Propose to lower Store IR

[manji...@gmail.com]

Currently, the Store IR has the barrier built-in, like Store[(kRepTaggedSigned : NoWriteBarrier)](10, 43, 46, 36, 59).

And rely one the OutOfLineRecordWrite to lower the barrier.

Should we change it into simple Store without any built-in, and call RecordWrite builtin using turbofan?

That approach will make RecordWrite's calling convention more consistent, rather than saving all the float-points registers itself or restrict the register allocation.

[Yang Guo]

+Michael Stanton +Tobias Tebbi +Georg Neis 

--
--
v8-dev mailing list
v8-...@googlegroups.com
http://groups.google.com/group/v8-dev
---
You received this message because you are subscribed to the Google Groups "v8-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to v8-dev+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/v8-dev/70d469ee-4ab7-452e-a5f8-04a80c438a5c%40googlegroups.com.

[Tobias Tebbi]

Of course, it would be more uniform if we didn't use a different calling convention for RecordWrite, but it's important for both performance on the fast-path and generated code size:

By using a mostly callee-save calling convention, we can avoid many spills and unspills in generated code and do them instead in the slow-path of the RecordWrite builtin.

We don't generally support a callee-saved calling convention since it's difficult to support that in the presence of GC-triggering calls, but RecordWrite can never trigger allocation, so it's easier to do a callee-saved calling convention there. The restricted register allocation is essentially a hack to ensure that most registers are not modified by RecordWrite while still using Turbofan to generate the code.

[manji...@gmail.com]

Current implementation stores all the registers used as parameters before the call, then restores after the call.
If it's a turbofan call, we may get a chance to tell the register allocated that the registers as parameters get clobbered. That may generate better code. And that's what I have done in my llvm turbofan compiler.

[manji...@gmail.com]

the register allocator, sorry.

[manji...@gmail.com]

And for other case, like a function only uses one float point register, and the RecordWrite has to save all the float point registers except C's csr.