Avoid constructing an unnecessary InferCtxt.

Currently, `search_for_structural_match_violation` constructs an `infcx`
from a `tcx` and then only uses the `tcx` within the `infcx`. This is
wasteful because `infcx` is a big type.

This commit changes it to use the `tcx` directly. When compiling
`pest-2.1.3`, this changes the memcpy stats reported by DHAT for a `check full`
build from this:
```
433,008,916 bytes (100%, 99,787.93/Minstr) in 2,148,668 blocks (100%, 495.17/Minstr), avg size 201.52 bytes
```
to this:
```
101,422,347 bytes (99.98%, 25,243.59/Minstr) in 1,318,407 blocks (99.96%, 328.15/Minstr), avg size 76.93 bytes
```
This translates to a 4.3% reduction in instruction counts.

Author: nnethercote

compiler/rustc_trait_selection/src/traits/structural_match.rs

@@ -58,10 +58,7 @@ pub fn search_for_structural_match_violation<'tcx>(
     tcx: TyCtxt<'tcx>,
     ty: Ty<'tcx>,
 ) -> Option<NonStructuralMatchTy<'tcx>> {
-    // FIXME: we should instead pass in an `infcx` from the outside.
-    tcx.infer_ctxt().enter(|infcx| {
-        ty.visit_with(&mut Search { tcx: infcx.tcx, span, seen: FxHashSet::default() }).break_value()
-    })
+    ty.visit_with(&mut Search { tcx, span, seen: FxHashSet::default() }).break_value()
 }
 
 /// This method returns true if and only if `adt_ty` itself has been marked as