[InstCombine] Failure to simplify add-of-negative+underflow check+null check of non-null base pointer

[LebedevRI]

Bugzilla Link	43259
Version	trunk
OS	Linux
CC	@RKSimon,@nikic,@rotateright,@vedantk
Fixed by commit(s)	372491, 372767, 372768

Extended Description

Again comes up in https://reviews.llvm.org/D67122
Much like what we have in 43251,
for

#include <cassert>
char* test(char& base, signed long offset) {
  __builtin_assume(offset < 0);
  return &base + offset;
}

We produce

https://godbolt.org/z/r40U47

and again those two icmp's can be merged:
https://rise4fun.com/Alive/8h2

[LebedevRI]

Hm, and there's also the predicate when we don't need is-not-zero check
https://rise4fun.com/Alive/Cjru
https://godbolt.org/z/WFnBO0

[LebedevRI]

Okay, looking into this now. For ule/ugt the folds are:

Name: 0
  %adjusted = add i8 %base, %offset
  %not_null = icmp ne i8 %adjusted, 0
  %no_underflow = icmp ule i8 %adjusted, %base
  %r = and i1 %not_null, %no_underflow
=>
  %neg_offset = sub i8 0, %offset
  %r = icmp ugt i8 %base, %neg_offset

Name: 1
  %adjusted = add i8 %base, %offset
  %not_null = icmp ne i8 %adjusted, 0
  %no_underflow = icmp ule i8 %adjusted, %base
  %r = or i1 %not_null, %no_underflow
=>
  %r = true

Name: 2
  %adjusted = add i8 %base, %offset
  %not_null = icmp eq i8 %adjusted, 0
  %no_underflow = icmp ule i8 %adjusted, %base
  %r = and i1 %not_null, %no_underflow
=>
  %r = %not_null

Name: 3
  %adjusted = add i8 %base, %offset
  %not_null = icmp eq i8 %adjusted, 0
  %no_underflow = icmp ule i8 %adjusted, %base
  %r = or i1 %not_null, %no_underflow
=>
  %r = %no_underflow


Name: 4
  %adjusted = add i8 %base, %offset
  %not_null = icmp ne i8 %adjusted, 0
  %no_underflow = icmp ugt i8 %adjusted, %base
  %r = and i1 %not_null, %no_underflow
=>
  %r = %no_underflow

Name: 5
  %adjusted = add i8 %base, %offset
  %not_null = icmp ne i8 %adjusted, 0
  %no_underflow = icmp ugt i8 %adjusted, %base
  %r = or i1 %not_null, %no_underflow
=>
  %r = %not_null

Name: 6
  %adjusted = add i8 %base, %offset
  %not_null = icmp eq i8 %adjusted, 0
  %no_underflow = icmp ugt i8 %adjusted, %base
  %r = and i1 %not_null, %no_underflow
=>
  %r = false

Name: 7
  %adjusted = add i8 %base, %offset
  %not_null = icmp eq i8 %adjusted, 0
  %no_underflow = icmp ugt i8 %adjusted, %base
  %r = or i1 %not_null, %no_underflow
=>
  %neg_offset = sub i8 0, %offset
  %r = icmp ule i8 %base, %neg_offset

[LebedevRI]

... and same is valid with flipped-strictness predicate if we know that one of the values being added is non-zero:
https://rise4fun.com/Alive/2AJ
https://rise4fun.com/Alive/YF4

[LebedevRI]

The instsimplify part for predicates that do not require knownbits were already handled by simplifyUnsignedRangeCheck() with the exception of a single case, fixed in r372491.