[llvm] [RISCV] Generaize reduction tree matching to all integer reductions (PR #68014)

[Alex Bradbury via llvm-commits]

asb wrote:

I reverted this in 824251c9b349d859a9169196cd9533c619a715ce because I was seeing test failures due to the unguarded call to `RISCVSubtarget::getELen` in `combineBinOpOfExtractToReduceTree`, which asserts when V isn't enabled.

I've found you actually can reproduce the crash by just running the fixed-vectors-reduction-formation.ll test without `+v`. `fixed-vectors-reduction-formation.ll` doesn't crash.

Alternatively, here is a bugpoint reduced example derived from 20050604-1.c in the GCC torture suite:
```ll
%union.anon.2.3.19.20 = type { <4 x i16> }

@u = external global %union.anon.2.3.19.20, align 8

declare void @foo() 

define void @main() {
entry:
  %u.promoted.i = load <4 x i16>, ptr @u, align 8
  %add.1.i = add <4 x i16> %u.promoted.i, <i16 24, i16 0, i16 0, i16 0>
  %0 = extractelement <4 x i16> %add.1.i, i64 2
  %tobool = icmp ne i16 %0, 0
  %or.cond = select i1 false, i1 true, i1 %tobool
  %1 = extractelement <4 x i16> %add.1.i, i64 3
  %tobool5 = icmp ne i16 %1, 0
  %or.cond6 = select i1 %or.cond, i1 true, i1 %tobool5
  br i1 %or.cond6, label %if.then, label %if.end

if.then:                                          ; preds = %entry
  unreachable

if.end:                                           ; preds = %entry
  ret void
}

declare void @abort()
```

One obvious fix would be to return immediately from `combineBinOpOfExtractToReduceTree` if `!Subtarget.hasVInstructions()`.

https://github.com/llvm/llvm-project/pull/68014