[llvm] VectorCombine: teach foldExtractedCmps about samesign (PR #122883)

[Florian Hahn via llvm-commits]

fhahn wrote:

It looks like this introduced a new crash:


```
target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-n32:64-S128-Fn32"
target triple = "arm64-apple-macosx15.0.0"

define fastcc float @test() {
entry:
  br label %if.end407

if.then374:                                       ; No predecessors!
  br label %if.end407

if.end407:                                        ; preds = %if.then374, %entry
  %overflow.1 = phi double [ 0.000000e+00, %entry ], [ 0.000000e+00, %if.then374 ]
  %0 = load i32, ptr null, align 4
  %cmp442 = fcmp ogt double %overflow.1, 0.000000e+00
  %cmp446 = icmp samesign ugt i32 %0, 0
  %or.cond238 = and i1 %cmp442, %cmp446
  ret float 0.000000e+00
}
```


```
$ bin/opt -p vector-combine reduced.ll
Unknown predicate!
UNREACHABLE executed at ../../../llvm/lib/IR/Instructions.cpp:3852!
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
```


https://github.com/llvm/llvm-project/pull/122883