[PATCH] D158825: [WIP] Helper script for analyzing mir dumps

[Markus Lavin via Phabricator via llvm-commits]

markus created this revision.
Herald added subscribers: mstorsjo, s.egerton, PkmX, simoncook, arichardson.
Herald added a project: All.
markus requested review of this revision.
Herald added subscribers: llvm-commits, wangpc.
Herald added a project: LLVM.

Some ideas for helper scripts for analyzing mir dumps.

The dump files generated by `llc -print-after-all` tend to be rather large and this results in a number of problems. When comparing two runs simply launching your favorite diff tool on two files in the range of ~100k lines is not ideal. Diffing one pass and a subsequent one requires copy&paste to put the fragments in different files. When searching for e.g. registers it is easy to accidentally get a match outside the current pass of interest.

A more sophisticated approach would be to develop a simple python library that creates an in-memory representation of the dump file. E.g. the dump file is represented by a `DumpFile` object and then you have `Fragment` objects that reference an interval of lines of this `DumpFile`. The `Fragment` class can then be specialized into `Pass`, `Function` and `BasicBlock`. Defining some operations (like compare, view and diff) on the fragments allows for a rather flexible framework that can be used to build custom tools to ease the task of dump analysis by eliminating cumbersome and error prone manual steps.

Example usage

  int dot(int *a, int *b, int n) {
    int sum = 0;
    for (int i = 0; i < n; i++)
      sum += a[i] * b[i];
    return sum;
  }
  
  $ clang --target=riscv32 dot.c -O4 -S -mllvm -print-after-all 2> dump.txt
  
  # Get an overview of where changes occurred
  $ utils/dumptool --primary=dump.txt --list-pruned
  * 0 riscv-isel
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 ]
  * 4 slotindexes
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 ]
  * 5 stack-coloring
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 ]
  * 10 machine-sink
          *dot [ *bb.0 *bb.4 bb.1 *bb.2 bb.3 ]
  * 21 unreachable-mbb-elimination
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 *bb.4 ]
  * 22 livevars
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 *bb.4 ]
  * 23 phi-node-elimination
          *dot [ bb.0 *bb.1 *bb.2 *bb.3 *bb.4 ]
  * 24 twoaddressinstruction
          *dot [ bb.0 bb.1 bb.2 bb.3 bb.4 ]
  * 25 slotindexes
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 *bb.4 ]
  * 26 liveintervals
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 *bb.4 ]
  * 27 register-coalescer
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 *bb.4 ]
  * 30 livedebugvars
          *dot [ bb.0 bb.1 bb.2 bb.3 bb.4 ]
  * 35 virtregrewriter
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 *bb.4 ]
  * 38 machine-cp
          *dot [ *bb.0 *bb.1 *bb.2 *bb.3 *bb.4 ]
  * 44 shrink-wrap
          *dot [ bb.0 bb.1 bb.2 *bb.3 *bb.4 *bb.5 ]
  * 47 branch-folder
          *dot [ *bb.0 *bb.1 *bb.2 ]
  * 50 postrapseudos
          *dot [ *bb.0 bb.1 *bb.2 ]
  * 64 machine-sanmd
          dot [ bb.0 bb.1 bb.2 ]
  * 65 stack-frame-layout
          dot [ bb.0 bb.1 bb.2 ]
  * 70 unpack-mi-bundles
          dot [ bb.0 bb.1 bb.2 ]
  
  # Examine bb.4 of function dot after machine-sink
  
  $ utils/dumptool --primary=dump.txt --primary-frag=10/dot/bb.4 --view
  
  # Diff function dot before and after machine sink
  
  $ utils/dumptool --primary=dump.txt --secondary=dump.txt --primary-frag=9/dot --secondary-frag=10/dot --diff


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D158825

Files:
  llvm/utils/dumptool
  llvm/utils/dumputils.py

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D158825.553411.patch
Type: text/x-patch
Size: 7951 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230825/3d4541e7/attachment.bin>