2012/7/17 Stefan Dösinger stefandoesinger@gmx.at:
Am Montag, 16. Juli 2012, 20:39:23 schrieb Matteo Bruni:
- HLSL_IR_BINOP_MUL,
...
- HLSL_IR_BINOP_DOT,
HLSL has 3 somewhat related multiplication operations: "*", "mul" and "dot". Dot is fairly straightforward, but "*" and "mul" have subtle differences. I recommend to write tests to find out the exact details, but as far as I remember "float4 * float4" gives you a float4 with a component-wise multiplication, whereas mul(float, float) returns a scalar like the dot product. "float4x4 * float4" is an error, but mul(float4x4, float4) performs a matrix-vector multiplication and returns a float4.
Indeed this is not very clear, but with HLSL_IR_BINOP_MUL I mean the translation of the "*" operator, while I'm planning to handle the mul() intrinsic with some code translating that into a bunch of MULs, ADDs and DOTs depending on the specific data types. mul() is apparently documented quite okay in MSDN but I know I'll have to test all of that anyway...
+static struct hlsl_type *expr_common_type(struct hlsl_type *t1, struct hlsl_type *t2,
struct source_location *loc)This type inference system might be a bit too limited, you have to take the operation into account. e.g. a dot(float3, float4) returns a scalar and crs(float4, float4) returns a float3.
Yeah, you're right, I'll need to take into account the operation in some cases, I consider this as a first "approximation". However I was planning to handle the "intrinsic function parameters / return value" case you mention in a different way, i.e. by explicitly generating all the intrinsic functions prototypes, with all the data type combinations allowed, and letting the (currently missing) "overloaded functions type checker/dispatcher" deal with the data types.
Not that I have written that code yet, so it may reveal to be a bad idea in the end. Stay tuned...