[InstCombine] Missed @llvm.usub.sat opportunity

[LebedevRI]

Bugzilla Link	45160
Version	trunk
OS	Linux
CC	@RKSimon,@nikic,@rotateright

Extended Description

Given https://godbolt.org/z/gGcybp

using size_type = unsigned;

size_type t0(size_type pos, size_type size) {
        size_type bytesRemaining = (pos < size) ? size - pos : 0;
    return ((bytesRemaining > 4)
                       ? (size - pos - 4)
                       : 0);
}

we currently produce

define dso_local i32 @_Z2t0jj(i32 %0, i32 %1) local_unnamed_addr #0 {
  %3 = tail call i32 @llvm.usub.sat.i32(i32 %1, i32 %0)
  %4 = icmp ugt i32 %3, 4
  %5 = sub i32 -4, %0
  %6 = add i32 %5, %1
  %7 = select i1 %4, i32 %6, i32 0
  ret i32 %7
}

declare i32 @llvm.usub.sat.i32(i32, i32) #1

but we should instead produce

define dso_local i32 @_Z6squarejj(i32 %0, i32 %1) local_unnamed_addr #0 {
  %3 = tail call i32 @llvm.usub.sat.i32(i32 %1, i32 %0)
  %4 = tail call i32 @llvm.usub.sat.i32(i32 %3, i32 4)
  ret i32 %4
}
declare i32 @llvm.usub.sat.i32(i32, i32) #1

----------------------------------------
define i8 @_Z6squarejj(i8 %0, i8 %1) {
%2:
  %3 = usub_sat i8 %1, %0
  %4 = icmp ugt i8 %3, 4
  %5 = sub i8 252, %0
  %6 = add i8 %5, %1
  %7 = select i1 %4, i8 %6, i8 0
  ret i8 %7
}
=>
define i8 @_Z6squarejj(i8 %0, i8 %1) {
%2:
  %3 = usub_sat i8 %1, %0
  %4 = usub_sat i8 %3, 4
  ret i8 %4
}
Transformation seems to be correct!

Summary:
  1 correct transformations
  0 incorrect transformations
  0 Alive2 errors

[LebedevRI]

Interestingly, this is also a CSE failure, because we do succeed for

using size_type = unsigned;

size_type t0(size_type pos, size_type size) {
        size_type bytesRemaining = (pos < size) ? size - pos : 0;
    return ((bytesRemaining > 4)
                       ? (bytesRemaining - 4)
                       : 0);
}

https://godbolt.org/z/u_ZR2P

[nikic]

This looks pretty tricky, I don't know how to fix it. IR at the time of instcombine is

define dso_local i32 @_Z2t0jj(i32 %pos, i32 %size) local_unnamed_addr {
entry:
  %sub = sub i32 %size, %pos
  %0 = call i32 @llvm.usub.sat.i32(i32 %size, i32 %pos)
  %cmp1 = icmp ugt i32 %0, 4
  %sub4 = add i32 %sub, -4
  %cond7 = select i1 %cmp1, i32 %sub4, i32 0
  ret i32 %cond7
}

We basically need to show that %0 and %sub are actually the same, because %sub is only used in cases where the subtraction is non-saturating.

If the addition of usub.sat caused a perf regression here, we may want to add a one-use check to prevent having both the sub and the usub.sat.