Missing calling convention support for vector types

[jsonn]

Bugzilla Link	46073
Version	trunk
OS	Linux
Attachments	Minimal test case
CC	@glaubitz,@jrtc27,@rorth

Extended Description

Try compiling the attached code, it will assert in SparcTargetLowering::LowerCall_32 respectively LowerCall_64 due to the missing lowering for 16x16 vector types.

[llvmbot]

This (unfortunately) happens in real world code and breaks firefox compilation for me.

[glaubitz]

Maybe we could look at how GCC does the lowering here.

[llvmbot]

AFAICT the ABI does not know about VIS data types at all, they are just passed like any other bunch of float (or double) values.

[glaubitz]

I would have been surprised to know that SPARC has any sorts of vector types.

[taiki-e]

According to calling conventions implemented by GCC:

• On SPARC32: 64-bit or smaller vector integer is passed using int reg (argument) / FP reg (return value)
• On SPARC64: 128-bit(argument)/256-bit(return value) or smaller vector integer/float are passed using FP reg

(sparc_pass_by_reference and sparc_return_in_memory have more comments about this.)