diff --git a/libs/vkd3d-shader/hlsl.y b/libs/vkd3d-shader/hlsl.y
index 5ddf4b6c..7dbb051f 100644
--- a/libs/vkd3d-shader/hlsl.y
+++ b/libs/vkd3d-shader/hlsl.y
@@ -2099,11 +2099,12 @@ static bool invert_swizzle_matrix(const struct hlsl_matrix_swizzle *swizzle,
 }
 
 static bool add_assignment(struct hlsl_ctx *ctx, struct hlsl_block *block, struct hlsl_ir_node *lhs,
-        enum parse_assign_op assign_op, struct hlsl_ir_node *rhs)
+        enum parse_assign_op assign_op, struct hlsl_ir_node *rhs, bool is_function_out_arg)
 {
     struct hlsl_type *lhs_type = lhs->data_type;
     unsigned int writemask = 0, width = 0;
     bool matrix_writemask = false;
+    bool first_cast = true;
 
     if (lhs->data_type->class == HLSL_CLASS_ERROR || rhs->data_type->class == HLSL_CLASS_ERROR)
     {
@@ -2149,7 +2150,7 @@ static bool add_assignment(struct hlsl_ctx *ctx, struct hlsl_block *block, struc
                 hlsl_fixme(ctx, &cast->loc, "Size change on the LHS.");
                 return false;
             }
-            if (hlsl_version_ge(ctx, 4, 0))
+            if (hlsl_version_ge(ctx, 4, 0) && (!is_function_out_arg || !first_cast))
             {
                 hlsl_error(ctx, &cast->loc, VKD3D_SHADER_ERROR_HLSL_INVALID_LVALUE,
                         "Base type casts are not allowed on the LHS for profiles >= 4.");
@@ -2159,6 +2160,8 @@ static bool add_assignment(struct hlsl_ctx *ctx, struct hlsl_block *block, struc
             lhs_type = lhs->data_type;
             if (lhs_type->class == HLSL_CLASS_VECTOR || (lhs_type->class == HLSL_CLASS_MATRIX && matrix_writemask))
                 lhs_type = hlsl_get_vector_type(ctx, lhs->data_type->e.numeric.type, width);
+
+            first_cast = false;
         }
         else if (lhs->type == HLSL_IR_SWIZZLE)
         {
@@ -2370,7 +2373,7 @@ static bool add_increment(struct hlsl_ctx *ctx, struct hlsl_block *block, bool d
         return false;
     hlsl_block_add_instr(block, one);
 
-    if (!add_assignment(ctx, block, lhs, decrement ? ASSIGN_OP_SUB : ASSIGN_OP_ADD, one))
+    if (!add_assignment(ctx, block, lhs, decrement ? ASSIGN_OP_SUB : ASSIGN_OP_ADD, one, false))
         return false;
 
     if (post)
@@ -3272,7 +3275,7 @@ static struct hlsl_ir_node *add_user_call(struct hlsl_ctx *ctx,
                 return NULL;
             hlsl_block_add_instr(args->instrs, &load->node);
 
-            if (!add_assignment(ctx, args->instrs, arg, ASSIGN_OP_ASSIGN, &load->node))
+            if (!add_assignment(ctx, args->instrs, arg, ASSIGN_OP_ASSIGN, &load->node, true))
                 return NULL;
         }
     }
@@ -5457,7 +5460,7 @@ static bool intrinsic_interlocked(struct hlsl_ctx *ctx, enum hlsl_interlocked_op
             hlsl_error(ctx, &orig_val->loc, VKD3D_SHADER_ERROR_HLSL_MODIFIES_CONST,
                     "Output argument to '%s' is const.", name);
 
-        if (!add_assignment(ctx, params->instrs, orig_val, ASSIGN_OP_ASSIGN, interlocked))
+        if (!add_assignment(ctx, params->instrs, orig_val, ASSIGN_OP_ASSIGN, interlocked, true))
             return false;
     }
 
@@ -6394,7 +6397,7 @@ static bool add_assignment_from_component(struct hlsl_ctx *ctx, struct hlsl_bloc
     if (!(load = hlsl_add_load_component(ctx, instrs, src, component, loc)))
         return false;
 
