On 14/11/2019 15:32, Kwok Cheung Yeung wrote:
This patch restricts non-kernel functions to using a maximum of 64 SGPRs
and 24 VGPRs.
Kernels can request various pieces of information from the HSA runtime,
and these will be loaded into the registers consecutively before the
kernel executes. These registers are normally fixed. Since non-kernel
functions cannot make these requests, they have to assume that the
default set of information has been requested. If a non-leaf kernel
requests information not in the defaults, a warning is now emitted as
pieces of info needed by callees may have shifted locations. A leaf
kernel can do whatever it wants.
I have setup FIXED_REGISTERS for the default case now - if a different
set of startup info is requested (which should be rare), then the set of
fixed registers will be adjusted accordingly by
gcn_conditional_register_usage. Compared to before, v0, s2 and s3 are
now unfixed (due to the newlib patch 'Stash reent marker in upper bits
of s1 on AMD GCN' and the first patch in this series).
Okay to commit?
I think this is fine, but please check the white-space and formatting.
There's something odd going on. Probably the patch got mangled by
Thunderbird. Please post future patches as attachments. I believe the
GCC guidelines request using the "inline" content disposition setting
you can only find in the TB config editor, but any kind of attachment is
OK with me.
Andrew
