accept explicit alignment for kernel arguments

[whchung]

I'm wondering would it be possible to let HSAIL backend support manually specifying alignment for kernel arguments?

For example, if C++ AMP emit the following IR:

define internal cc76 void @foo(i32 align 8) {
ret void
}

then HSAIL backend will generate the following HSAIL:

prog kernel &foo(
align(8) kernarg_u32 %__arg_p1
) {
ret;
}

The rationale is because in C++AMP runtime we use std::vector to hold all kernel arguments prior to kernel launch. This is copied from Okra runtime implementation. And right now we use "sed" to hack HSAIL emitted by HSAIL backend to make sure all kernargs are 8-bytes aligned. But if we move to BRIG then such hack won't work anymore.

[alienanthill]

Currently, is the back-end stripping the alignments information?

[whchung]

Yes, the alignment information is ignored by HSAIL backend right now.