Commit Graph

4 Commits

Author SHA1 Message Date
Sanjay Patel e77ba263fe [InstSimplify] peek through 'not' operand in logic-of-icmps fold
This extends D78430 to solve cases like:
https://llvm.org/PR47858

There are still missed opportunities shown in the tests,
and as noted in the earlier patches, we have related
functionality in InstCombine, so we may want to extend
other folds in a similar way.

A semi-random sampling of test diff proofs in this patch:
https://rise4fun.com/Alive/sS4C
2020-10-25 11:13:30 -04:00
Sanjay Patel 7de2add829 [InstSimplify] add tests for logic-of-cmps with not op; NFC
One variant of this is shown in:
https://llvm.org/PR47858
2020-10-25 11:13:30 -04:00
Sanjay Patel e86eff0e82 [InstSimplify] fold and/or of compares with equality to min/max constant
I found 12 (6 if we compress the DeMorganized forms) patterns for logic-of-compares
with a min/max constant while looking at PR45510:
https://bugs.llvm.org/show_bug.cgi?id=45510

The variations on those forms multiply the test cases by 8 (unsigned/signed, swapped
compare operands, commuted logic operands).
We have partial logic to deal with these for the unsigned min (zero) case, but
missed everything else.

We are deferring the majority of these patterns to InstCombine to allow more general
handling (see D78582).

We could use ConstantRange instead of predicate+constant matching here. I don't
expect there's any noticeable compile-time impact for either form.

Here's an abuse of Alive2 to show the 12 basic signed variants of the patterns in
one function:
http://volta.cs.utah.edu:8080/z/5Vpiyg

declare void @use(i1, i1, i1, i1, i1, i1, i1, i1, i1, i1, i1, i1)
define void @src(i8 %x, i8 %y)  {
  %m1 = icmp eq i8 %x, 127
  %c1 = icmp slt i8 %x, %y
  %r1 = and i1 %m1, %c1   ; (X == MAX) && (X < Y) --> false

  %m2 = icmp ne i8 %x, 127
  %c2 = icmp sge i8 %x, %y
  %r2 = or i1 %m2, %c2    ; (X != MAX) || (X >= Y) --> true

  %m3 = icmp eq i8 %x, -128
  %c3 = icmp sgt i8 %x, %y
  %r3 = and i1 %m3, %c3   ; (X == MIN) && (X > Y) --> false

  %m4 = icmp ne i8 %x, -128
  %c4 = icmp sle i8 %x, %y
  %r4 = or i1 %m4, %c4    ; (X != MIN) || (X <= Y) --> true

  %m5 = icmp eq i8 %x, 127
  %c5 = icmp sge i8 %x, %y
  %r5 = and i1 %m5, %c5   ; (X == MAX) && (X >= Y) --> X == MAX

  %m6 = icmp ne i8 %x, 127
  %c6 = icmp slt i8 %x, %y
  %r6 = or i1 %m6, %c6   ; (X != MAX) || (X < Y) --> X != MAX

  %m7 = icmp eq i8 %x, -128
  %c7 = icmp sle i8 %x, %y
  %r7 = and i1 %m7, %c7   ; (X == MIN) && (X <= Y) --> X == MIN

  %m8 = icmp ne i8 %x, -128
  %c8 = icmp sgt i8 %x, %y
  %r8 = or i1 %m8, %c8   ; (X != MIN) || (X > Y) --> X != MIN

  %m9 = icmp ne i8 %x, 127
  %c9 = icmp slt i8 %x, %y
  %r9 = and i1 %m9, %c9    ; (X != MAX) && (X < Y) --> X < Y

  %m10 = icmp eq i8 %x, 127
  %c10 = icmp sge i8 %x, %y
  %r10 = or i1 %m10, %c10    ; (X == MAX) || (X >= Y) --> X >= Y

  %m11 = icmp ne i8 %x, -128
  %c11 = icmp sgt i8 %x, %y
  %r11 = and i1 %m11, %c11    ; (X != MIN) && (X > Y) --> X > Y

  %m12 = icmp eq i8 %x, -128
  %c12 = icmp sle i8 %x, %y
  %r12 = or i1 %m12, %c12    ; (X == MIN) || (X <= Y) --> X <= Y

  call void @use(i1 %r1, i1 %r2, i1 %r3, i1 %r4, i1 %r5, i1 %r6, i1 %r7, i1 %r8, i1 %r9, i1 %r10, i1 %r11, i1 %r12)
  ret void
}

define void @tgt(i8 %x, i8 %y)  {
  %m5 = icmp eq i8 %x, 127
  %m6 = icmp ne i8 %x, 127
  %m7 = icmp eq i8 %x, -128
  %m8 = icmp ne i8 %x, -128
  %c9 = icmp slt i8 %x, %y
  %c10 = icmp sge i8 %x, %y
  %c11 = icmp sgt i8 %x, %y
  %c12 = icmp sle i8 %x, %y
  call void @use(i1 0, i1 1, i1 0, i1 1, i1 %m5, i1 %m6, i1 %m7, i1 %m8, i1 %c9, i1 %c10, i1 %c11, i1 %c12)
  ret void
}

Differential Revision: https://reviews.llvm.org/D78430
2020-04-23 09:16:10 -04:00
Sanjay Patel 02b070ed49 [InstSimplify] add tests for logic-of-icmp with min/max constant; NFC
See PR45510:
https://bugs.llvm.org/show_bug.cgi?id=45510

We had partial coverage for some of these patterns, so removing duplicate tests
with the complete set in the new test file.
2020-04-19 08:24:38 -04:00