Cannot lower "icmp uge <4 x i32> zeroinitializer, ..." for ARMv7

[aaronpuchert]

This is reduced from a larger file:

define <4 x i32> @vcmpz(<4 x i32> %0) {
  %2 = icmp uge <4 x i32> zeroinitializer, %0
  %3 = sext <4 x i1> %2 to <4 x i32>
  ret <4 x i32> %3
}

Compilation with llc -O2 -march=arm --float-abi=hard -mattr=+armv7-a fails with:

LLVM ERROR: Cannot select: t17: v4i32 = ARMISD::VCMPZ t3, Constant:i32<2>
  t3: v4i32 = bitcast t2
    t2: v2f64,ch = CopyFromReg t0, Register:v2f64 %0
      t1: v2f64 = Register %0
  t16: i32 = Constant<2>

The larger file had a more complex tree, but I think it boils down to the VCMPZ. Other comparisons are not (necessarily) an issue, e.g. for the opposite icmp ult we get

        vceq.i32        q8, q0, #0
        vmvn    q0, q8
        bx      lr

[llvmbot]

@llvm/issue-subscribers-backend-arm

[davemgreen]

The backend is probably assuming that %2 = icmp uge <4 x i32> zeroinitializer, %0 has been simplified into %2 = icmp eq <4 x i32> %0, zeroinitializer, but that is not always reliably done.

There is a fix in https://reviews.llvm.org/D136447

[Vogtinator]

That fits.

This particular failure was caused by Mesa not doing optimization passes with LLVM 15: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19217

With the instcombine pass the selection failure is gone.

For completeness, the downstream bug report: http://bugzilla.opensuse.org/show_bug.cgi?id=1204267

[davemgreen]

I see. Makes sense.

FYI, You likely don't need instsimplify, instcombine should be a super-set of it. But from the look of "After:" in the PR I don't see instsimplify running between PromotePass and InstCombinePass anyway.

[aaronpuchert]

This particular failure was caused by Mesa not doing optimization passes with LLVM 15

Adding optimizations could make it disappear, but backends should be able to lower non-optimized code.