[AArch64] AND optimization does not work if there is a freeze between AND and the cmp

[AZero13]

Here is a case from the real world where this happens:

define i1 @test_disjoint(i1 %0, i32 %1, i32 %2) {
  %3 = or disjoint i32 %2, 8388608
  %4 = shl nuw i32 1, %1
  %5 = and i32 %3, %4
  %6 = icmp eq i32 %5, 0
  %7 = select i1 %0, i1 true, i1 %6
  ret i1 %7
}

We get:

test_disjoint:                          // @test_disjoint
        mov     w8, #1                          // =0x1
        orr     w9, w2, #0x800000
        lsl     w8, w8, w1
        and     w8, w9, w8
        cmp     w8, #0
        cset    w8, eq
        orr     w8, w0, w8
        and     w0, w8, #0x1
        ret

There are many ways to solve this problem.

In the short term, I guess we can fix this by special casing it. However, I am not sure of if this will be safe every time.

In all honesty, I think that perhaps this optimization should not be done at all in the long term. Why? This seems more like a peephole optimizer job.

So, my solution to this problem, and yes this definitely will not catch all cases because this works on a different domain to the SelDAG, is #153608, but also we don't do the replace all users in GISel anyway, so this benefits that too.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: AZero13 (AZero13)

Here is a case from the real world where this happens: define i1 @test_disjoint(i1 %0, i32 %1, i32 %2) { %3 = or disjoint i32 %2, 8388608 %4 = shl nuw i32 1, %1 %5 = and i32 %3, %4 %6 = icmp eq i32 %5, 0 %7 = select i1 %0, i1 true, i1 %6 ret i1 %7 }

We get:

test_disjoint: // @test_disjoint
mov w8, #1 // =0x1
orr w9, w2, #0x800000
lsl w8, w8, w1
and w8, w9, w8
cmp w8, #0
cset w8, eq
orr w8, w0, w8
and w0, w8, #0x1
ret

There are many ways to solve this problem.

In the short term, I guess we can fix this by special casing it. However, I am not sure of if this will be safe every time.

In all honesty, I think that perhaps this optimization should not be done at all in the long term. Why? This seems more like a peephole optimizer job.

So, my solution to this problem, and yes this definitely will not catch all cases because this works on a different domain to the SelDAG, is #153608, but also we don't do the replace all users in GISel anyway, so this benefits that too.

[c-rhodes]

"AND optimization does not work" to me suggests it isn't happening when it should, do you mean it is happening like in https://godbolt.org/z/cvvrxKron, but it's not safe?

[AZero13]

"AND optimization does not work" to me suggests it isn't happening when it should, do you mean it is happening like in https://godbolt.org/z/cvvrxKron, but it's not safe?

It should be happening.

[c-rhodes]

"AND optimization does not work" to me suggests it isn't happening when it should, do you mean it is happening like in https://godbolt.org/z/cvvrxKron, but it's not safe?

It should be happening.

ok, perhaps I'm missing something, but there is no freeze in your example nor in the tests on your PR #153608?

[AZero13]

"AND optimization does not work" to me suggests it isn't happening when it should, do you mean it is happening like in https://godbolt.org/z/cvvrxKron, but it's not safe?

It should be happening.

ok, perhaps I'm missing something, but there is no freeze in your example nor in the tests on your PR #153608?

That's because it's produced by the disjoint.

It's not in the test itself.

And the test I did was a real world example of this.

[AZero13]

Fix: #153608