On Tue, 14 May 2019, Richard Sandiford wrote:
> Several SVE patterns need define_insn_and_splits that generate the
> same insn_code, but with different operands. That's probably a
> niche requirement, but it's cropping up often enough on the ACLE
> branch that I think it would be good to have a
On 5/17/19 12:33 PM, Janne Blomqvist wrote:
--- snip ---
And yes, while I think one year might be a quite optimistic timeframe
to get this fixed, I agree we shouldn't keep the workaround
indefinitely either. I think the best way would be if CBLAS & LAPACKE
would be fixed, and then we could tell
On Sat, May 18, 2019 at 08:37:06PM +0200, Thomas Koenig wrote:
>
> the attached patch puts the workaround introduced by Jakub behind
> a flag. I debated with myself what the name should be.
> -fbroken-c-interfaces-die-die-die came to mind, but that would
> have penalized people who wanted to
On Sat, May 18, 2019 at 09:15:00PM +0200, Thomas Koenig wrote:
> Am 11.05.19 um 15:10 schrieb Thomas Koenig:
> > Hello world,
> >
> > this new version of the inlie argument packing patch (PR 88821)
> > avoids the ICE on the test case for PR 61968. Otherwise it is
> > unchanged.
> >
> >
This turns out to be obvious. The component of the dummy argument
cannot be replaced by the hidden descriptor (actually the class
object) and so an ICE occurs. The fix is to add a guard against the
expression being a COMPONENT_REF.
I will commit to all three branches tomorrow unless there are any
Am 11.05.19 um 15:10 schrieb Thomas Koenig:
Hello world,
this new version of the inlie argument packing patch (PR 88821)
avoids the ICE on the test case for PR 61968. Otherwise it is
unchanged.
Regression-tested. OK for trunk?
Ping?
Hello world,
the attached patch puts the workaround introduced by Jakub behind
a flag. I debated with myself what the name should be.
-fbroken-c-interfaces-die-die-die came to mind, but that would
have penalized people who wanted to disable the option :-)
Therefore, I have settled on
(@Feng Xue, it is better to include testcases in your patches)
I'm not a big fan of this patch. I'd rather be looking at either
improving our analysis
Better analysis cannot hurt.
or adding attributes to the loops to help the analysis bits prove
termination.
There may not be a loop in
On Wed, 15 May 2019, Richard Biener wrote:
As you write the heuristic you could as well remove the malloc result
points-to set from the others after points-to analysis is finished?
Looking at the vector testcase:
#include
#include
#include
inline void* operator new(std::size_t n){return
Hi Jakub,
Could we easily detect at resolve time whether a subroutine/function
makes any implicitly prototyped calls and limit this workaround to those
cases (i.e. only old-style Fortran)? I've mentioned it in the PR, but not
sure how exactly to check that.
I think we could to that. We
Darwin is able to use two runtimes for objective-c; the default
is its native "NeXT" runtime, but also it can build code using the
"gnu-runtime". In order to do this, we have to be able to find the
gnu-runtime headers (which are installed into the compiler's tree).
The process to do this has
Small part of catching up with changes made to the objective-c language.
The instancetype has been added to other objective-c implementations
as a typedef alias to id in order to allow diagnosis of cases where a class is
used or returned where an instance is expected.
This adds the typedef, and
One I’ve had hanging around in my local trees for a long time.
We have been emitting two section switches in the Darwin's
file end function by outputting the asm directly. This means that
varasm’s tracking of the section switched is not updated
which could matter if we elect to reorder some of
13 matches
Mail list logo
