[all-commits] [llvm/llvm-project] 3562ad: BPF: implement isLegalAddressingMode() properly

[yonghong-song via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 3562ad3ebeb29d0cf320bfcf51f64372e9b37c95
      https://github.com/llvm/llvm-project/commit/3562ad3ebeb29d0cf320bfcf51f64372e9b37c95
  Author: Yonghong Song <yhs at fb.com>
  Date:   2021-09-30 (Thu, 30 Sep 2021)

  Changed paths:
    M llvm/lib/Target/BPF/BPFISelLowering.cpp
    M llvm/lib/Target/BPF/BPFISelLowering.h

  Log Message:
  -----------
  BPF: implement isLegalAddressingMode() properly

Latest upstream llvm caused the kernel bpf selftest emitting the
following warnings:

  In file included from progs/profiler3.c:6:
  progs/profiler.inc.h:489:2: warning: loop not unrolled:
    the optimizer was unable to perform the requested transformation;
    the transformation might be disabled or specified as part of an unsupported
    transformation ordering [-Wpass-failed=transform-warning]
          for (int i = 0; i < MAX_PATH_DEPTH; i++) {
          ^

Further bisecting shows this SimplifyCFG patch ([1]) changed
the condition on how to fold branch to common dest. This caused
some unroll pragma is not honored in selftests/bpf.

The patch [1] test getUserCost() as the condition to
perform the certain basic block folding transformation.
For the above example, before the loop unroll pass, the control flow
looks like:
    cond_block:
       branch target: body_block, cleanup_block
    body_block:
       branch target: cleanup_block, end_block
    end_block:
       branch target: cleanup_block, end10_block
    end10_block:
       %add.ptr = getelementptr i8, i8* %payload.addr.0, i64 %call2
       %inc = add nuw nsw i32 %i.0, 1
       branch target: cond_block

In the above, %call2 is an unknown scalar.

Before patch [1], end10_block will be folded into end_block, forming
the code like
    cond_block:
       branch target: body_block, cleanup_block
    body_block:
       branch target: cleanup_block, end_block
    end_block:
       branch target: cleanup_block, cond_block
and the compiler is happy to perform unrolling.

With patch [1], getUserCost(), which calls getGEPCost(), which calls
isLegalAddressingMode() in TargetLoweringBase.cpp, considers IR
  %add.ptr = getelementptr i8, i8* %payload.addr.0, i64 %call2
is free, so the above basic block folding transformation is not performed
and unrolling does not happen.

For BPF target, the IR
  %add.ptr = getelementptr i8, i8* %payload.addr.0, i64 %call2
is not free and we don't have ld/st instruction address with 'r+r' mode.

This patch implemented a BPF hook for isLegalAddressingMode(), which is
identical to Mips isLegalAddressingMode() implementation where
the address pattern like 'r+r', 'r+r+i' or '2*r' are not allowed.
With testing kernel bpf selftests, all loop not unrolled warnings
are gone and all selftests run successfully.

  [1] https://reviews.llvm.org/D108837

Differential Revision: https://reviews.llvm.org/D110789