SSE XMM0 ABI #923
SSE XMM0 ABI #923
Conversation
…rs. This new casting operation is used to extract xmm registers from ymm registers in order to support monomorphic functions in sse. The SysV ABI is updated to recognize functions that read and write to xmm0.
There was a problem hiding this comment.
The flt function may terminate with an out-of-the-bound exception, we need to fix it.
Additionally, if you think that we need float right now, then it is easy to extend the code with it, the only difference is that we need to cast them to 32 bit instead of 64.
plugins/x86/x86_abi.ml
Outdated
| let flt = function | ||
| | Ret_0 -> xmm 0 | ||
| | Ret_1 -> xmm 1 | ||
| | Arg n -> xmm n |
There was a problem hiding this comment.
This function will blow up with an exception if we will try to pass more than 16 floats, and will be just incorrect if we will try to pass more than 8, as according to SysV AMD64 ABI only xmms0-xmms7 are used for arguments passing, while the rest XMM8-XMM15 are used as scratch registers. So it should be something like this:
| Arg n -> if n < 8 then xmm n else stack Int.(n-8)Note, this is still an under-approximation, as it will only work for monomorphic functions, but better than terminating analysis with an exception :)
The following code allows ymm registers to be cast to smaller register. This new casting operation is used to extract xmm registers from ymm registers in order to support monomorphic functions in sse. The SysV ABI is extended to receive more arguments and can now recognize functions that read and write to xmm0.