[llvm] [InstCombine] Tighten use constraint in factorization transforms (PR #102943)

[Yingwei Zheng via llvm-commits]

dtcxzyw wrote:

Regression:
```
; bin/opt -passes=instcombine test.ll -S
define i64 @test(i64 %23, i64 %24) {
entry:
   %.idx14.i = shl nsw i64 %23, 3
   %.idx.i = shl nsw i64 %24, 3
   %gepdiff.i = sub nsw i64 %.idx.i, %.idx14.i
   %25 = ashr exact i64 %gepdiff.i, 3
   call void @use(i64 %.idx14.i)
   ret i64 %25
}

declare void @use(i64)
```

It can be folded into:
```
define i64 @test(i64 %0, i64 %1) {
entry:
  %.idx14.i = shl nsw i64 %0, 3
  %2 = sub nsw i64 %1, %0
  call void @use(i64 %.idx14.i)
  ret i64 %2
}
```


https://github.com/llvm/llvm-project/pull/102943