-    if (!add_assignment(ctx, instrs, dest, ASSIGN_OP_ASSIGN, load))
+    if (!add_assignment(ctx, instrs, dest, ASSIGN_OP_ASSIGN, load, false))
         return false;
 
     return true;
@@ -6550,7 +6553,7 @@ static bool add_getdimensions_method_call(struct hlsl_ctx *ctx, struct hlsl_bloc
             return false;
         hlsl_block_add_instr(block, sample_info);
 
-        if (!add_assignment(ctx, block, args[ARG_SAMPLE_COUNT], ASSIGN_OP_ASSIGN, sample_info))
+        if (!add_assignment(ctx, block, args[ARG_SAMPLE_COUNT], ASSIGN_OP_ASSIGN, sample_info, false))
             return false;
     }
 
@@ -9839,7 +9842,7 @@ assignment_expr:
                 hlsl_error(ctx, &@2, VKD3D_SHADER_ERROR_HLSL_MODIFIES_CONST, "Statement modifies a const expression.");
             hlsl_block_add_block($3, $1);
             destroy_block($1);
-            if (!add_assignment(ctx, $3, lhs, $2, rhs))
+            if (!add_assignment(ctx, $3, lhs, $2, rhs, false))
                 YYABORT;
             $$ = $3;
         }
diff --git a/tests/hlsl/function-cast.shader_test b/tests/hlsl/function-cast.shader_test
index e2593444..99de5a92 100644
--- a/tests/hlsl/function-cast.shader_test
+++ b/tests/hlsl/function-cast.shader_test
@@ -25,7 +25,7 @@ probe (0, 0) rgba (-1.0, -1.0, 2.0, 4.0)
 % In SM 6 a cast seems to implicitly promote the type to const,
 % so it fails to match the parameter of func().
 
-[pixel shader fail(sm>=6) todo(sm>=4)]
+[pixel shader fail(sm>=6)]
 
 uniform float4 f;
 
@@ -43,12 +43,12 @@ float4 main() : sv_target
 
 [test]
 uniform 0 float4 -1.9 -1.0 2.9 4.0
-todo(sm>=4) draw quad
+draw quad
 probe (0, 0) rgba (-1.0, -1.0, 2.0, 4.0)
 
 % As above, but declare "x" as float4 and cast it to int4.
 
-[pixel shader fail(sm>=6) todo(sm>=4)]
+[pixel shader fail(sm>=6)]
 
 uniform float4 f;
 
@@ -66,7 +66,7 @@ float4 main() : sv_target
 
 [test]
 uniform 0 float4 -1.9 -1.0 2.9 4.0
-todo(sm>=4) draw quad
+draw quad
 probe (0, 0) rgba (-1.0, -1.0, 2.0, 4.0)
 
 
@@ -93,7 +93,7 @@ probe (0, 0) rgba (-1.0, 0.0, 1.0, -3000000.0) 4
 
 
 % An explicit cast gets applied right before assignment, as if it was on the lhs.
-[pixel shader fail(sm>=6) todo(sm>=4)]
+[pixel shader fail(sm>=6)]
 void fun(out float4 f)
 {
     f = float4(1.4, 2.6, 3.9, 4.3);
@@ -108,7 +108,7 @@ float4 main() : sv_target
 }
 
 [test]
-todo(sm>=4) draw quad
+draw quad
 probe (0, 0) rgba(1, 2, 3, 4)
 
 
diff --git a/tests/hlsl/trigonometry.shader_test b/tests/hlsl/trigonometry.shader_test
index b966a8aa..1bce3628 100644
--- a/tests/hlsl/trigonometry.shader_test
+++ b/tests/hlsl/trigonometry.shader_test
@@ -162,7 +162,7 @@ float4 main() : sv_target
     return float4(6789);
 }
 
-[pixel shader todo(sm>=4)]
+[pixel shader]
 float4 main() : sv_target
 {
     int sin_out, cos_out;
@@ -171,5 +171,5 @@ float4 main() : sv_target
 }
 
 [test]
-todo(sm>=4 & sm<6) draw quad
+todo(glsl | msl) draw quad
 probe (0, 0) rgba (0, 0, 0, 0);