Page MenuHomePhabricator

[RISCV] Add LLVM intrinsics for the Bit Manipulation extension
Needs ReviewPublic

Authored by s.egerton on Aug 20 2019, 8:18 AM.

Details

Summary

This adds IR level intrinsics for the proposed Bit Manipulation extension.
Details on the extension can be found here: https://raw.githubusercontent.com/riscv/riscv-bitmanip/master/bitmanip-0.90.pdf

Event Timeline

s.egerton created this revision.Aug 20 2019, 8:18 AM
Herald added a project: Restricted Project. · View Herald TranscriptAug 20 2019, 8:18 AM
s.egerton edited the summary of this revision. (Show Details)Aug 21 2019, 1:52 AM
s.egerton added inline comments.
llvm/include/llvm/IR/IntrinsicsRISCV.td
75

These intrinsics all use llvm_anyint_ty. The arguments can currently only be either 32 bit or 64 bit and nothing else. Is there better way to achieve this without using llvm_anyint_ty?

99

I haven't added intrinsics for every new instruction. Either because they are too simple (andn, orn, xnor, min and max) or because there is already an IR intrinsic that has the same function (clz, ctz, pcnt, ror, rol, fsl and fsr).
Are there any other IR intrinsics that it would be better to remove?

llvm/lib/Target/RISCV/RISCVInstrInfoB.td
475

I considered using a multiclass for these GREVI patterns and a few other similar patterns in this file. Would this be a better approach?

xbolva00 added inline comments.
llvm/include/llvm/IR/IntrinsicsRISCV.td
99

InstCombine could benefit from smin/smax/umin/umax intrinsic

@spatel ?

spatel added inline comments.
llvm/include/llvm/IR/IntrinsicsRISCV.td
99

I agree, and I raised that possibility ~3 years ago (and it was debated well before that time too):
http://lists.llvm.org/pipermail/llvm-dev/2016-November/106868.html
...but the consensus was that we could pattern match our way out of it (and adjust the cost models to deal with it), so the benefit of adding intrinsics was not worth the cost.

If there's new evidence to suggest we should have those intrinsics, please send a note to llvm-dev. Then, it can be seen by more people. It's not clear to me how this patch review would be changed.

jeremybennett added inline comments.
llvm/include/llvm/IR/IntrinsicsRISCV.td
99

I note that GCC support for RISC-V bit manipulation extension defines a set of builtins (not yet upstream, available at https://github.com/riscv/riscv-bitmanip/blob/master/cproofs/rvintrin.h), and that this set of builtins has been agreed with Clifford Wolf, chair of the Bit Manipulation Task Group and author of the instruction set extension definition document.

Clang/LLVM should support this set as a minimum in order to ensure that code that works on GCC works out of the box on Clang/LLVM.

lebedev.ri added inline comments.
llvm/include/llvm/IR/IntrinsicsRISCV.td
99

A frontend(clang) intrinsic, LLVM IR instruction/intrinsic, backend RISCVISD node and finally RISCV instruction
are not strongly interdependant.

In order to have a frontend(clang) intrinsic/builtin, you don't need anything from the rest of the compiler,
you can either just write it in C, or add a builtin to clang that will emit some appropriate LLVM IR.

You don't need to have any specific LLVM IR instruction/intrinsic for every ISA extension,
that would be extremely intrusive; you at most want an RISCVISD node, and patternmatch it from
normal ISD nodes in RISCVISelLowering.cpp

s.egerton marked an inline comment as done.Sep 3 2019, 6:49 AM
s.egerton added inline comments.
llvm/include/llvm/IR/IntrinsicsRISCV.td
99

You raise a good point.
This has been discussed and we have decided to revisit our implementation strategy. There will be future patches to support this.

For future reference: Here is a link to similar discussions on the mailing list: https://lists.llvm.org/pipermail/llvm-dev/2019-August/134791.html

khchen added a subscriber: khchen.Dec 7 2020, 8:11 AM
rkruppe removed a subscriber: rkruppe.Dec 7 2020, 8:26 AM
lenary resigned from this revision.Jan 14 2021, 10:27 AM