[PATCH] D128123: [SDAG] try to replace subtract-from-constant with xor

[Bjorn Pettersson via Phabricator via llvm-commits]

bjope added a comment.

Not sure really if this ends up as a regression or not, but with this source we can see some differences if for example using heaxagon or systemz as targets.
(This example is related to doing sub->xor in instcombine so not exactly the patch in this review.)

Input C program (sub_xor.c):

  static unsigned ARR[100];
  
  extern void populate(unsigned*);
  
  unsigned foo(unsigned *a) {
      populate(ARR);
      unsigned sum = 0;
      for (int j = 1; j < 4; ++j) {
          unsigned *ptr = &ARR[99];
          for (int i = 0; i < 70; ++i) {
              sum += j * *(ptr - i);
          }
      }
      return sum;
  }

Compile with `clang -target hexagon -O2 sub_xor.c -o sub_xor.ll -emit-llvm -S` and the only difference will be some sub:s ending up as xor (depending on if the instcombine transform from commit 79bb915fb60b2cd2 <https://reviews.llvm.org/rG79bb915fb60b2cd220d89e3bb54f67abb8cdb7bd> is reverted or not).
If also comparing the result of `llc -O2 sub_xor.ll -o - -print-after-all`, given those two different versions of sub_xor.ll from prior step, we can see that Loop Strength Reduce can eliminate the sub but it does not eliminate the xor.

Well, this is the best in-tree example I got so far (when aiming at having a full test from C code, I guess this could be reduced to a much smaller test just running LSR to show that we get different result for sub vs xor).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D128123/new/

https://reviews.llvm.org/D128123