Combine and-mask+icmp eq -> icmp ule when mask was all-ones in low bits

[LebedevRI]

Bugzilla Link	38123
Resolution	FIXED
Resolved on	Aug 31, 2018 06:28
Version	trunk
OS	Linux
CC	@topperc,@RKSimon,@rotateright

Extended Description

https://rise4fun.com/Alive/W2u
https://godbolt.org/g/sCX9A6

This pattern will be produced by Implicit Integer Trucation sanitizer,
(https://reviews.llvm.org/D48958)
in unsigned case, therefore it is probably a good idea to improve it.

[LebedevRI]

assigned to @LebedevRI

[LebedevRI]

https://godbolt.org/g/sCX9A6
That was of course should have been https://godbolt.org/g/eMV9aw

[rotateright]

This seems less likely, but there's always some kind of symmetrical pattern if you solve the problem generally:

Name: range using high bits
%masked = or i8 %arg, 192
%r = icmp ne i8 %masked, %arg
=>
%r = icmp ult i8 %arg, 192

?

[LebedevRI]

This seems less likely, but there's always some kind of symmetrical pattern
if you solve the problem generally:

Name: range using high bits
%masked = or i8 %arg, 192
%r = icmp ne i8 %masked, %arg
=>
%r = icmp ult i8 %arg, 192

?

Heh, yes, i was planning on handling more patterns.
These are all the patterns with low-bit-mask i can think of:
https://rise4fun.com/Alive/Rny

The patterns with high bit mask will be similar, but not everything at once..

[rotateright]

Not sure if we wanted this bug to cover all of the predicates, but let me add the example from D49179 using signed compares, so we don't forget it:

Pre: isPowerOf2(C1+1)
%masked = and i32 %arg, C1
%truncheck = icmp sge i32 %masked, %arg
=>
%truncheck = icmp sle i32 %arg, C1

https://rise4fun.com/Alive/K8H

[LebedevRI]

Not sure if we wanted this bug to cover all of the predicates, but let me
add the example from D49179 using signed compares, so we don't forget it:

Pre: isPowerOf2(C1+1)
%masked = and i32 %arg, C1
%truncheck = icmp sge i32 %masked, %arg
=>
%truncheck = icmp sle i32 %arg, C1

https://rise4fun.com/Alive/K8H

Ok, finished up the rest of unsigned predicates (for when mask is (-1 >> y)).

https://rise4fun.com/Alive/I3O <- following 4 folds are possible for signed predicates.

[LebedevRI]

Not sure if we wanted this bug to cover all of the predicates, but let me
add the example from D49179 using signed compares, so we don't forget it:

Pre: isPowerOf2(C1+1)
%masked = and i32 %arg, C1
%truncheck = icmp sge i32 %masked, %arg
=>
%truncheck = icmp sle i32 %arg, C1

https://rise4fun.com/Alive/K8H

Ok, finished up the rest of unsigned predicates (for when mask is (-1 >> y)).

https://rise4fun.com/Alive/I3O <- following 4 folds are possible for signed
predicates.

Done, too. In this case with signed predicates, we had to explicitly ignore
the versions with "commuted" sides of the comparison.
That was a lot of commits, sorry for the noise.

The part (-1 << y) is TODO.

Also, i wonder if we can/should do anything about
patterns like (-1 >> y) | (-1 << (width - 1))
I.e. a mask with some low bits set, and always with a sign bit set, too.

[LebedevRI]

This pattern with low-bit-mask was fixed via numerous commits.

I've filed llvm/llvm-bugzilla-archive#38791 to track the opposite pattern.

[LebedevRI]

mentioned in issue llvm/llvm-bugzilla-archive#38791