Improve vectorization with conditionals around sin/cos

Recognize a loop with half-size predicate conditional and body
computation using sin or cosine intrinsics.

This change provides enhanced capabilities for such loops.
Previously we tried to prevent vectorization due to the nature of
sin/cos handling, and now these loops correctly vectorize.

Author: gklimowicz

tools/flang2/flang2exe/cgmain.cpp

@@ -7473,10 +7473,21 @@ gen_llvm_vsincos_call(int ilix)
   char sincosName[36]; /* make_math_name buffer is 32 */
   int vecLen = vecTy->sub_elements;
   int opndCount = ili_get_vect_arg_count(ilix);
+  int mask_arg_ili = ILI_OPND(ilix, opndCount - 1);
   bool hasMask = false;
+
   /* Mask operand is always the one before the last operand */
-  if (ILI_OPC(ILI_OPND(ilix, opndCount - 1)) != IL_NULL) {
-    opnd->next = gen_llvm_expr(ILI_OPND(ilix, opndCount - 1), vecTy);
+  if (ILI_OPC(mask_arg_ili) != IL_NULL) {
+    if((ILI_OPC(mask_arg_ili) == IL_VPERMUTE ||
+       (ILI_OPC(mask_arg_ili) == IL_VNOT &&
+        ILI_OPC(ILI_OPND(mask_arg_ili, 1)) == IL_VPERMUTE)) &&
+        internal_masked_intrinsic) {
+      internal_masked_intrinsic = false;
+      opnd->next = gen_llvm_expr(mask_arg_ili, vecTy);
+      internal_masked_intrinsic = true;
+    }
+    else
+    opnd->next = gen_llvm_expr(mask_arg_ili, vecTy);
     hasMask = true;
   }
   llmk_math_name(sincosName, MTH_sincos, vecLen, hasMask, dtypeName);