Fix VPERMUTE for input and returned vectors having different sizes

Example: VPERMUTE <2 x float>, <2 x float>, (0, 1, 2, 3), <4
x float>  <- return dtype This example shuffles the contents
from both <2 x float> vectors into a single <4 x float> vector.
Previously, in this example, the <2 x float>s were bitcasted
into <4 x float>s, which triggers an internal error. This
change prevents that bitcast from happening.

Author: gklimowicz

tools/flang2/flang2exe/cgmain.cpp

@@ -9146,7 +9146,7 @@ gen_llvm_expr(int ilix, LL_Type *expected_type)
     break;
   case IL_VPERMUTE: {
     OPERAND *op1;
-    LL_Type *vect_lltype, *int_type, *select_type;
+    LL_Type *vect_lltype, *int_type, *select_type, *op_lltype;
     DTYPE vect_dtype = ili_get_vect_dtype(ilix);
     int mask_ili, num_elem;
 
@@ -9159,6 +9159,12 @@ gen_llvm_expr(int ilix, LL_Type *expected_type)
     lhs_ili = ILI_OPND(ilix, 1);
     rhs_ili = ILI_OPND(ilix, 2);
     mask_ili = ILI_OPND(ilix, 3);
+    op_lltype = make_lltype_from_dtype(ili_get_vect_dtype(lhs_ili));
+
+    if ((*vect_lltype->sub_types)->data_type != (*op_lltype->sub_types)->data_type) {
+      assert(0, "VPERMUTE: result and operand dtypes must have matching base type.", 0, ERR_Severe);
+    }
+
     if (expected_type && ILI_OPC(rhs_ili) == IL_NULL &&
         ILI_OPC(lhs_ili) == IL_VCMP && !internal_masked_intrinsic) {
       num_elem = expected_type->sub_elements;
@@ -9167,12 +9173,16 @@ gen_llvm_expr(int ilix, LL_Type *expected_type)
       vect_lltype = select_type;
     } else
       select_type = NULL;
-    op1 = gen_llvm_expr(lhs_ili, vect_lltype);
+    op1 = gen_llvm_expr(lhs_ili, op_lltype);
     if (ILI_OPC(rhs_ili) == IL_NULL) /* a don't care, generate an undef */
       op1->next = make_undef_op(op1->ll_type);
     else
-      op1->next = gen_llvm_expr(rhs_ili, vect_lltype);
+      op1->next = gen_llvm_expr(rhs_ili, op_lltype);
     op1->next->next = gen_llvm_expr(mask_ili, 0);
+    if (vect_lltype->sub_elements != op1->next->next->ll_type->sub_elements) {
+      assert(0, "VPERMUTE: result and mask must have the same number of elements.", 
+             vect_lltype->sub_elements, ERR_Severe);
+    }
     operand = ad_csed_instr(I_SHUFFVEC, ilix, vect_lltype, op1,
                             InstrListFlagsNull, true);
     /* This next case is where the VPERMUTE is used to expand a half-size

tools/flang2/utils/ilitp/aarch64/ilitp.n

@@ -5239,11 +5239,13 @@ Vector FMA for LLVM intrinsic - -lnk1*lnk2-lnk3, with stc the dtype
 .AT arth null lnk cse vect
 .CG notCG
 .IL VPERMUTE lnk lnk lnk stc
-Shuffle contents of vector registers. lnk1 and link2 can be same vector,
-with result in link2. stc4 is the dtype, link3 is a vector constant
+Shuffle contents of vector registers. lnk1 and lnk2 can be the same vector
+or lnk2 can be null. lnk1 dtype is used as dtype for both lnk1 and lnk2,
+unless lnk2 is null. stc is the result dtype, lnk3 is a vector constant
 representing a mask where each field represents which L-to-R element of
-concatenated <link1,link2> vector is to be placed in corresponding result
-field.
+concatenated <lnk1,lnk2> vector is to be placed in corresponding result
+field. lnk3 size must match the size of the result vector, but can be
+different than lnk1 and lnk2's size.
 .AT other null lnk vect
 .CG notCG
 .IL VBLEND lnk lnk lnk stc

tools/flang2/utils/ilitp/ppc64le/ilitp.n

@@ -5219,11 +5219,13 @@ Vector FMA for LLVM intrinsic - -lnk1*lnk2-lnk3, with stc the dtype
 .AT arth null lnk cse vect
 .CG notCG
 .IL VPERMUTE lnk lnk lnk stc
-Shuffle contents of vector registers. lnk1 and link2 can be same vector,
-with result in link2. stc4 is the dtype, link3 is a vector constant
+Shuffle contents of vector registers. lnk1 and lnk2 can be the same vector
+or lnk2 can be null. lnk1 dtype is used as dtype for both lnk1 and lnk2,
+unless lnk2 is null. stc is the result dtype, lnk3 is a vector constant
 representing a mask where each field represents which L-to-R element of
-concatenated <link1,link2> vector is to be placed in corresponding result
-field.
+concatenated <lnk1,lnk2> vector is to be placed in corresponding result
+field. lnk3 size must match the size of the result vector, but can be
+different than lnk1 and lnk2's size.
 .AT other null lnk vect
 .CG notCG
 .IL VBLEND lnk lnk lnk stc

tools/flang2/utils/ilitp/x86_64/ilitp.n

@@ -6239,11 +6239,13 @@ Vector FMA for LLVM intrinsic - -lnk1*lnk2-lnk3, with stc the dtype
 .AT arth null lnk cse vect
 .CG notCG
 .IL VPERMUTE lnk lnk lnk stc
-Shuffle contents of vector registers. lnk1 and lnk2 can be same vector,
-with result in lnk2. stc is the dtype, lnk3 is a vector constant
+Shuffle contents of vector registers. lnk1 and lnk2 can be the same vector
+or lnk2 can be null. lnk1 dtype is used as dtype for both lnk1 and lnk2, 
+unless lnk2 is null. stc is the result dtype, lnk3 is a vector constant
 representing a mask where each field represents which L-to-R element of
 concatenated <lnk1,lnk2> vector is to be placed in corresponding result
-field.
+field. lnk3 size must match the size of the result vector, but can be 
+different than lnk1 and lnk2's size.
 .AT other null lnk vect
 .CG notCG
 .IL VBLEND lnk lnk lnk stc