Refactor Compile::Output() to its own Phase

[Vladimir Kozlov]

First, I like the idea of this refactoring.

I am fine with moving _orig_pc_slot intialization to PhaseOutput where it is only used. fixed_slots() should not be 
modified before Output() is called.

I have concern about indirections you are adding:

C->output()->

We need Claes opinion on this. On other hand it affect only code generation speed.

Thanks,
Vladimir

On 3/5/20 2:27 AM, Erik Österlund wrote:
> Hi,
> 
> The Compile::Output function comes with a lot of baggage (state and related functions), that is
> mostly unrelated to other parts of a compilation unit. This cluttering of unrelated stuff on
> Compile makes the code messy.This is a proposal to break out that state and embed it into a
> separate Phase.
> 
> This patch moves ~800 LOC away from Compile. There are more opportunities for refactoring further,
> but my intention with this patch is to focus on things belonging to Output() in particular. I tried
> to be as mechanical as possible, and not change logic.
> 
> One slight difference in logic is that we used to advance the frame_slots before calling Code_Gen(),
> and stash away the _orig_pc_slot to the value before it was advanced, way before output is called.
> In order to have PhaseOutput live only across Output(), I changed this to advancing the frame_slots
> as usual before Code_Gen(), and during PhaseOutput, I set the _orig_pc_slot to the previous frame_slot
> (which now points at the sp). That bit of untangling allowed PhaseOutput to live only across the Output()
> call, followed by code installation (which I moved into Code_Gen, to narrow down the life time of
> PhaseOutput).
> 
> People maintaining non-Oracle platforms might want to have a go at compiling this to see if I managed
> to catch all code motion in the AD files.
> 
> Bug:
> https://bugs.openjdk.java.net/browse/JDK-8240363
> 
> Webrev:
> http://cr.openjdk.java.net/~eosterlund/8240363/webrev.00/
> 
> Testing:
> hs-tier{1-3}
> 
> Thanks,
> /Erik