[llvm] [CodeGen] Fix MachineInstr::isSafeToMove handling of inline asm. (PR #126807)

[Phoebe Wang via llvm-commits]

phoebewang wrote:

> Making inline asm "speculatable" by default seems like it has weird effects in all sorts of scenarios:

I think that's how GCC [documents](https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#Volatile-1) it and what's user expected https://stackoverflow.com/a/14449998
 
> * You can't guard an instruction which is only valid on some CPUs with a runtime check.  For example, if you use cpuid.h to check whether an instruction is legal, you can only use the instruction in a volatile asm.
> * You can't use an instruction which is only valid with certain operands, because speculation can change the operands to anything.  For example, cpuid.h is broken because it doesn't mark the asm volatile.

I don't get the point of the example. If the use of cpuid instruction may trap, why don't add volatile? Without it, the example shown in GCC document is already broken.
 
> It's hard enough to use inline asm correctly without adding more weird edge cases.
> 
> We're not even getting much benefit from speculating inline asm, currently; we only do it in a couple obscure optimizations in the backend. Middle-end LLVM IR analysis already treats inline asm as potentially trapping.

I don't see why we can't go along the opposite direction, i.e., optimize in the middle-end.

https://github.com/llvm/llvm-project/pull/126807