Inverting a boolean should use eor about llvm-mos

[mysterymath]

Yep, fair criticism. I ran this through godbolt, looks like it's coming down from the frontend as a 1-bit XOR:

# Machine code for function invert: IsSSA, TracksLiveness

bb.1.entry:
  liveins: $a
  %1:_(s8) = COPY $a
  %2:_(s8) = G_ASSERT_ZEXT %1:_, 1
  %0:_(s1) = G_TRUNC %2:_(s8)
  %3:_(s1) = G_CONSTANT i1 true
  %4:_(s1) = G_XOR %0:_, %3:_; example.c:3:12
  %5:_(s8) = G_ZEXT %4:_(s1); example.c:3:5
  $a = COPY %5:_(s8); example.c:3:5
  RTS implicit $a; example.c:3:5

# End machine code for function invert.

Our legalizer is lowering it to the following select:

# Machine code for function invert: IsSSA, TracksLiveness, Legalized

bb.1.entry:
  liveins: $a
  %1:_(s8) = COPY $a
  %2:_(s8) = G_ASSERT_ZEXT %1:_, 1
  %0:_(s1) = G_TRUNC %2:_(s8)
  %3:_(s1) = G_CONSTANT i1 true
  %6:_(s1) = G_CONSTANT i1 false
  %4:_(s1) = G_SELECT %0:_(s1), %6:_, %3:_; example.c:3:12
  %7:_(s8) = G_CONSTANT i8 1
  %8:_(s8) = G_CONSTANT i8 0
  %5:_(s8) = G_SELECT %4:_(s1), %7:_, %8:_; example.c:3:5
  $a = COPY %5:_(s8); example.c:3:5
  RTS implicit $a; example.c:3:5

# End machine code for function invert.

This is because 1-bit XORs are just how the IR happens to represent negation of boolean flag values, and it's usually better to lower those to control flow (which usually folds away to nothing) than to widen to 8 bits, XOR, then truncate.

But, in this case, the 1-bit XOR is immediately zero-extended (ZEXT) to 8 bits, so we might as well just compute an 8-bit XOR. Should be a relatively simple thing to fix.

[mysterymath]

Fixed by 2b1d155

[mysterymath]

Had to revert the fix to this, as it causes a pretty severe regression elsewhere.

This will take some doing; the game here is surprisingly tricky, and I don't fully understand all the behavior of all the pieces yet.