Rollup merge of #104121 - Lokathor:mir-opt-when-instruction-set-missing-on-callee, r=tmiasko

Refine `instruction_set` MIR inline rules

Previously an exact match of the `instruction_set` attribute was required for an MIR inline to be considered. This change checks for an exact match *only* if the callee sets an `instruction_set` in the first place. When the callee does not declare an instruction set then it is considered to be platform agnostic code and it's allowed to be inline'd into the caller.

cc ``@oli-obk``

[Edit] Zulip Context: https://rust-lang.zulipchat.com/#narrow/stream/189540-t-compiler.2Fwg-mir-opt/topic/What.20exactly.20does.20the.20MIR.20optimizer.20do.3F

1 files changed, 15 insertions, 1 deletions

diff --git a/compiler/rustc_mir_transform/src/inline.rs b/compiler/rustc_mir_transform/src/inline.rs
index d7dd5fc8528..9174f04887e 100644
--- a/compiler/rustc_mir_transform/src/inline.rs
+++ b/compiler/rustc_mir_transform/src/inline.rs
@@ -375,7 +375,12 @@ impl<'tcx> Inliner<'tcx> {
             return Err("incompatible sanitizer set");
         }
 
-        if callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set {
+        // Two functions are compatible if the callee has no attribute (meaning
+        // that it's codegen agnostic), or sets an attribute that is identical
+        // to this function's attribute.
+        if callee_attrs.instruction_set.is_some()
+            && callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set
+        {
             return Err("incompatible instruction set");
         }
 
@@ -453,6 +458,15 @@ impl<'tcx> Inliner<'tcx> {
                 if ty.needs_drop(tcx, self.param_env) && let Some(unwind) = unwind {
                         work_list.push(unwind);
                     }
+            } else if callee_attrs.instruction_set != self.codegen_fn_attrs.instruction_set
+                && matches!(term.kind, TerminatorKind::InlineAsm { .. })
+            {
+                // During the attribute checking stage we allow a callee with no
+                // instruction_set assigned to count as compatible with a function that does
+                // assign one. However, during this stage we require an exact match when any
+                // inline-asm is detected. LLVM will still possibly do an inline later on
+                // if the no-attribute function ends up with the same instruction set anyway.
+                return Err("Cannot move inline-asm across instruction sets");
             } else {
                 work_list.extend(term.successors())
             }