[llvm] [AMDGPU] Add inreg support for SGPR arguments (PR #67182)

[Matt Arsenault via llvm-commits]

================
@@ -2570,8 +2569,7 @@ bool isArgPassedInSGPR(const CallBase *CB, unsigned ArgNo) {
     return CB->paramHasAttr(ArgNo, Attribute::InReg) ||
            CB->paramHasAttr(ArgNo, Attribute::ByVal);
   default:
-    // TODO: Should calls support inreg for SGPR inputs?
-    return false;
+    return CB->paramHasAttr(ArgNo, Attribute::InReg);
----------------
arsenm wrote:

No, I mean for calls. e.g. `call @foo(i32 inreg %outgoing0, <2 x double> inreg %outgoing1)`

https://github.com/llvm/llvm-project/pull/67182