On Tue, Sep 16, 2014 at 3:03 PM, James Greenhalgh <james.greenha...@arm.com> wrote: > On Tue, Sep 16, 2014 at 10:36:08PM +0100, Andrew Pinski wrote: >> On Thu, Sep 4, 2014 at 1:18 AM, James Greenhalgh >> <james.greenha...@arm.com> wrote: >> > On Thu, Sep 04, 2014 at 08:42:31AM +0100, Venkataramanan Kumar wrote: >> >> Hi maintainers, >> >> >> >> I just added "=r" and retested it. >> > >> > I had a very similar patch to this sitting in my local tree. However, >> > I am surprised you have left operand 3 as an output operand. In my tree >> > I had marked operand 3 as "&r". >> > >> > What do you think? >> >> The clobber needs to be "=&r" as you are writing to the register and >> not just reading from it. I think this is causing some issues >> including linaro bugzilla #667 >> (https://bugs.linaro.org/show_bug.cgi?id=667). > > (+CC Matthias Klose and Steve McIntyre who have also been in contact with me > regarding this bug) > > I've seen this bug locally, and had considered sending the patch you > suggested, which does indeed fix the bug. However, it feels wrong as > the operand is not a formal output of the pattern. It is clobbered - and > indeed earlyclobbered - so yes it is written to, but it isn't an output. > This makes the fix look like a band-aid around the real problem. > > The bug looks similar to pr52573 - regrename fails to spot that it should > not rename to a register used in an earlyclobber operand of any type, rather > than just an output+earlyclobber operand as it does now. > > I've played about with a fix that sits in regrename, and forces it to think > of all earlyclobber operands as starting and ending chains but this didn't > bootstrap clean - we end up with what I believe are false reports of stack > smashing in libstdc++. > > I was planning to look again at my approach tomorrow, I would like to > convince myself that this isn't a deficiency in regrename before I would > support just marking this operand "=&r". > > If you have any other suggestions, or if "=&r" is actually correct and > I am misreading the documentation please let me know.
I think you misread the documentation. Or rather the documentation is not fully clear here. https://gcc.gnu.org/onlinedocs/gccint/Modifiers.html says this for "&": ‘&’ does not obviate the need to write ‘=’ or ‘+’. Which means you need "=" if you write to the register. Match_scratch is the same as match_operand except it is able to be added back during combine. Also "=" means: Means that this operand is write-only for this instruction: the previous value is discarded and replaced by output data. That is not necessary a formal output of the pattern. Thanks, Andrew Thanks, Andrew Pinski > > Thanks, > James >
