[clang] [llvm] [PowerPC] Add restriction for rldimi builtin (PR #85040)

[Chen Zheng via llvm-commits]

================
@@ -5093,9 +5094,33 @@ bool Sema::CheckPPCBuiltinFunctionCall(const TargetInfo &TI, unsigned BuiltinID,
   case PPC::BI__builtin_ppc_rlwnm:
     return SemaValueIsRunOfOnes(TheCall, 2);
   case PPC::BI__builtin_ppc_rlwimi:
-  case PPC::BI__builtin_ppc_rldimi:
     return SemaBuiltinConstantArg(TheCall, 2, Result) ||
            SemaValueIsRunOfOnes(TheCall, 3);
+  case PPC::BI__builtin_ppc_rldimi: {
+    llvm::APSInt SH;
+    if (SemaBuiltinConstantArg(TheCall, 2, SH) ||
+        SemaBuiltinConstantArg(TheCall, 3, Result))
+      return true;
+    if (SH > 63)
+      return Diag(TheCall->getBeginLoc(), diag::err_argument_invalid_range)
+             << toString(Result, 10) << 0 << 63 << TheCall->getSourceRange();
+    unsigned MB = 0, ML = 0;
+    if (Result.isShiftedMask(MB, ML)) {
+      MB = 64 - MB - ML;
+    } else if ((~Result).isShiftedMask(MB, ML)) {
+      MB = 64 - MB;
+      ML = 64 - ML;
+    } else {
+      return Diag(TheCall->getBeginLoc(),
+                  diag::err_argument_not_contiguous_bit_field)
+             << 3 << TheCall->getSourceRange();
+    }
+    if ((MB + ML - 1) % 64 != 63 - SH.getZExtValue())
----------------
chenzheng1030 wrote:

I am a little worried about this check. This adds more restriction on the __rldimi builtin than IBM XLC does. And what do you expect the mask looks like if the mask is wrapped? i.e., MB > 64 -SH.

How about we accept all continuous 1 masks, and in the backend while lowering the intrinsic if the mask matches the `rlwinm` instruction encoding requirement, we use only one rlwinm instructions. Otherwise we use two instructions, like:

```
__rldimi(a, b, SH, mask)
```
->
```
rldicl     rx,a,SH,0
rldimi    b,rx,0,mask
```

https://github.com/llvm/llvm-project/pull/85040