[llvm-dev] Mips unconditionally uses fast-isel?

[Daniel Sanders via llvm-dev]

> -----Original Message-----

> From: Robinson, Paul [mailto:Paul_Robinson at playstation.sony.com]

> Sent: 17 November 2015 22:58

> To: Daniel Sanders; llvm-dev at lists.llvm.org

> Subject: RE: Mips unconditionally uses fast-isel?

>

> > > The other thing that might work, is having TargetMachine remember how

> > > the fast-isel option got set, and make OptLevelChanger do the right

> > > thing. But that seems like a hack to work around Mips not obeying the

> > > specified optimization level, honestly.

> >

> > I think we should do that as well. I don't think it's right that optnone

> > enables Fast ISel even when it's been explicitly disabled. It should do

> > the same checks as addPassesToGenerateCode() does.

>

> Hm?  What you're asking for is that "-O2" and "-O2 -fast-isel=none" are

> identical, unless you have an 'optnone' function. Do you really have a

> use-case for controlling the codegen path for an 'optnone' function?

> The whole point of 'optnone' is to avoid optimizations.

> --paulr

>



No, that's already true. -O2 doesn't try to enable Fast ISel (unless the optnone attribute is given) so -fast-isel=false has no effect.



I'm saying that optnone means 'use -O0 for this function' and that optnone should

respect non-default values of the -fast-isel flag like -O0 does. This is the behaviour I'd expect:




-fast-isel=false


-fast-isel=default


-fast-isel=true


-O0


SelectionDAG


FastISel


FastISel


-O0 + optnone attribute


SelectionDAG*


FastISel


FastISel


-O1 + optnone attribute


SelectionDAG*


FastISel


FastISel


-O2 + optnone attribute


SelectionDAG*


FastISel


FastISel


-O3 + optnone attribute


SelectionDAG*


FastISel


FastISel


-O1


SelectionDAG


SelectionDAG


FastISel


-O2


SelectionDAG


SelectionDAG


FastISel


-O3


SelectionDAG


SelectionDAG


FastISel


The cells marked with '*' differ from the current behaviour.



In terms of code, I think this part of OptLevelChanger::OptLevelChanger():

   if (NewOptLevel == CodeGenOpt::None) {

      DEBUG(dbgs() << "\nEnable FastISel for Function "

            << IS.MF->getFunction()->getName() << "\n");

      IS.TM.setFastISel(true);

    }

Should be:

    if (NewOptLevel == CodeGenOpt::None) {

      DEBUG(dbgs() << "\nEnable FastISel for Function "

            << IS.MF->getFunction()->getName() << "\n");

      IS.TM.setFastISel(EnableFastISelOption != cl::BOU_FALSE);

    }

Where EnableFastISelOption has the same value as the global in LLVMTargetMachine.cpp



The main reason I'm asking for this is that I think it's weird to for optnone to use a different code generator

than  -O0. These hidden overrides exist to help us debug code generation problems and, faced with a code

generation bug, -fast-isel=false is useful for quickly determining whether it's in FastISel or somewhere else.

The current behaviour allows optnone to overrule the hidden option to force-disable FastISel which will give

misleading guidance for bugs that lie in functions with optnone.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20151118/3196323f/attachment-0001.html>