[llvm] [SystemZ] Remove custom implementation of optimizeLoadInst [NFC] (PR #133123)

[Ulrich Weigand via llvm-commits]

uweigand wrote:

> > I _do_ wonder about multiple uses - if a loaded value is used in multiple instructions, and some of those instructions are changed to fold in the load, this will increase the total number of memory accesses.
> 
> The target specific foldMemoryOperandImpl is still in play here, and can chose to avoid cases with multiple operands if desired.
> 
> Edit: Note that the previous target specific optimizeLoadInst would only pass the first operand if multiple were potentially foldable. This is suspect correctness wise. The target specific foldMemoryOperandImpl hook will bail out with two or more ops, so this is converting a possible correctness issue into a missed optimization at worst.

I agree with the case of multiple operands into the same instruction.

I was concerned about the case of a single load whose value is used by multiple *different* instructions.

https://github.com/llvm/llvm-project/pull/133123