[llvm-dev] [RFC] Target-Independent MC-Layer Compiler Barrier

[Sam Elliott via llvm-dev]

Hi LLVM-Dev,

I recently had a bug report on the RISC-V backend that Rust's `compiler_fence` (which generates the same IR as C/C++'s `atomic_signal_fence`) was emitting machine code when it didn't need to.

Most targets seem to handle this the same way, in that they introduce a custom node (often called MEMBARRIER) that expands to an assembly comment, in order to implement this. I could have done the same for RISC-V, but it felt a bit like we were copy/pasting what other backends do. 

Instead, I prepared a patch which adds a TargetOpcode::COMPILER_BARRIER, to be used as a target-independent compiler barrier, which just expands to an assembly comment (but no instructions). I've tried to implement this in a way that allows backends with more complex memory models to avoid introducing no instructions where that would be incorrect (for instance, AMDGPU has more complex requirements on what syncscope::singlethread means).

I have prepared a patch, here: https://reviews.llvm.org/D92842 

This patch adds the feature, and moves the RISC-V, x86, Arm and AArch64 backends to use this feature, instead of their target-specific equivalents. 

I would welcome feedback on the approach in the patch, or any concerns you have about this approach.

Thanks in advance,

Sam

--
Sam Elliott
Software Team Lead
Senior Software Developer - LLVM and OpenTitan
lowRISC CIC

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201210/78180a57/attachment-0001.html>