[llvm] [AggressiveInstCombine] Implement store merge optimization (PR #147540)

[Nikita Popov via llvm-commits]

nikic wrote:

@dtcxzyw We always run these in the order `instcombine,aggressive-instcombine`, not `aggressive-instcombine,instcombine`. If instcombine runs first, then it will produce
```llvm
define void @src(ptr %0, ptr %.sroa.0.010) {
  %2 = load i16, ptr %.sroa.0.010, align 2
  %3 = getelementptr inbounds nuw i8, ptr %.sroa.0.010, i64 2
  %4 = load i16, ptr %3, align 4
  %5 = getelementptr inbounds nuw i8, ptr %0, i64 2
  store i16 %2, ptr %5, align 2
  %6 = getelementptr inbounds nuw i8, ptr %0, i64 4
  store i16 %4, ptr %6, align 4
  ret void
}
```
and then aggressive-instcombine will not change the IR.

https://github.com/llvm/llvm-project/pull/147540