[llvm] [SLP]Improve final minbitwidth analysis attempt. (PR #87786)

[Arthur Eubanks via llvm-commits]

aeubanks wrote:

We have a test that started failing after this patch. I seem to have narrowed it down to the following diff after SLPVectorizer with and without this patch:

```
$ diff /tmp/good.ll /tmp/bad.ll 
992,1011c992,1009
<   %8 = zext <8 x i16> %0 to <8 x i32>
<   %9 = zext <8 x i16> %1 to <8 x i32>
<   %10 = sub nsw <8 x i32> %9, %8
<   %11 = add nsw <8 x i32> %10, <i32 3329, i32 3329, i32 3329, i32 3329, i32 3329, i32 3329, i32 3329, i32 3329>
<   %12 = insertelement <8 x i32> poison, i32 %zext795, i32 0
<   %13 = shufflevector <8 x i32> %12, <8 x i32> poison, <8 x i32> zeroinitializer
<   %14 = mul <8 x i32> %11, %13
<   %15 = zext <8 x i32> %14 to <8 x i64>
<   %16 = mul nuw nsw <8 x i64> %15, <i64 5039, i64 5039, i64 5039, i64 5039, i64 5039, i64 5039, i64 5039, i64 5039>
<   %17 = lshr <8 x i64> %16, <i64 24, i64 24, i64 24, i64 24, i64 24, i64 24, i64 24, i64 24>
<   %18 = trunc <8 x i64> %17 to <8 x i32>
<   %19 = mul <8 x i32> %18, <i32 62207, i32 62207, i32 62207, i32 62207, i32 62207, i32 62207, i32 62207, i32 62207>
<   %20 = add <8 x i32> %19, %14
<   %21 = trunc <8 x i32> %20 to <8 x i16>
<   %22 = add <8 x i16> %21, <i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329>
<   %23 = icmp slt <8 x i16> %22, zeroinitializer
<   %24 = select <8 x i1> %23, <8 x i16> %21, <8 x i16> zeroinitializer
<   %25 = call <8 x i16> @llvm.smax.v8i16(<8 x i16> %22, <8 x i16> zeroinitializer)
<   %26 = or <8 x i16> %24, %25
<   store <8 x i16> %26, ptr %getelementptr797, align 2
---
>   %8 = sub <8 x i16> %1, %0
>   %9 = add <8 x i16> %8, <i16 3329, i16 3329, i16 3329, i16 3329, i16 3329, i16 3329, i16 3329, i16 3329>
>   %10 = insertelement <8 x i32> poison, i32 %zext795, i32 0
>   %11 = shufflevector <8 x i32> %10, <8 x i32> poison, <8 x i32> zeroinitializer
>   %12 = trunc <8 x i32> %11 to <8 x i16>
>   %13 = mul <8 x i16> %9, %12
>   %14 = zext <8 x i16> %13 to <8 x i64>
>   %15 = mul nuw nsw <8 x i64> %14, <i64 5039, i64 5039, i64 5039, i64 5039, i64 5039, i64 5039, i64 5039, i64 5039>
>   %16 = lshr <8 x i64> %15, <i64 24, i64 24, i64 24, i64 24, i64 24, i64 24, i64 24, i64 24>
>   %17 = trunc <8 x i64> %16 to <8 x i16>
>   %18 = mul <8 x i16> %17, <i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329>
>   %19 = add <8 x i16> %18, %13
>   %20 = add <8 x i16> %19, <i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329, i16 -3329>
>   %21 = icmp slt <8 x i16> %20, zeroinitializer
>   %22 = select <8 x i1> %21, <8 x i16> %19, <8 x i16> zeroinitializer
>   %23 = call <8 x i16> @llvm.smax.v8i16(<8 x i16> %20, <8 x i16> zeroinitializer)
>   %24 = or <8 x i16> %22, %23
>   store <8 x i16> %24, ptr %getelementptr797, align 2
```

but I'm trying to understand what's going wrong

https://github.com/llvm/llvm-project/pull/87786