[clang] [CIR] Move records through memory in x86_64 call lowering (PR #231663)

[Adam Smith via cfe-commits]

================
@@ -88,7 +88,6 @@ cir.func @takes_dim3(%arg0: !rec_Dim3)
 }
 
 // CHECK:      cir.func{{.*}} @takes_dim3(%[[D0:.*]]: !s64i, %[[D1:.*]]: !s32i)
-// CHECK:        cir.alloca "coerce" {{.*}} : !cir.ptr<{{.*}}, target_address_space(5)>
----------------
adams381 wrote:

Yes, that alloca is gone.  We used to store the fields into one slot, load the coerced struct back out, store it into a second slot, and then load the record from there.  Now the fields go into a single slot and the body reads the record through a cast of it.  The test never checked that load/store pair, which is why nothing else changed.  I added checks for the cast and for there being no record load after it.

https://github.com/llvm/llvm-project/pull/231663