[AVR] Rewrite the function calling convention.
authorDylan McKay <me@dylanmckay.io>
Fri, 19 Jun 2020 11:26:00 +0000 (23:26 +1200)
committerDylan McKay <me@dylanmckay.io>
Tue, 23 Jun 2020 09:36:18 +0000 (21:36 +1200)
commitb9c26a9cfe53da7ef96e13ef1aa7fe793b1b5d28
tree0e6389626733f0d897b51bfe5c4047af3613e268
parent82a882db083236a02f7d12fc7016bddb2b47488d
[AVR] Rewrite the function calling convention.

Summary:
The previous version relied on the standard calling convention using
std::reverse() to try to force the AVR ABI. But this only works for
simple cases, it fails for example with aggregate types.

This patch rewrites the calling convention with custom C++ code, that
implements the ABI defined in https://gcc.gnu.org/wiki/avr-gcc.

To do that it adds a few 16-bit pseudo registers for unaligned argument
passing, such as R24R23. For example this function:

    define void @fun({ i8, i16 } %a)

will pass %a.0 in R22 and %a.1 in R24R23.

There are no instructions that can use these pseudo registers, so a new
register class, DREGSMOVW, is defined to make them apart.

Also the ArgCC_AVR_BUILTIN_DIV is no longer necessary, as it is
identical to the C++ behavior (actually the clobber list is more strict
for __div* functions, but that is currently unimplemented).

Reviewers: dylanmckay

Subscribers: Gaelan, Sh4rK, indirect, jwagen, efriedma, dsprenkels, hiraditya, Jim, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D68524

Patch by Rodrigo Rivas Costa.
llvm/lib/Target/AVR/AVRCallingConv.td
llvm/lib/Target/AVR/AVRISelLowering.cpp
llvm/lib/Target/AVR/AVRISelLowering.h
llvm/lib/Target/AVR/AVRInstrInfo.cpp
llvm/lib/Target/AVR/AVRRegisterInfo.td
llvm/test/CodeGen/AVR/calling-conv/c/basic_aggr.ll [new file with mode: 0644]
llvm/test/CodeGen/AVR/calling-conv/c/call.ll [new file with mode: 0644]
llvm/test/CodeGen/AVR/calling-conv/c/call_aggr.ll [new file with mode: 0644]
llvm/test/CodeGen/AVR/calling-conv/c/return_aggr.ll [new file with mode: 0644]