Rollup merge of #142285 - workingjubilee:dont-use-bad-assembly, r=nnethercote

tests: Do not run afoul of asm.validity.non-exhaustive in input-stats

This addresses one of the three powerpc64-unknown-linux-musl test failures in #142280

I was motivated to cover it myself because technically this is also compile-time UB if we compile a program that has `asm!` with x86-64-specific instructions on another platform. That'll only mean something if this is ever switched to build-pass, or if checking emits object code, but conveniently "nop" is valid assembly on all platforms anyone has implemented Rust codegen for. Even the weird ones LLVM doesn't support, like PA-RISC or Common Intermediate Language.

...except GPUs. Not sure about those.

r? `@nnethercote`

Author: fmease

tests/ui/stats/input-stats.rs

@@ -1,6 +1,7 @@
 //@ check-pass
 //@ compile-flags: -Zinput-stats
 //@ only-64bit
+//@ needs-asm-support
 // layout randomization affects the hir stat output
 //@ needs-deterministic-layouts
 //
@@ -49,5 +50,7 @@ fn main() {
         _ => {}
     }
 
-    unsafe { asm!("mov rdi, 1"); }
+    // NOTE(workingjubilee): do GPUs support NOPs? remove this cfg if they do
+    #[cfg(not(any(target_arch = "nvptx64", target_arch = "spirv", target_arch = "amdgpu")))]
+    unsafe { asm!("nop"); }
 }