Re: RISC-V glibc port, v5

[Jim Wilson <jimw at sifive dot com>]

On 01/25/2018 09:06 AM, Joseph Myers wrote:
> On Thu, 25 Jan 2018, DJ Delorie wrote:
>
> > > In my experience, testing MIPS with QEMU takes about three hours per ABI.
> >
> > For reference, on the fpga (quad core 64-bit) building glibc natively
> > took a day, and my tests ran for four days before finally hanging
>
> I don't suggest testing natively on QEMU or FPGA; just executing tests
> there but compiling on a suitably faster system.

There is a known problem with the RISC-V system qemu.  It is flushing 
the TLBs unnecessarily, and this causes it to be much slower than it 
should be.  I did a native gcc-7-branch bootstrap on a fedora disk image 
running on qemu.  It took over 3 days.  There is a pending patch to fix 
the TLB flushing issue, which claims to make dd run 10 times as fast, 
but there is some debate about whether the patch is correct.  Also, the 
underlying code is being rewritten while we try to fix it as we are also 
trying to upstream QEMU at the same time, which requires merging into a 
newer qemu version, which makes the fix more complicated.  This doesn't 
affect user mode QEMU, but glibc testing is working poorly under user 
mode qemu.  Meanwhile, glibc testing on a RISC-V system qemu is working 
OK, but is much slower than it should be, even with the compiles on a 
fast x86_64 machine.

Jim