================
@@ -267,45 +267,38 @@ Examples:
@llvm.dx.handle.fromHeap.tdx.RawBuffer_v4f32_1_0(
i32 2, i1 false)
-Buffer Loads and Stores
------------------------
-
-*relevant types: Buffers*
-
-We need to treat buffer loads and stores from "dx.TypedBuffer" and
-"dx.RawBuffer" separately. For TypedBuffer, we have ``llvm.dx.typedBufferLoad``
-and ``llvm.dx.typedBufferStore``, which load and store 16-byte "rows" of data
-via a simple index. For RawBuffer, we have ``llvm.dx.rawBufferPtr``, which
-return a pointer that can be indexed, loaded, and stored to as needed.
-
-The typed load and store operations always operate on exactly 16 bytes of data,
-so there are only a few valid overloads. For types that are 32-bits or smaller,
-we operate on 4-element vectors, such as ``<4 x i32>``, ``<4 x float>``, or
-``<4 x half>``. Note that in 16-bit cases each 16-bit value occupies 32-bits of
-storage. For 64-bit types we operate on 2-element vectors - ``<2 x double>`` or
-``<2 x i64>``. When a type like `Buffer<float>` is used at the HLSL level, it
-is expected that this will operate on a single float in each 16 byte row - that
-is, a load would use the ``<4 x float>`` variant and then extract the first
-element.
-
-.. note:: In DXC, trying to operate on a ``Buffer<double4>`` crashes the
- compiler. We should probably just reject this in the frontend.
-
-The TypedBuffer intrinsics are lowered to the `bufferLoad`_ and `bufferStore`_
-operations, and the operations on the memory accessed by RawBufferPtr are
-lowered to `rawBufferLoad`_ and `rawBufferStore`_. Note that if we want to
-support DXIL versions prior to 1.2 we'll need to lower the RawBuffer loads and
-stores to the non-raw operations as well.
-
-.. note:: TODO: We need to account for `CheckAccessFullyMapped`_ here.
-
- In DXIL the load operations always return an ``i32`` status value, but this
- isn't very ergonomic when it isn't used. We can (1) bite the bullet and have
- the loads return `{%ret_type, %i32}` all the time, (2) create a variant or
- update the signature iff the status is used, or (3) hide this in a sideband
- channel somewhere. I'm leaning towards (2), but could probably be convinced
- that the ugliness of (1) is worth the simplicity.
-
+16-byte Loads, Samples, and Gathers
+-----------------------------------
+
+*relevant types: TypedBuffer, CBuffer, and Textures*
+
+TypedBuffer, CBuffer, and Texture loads, as well as samples and gathers, can
+return 1 to 4 elements from the given resource, to a maximum of 16 bytes of
+data. DXIL's modeling of this is influenced by DirectX and DXBC's history and
+it generally treats these operations as returning 4 32-bit values. For 16-bit
+elements the values are 16-bit values, and for 64-bit values the operations
+return 4 32-bit integers and combine them with further operations.
----------------
bogner wrote:
Updated to say that it emits code to construct a double
https://github.com/llvm/llvm-project/pull/104252
_______________________________________________
llvm-branch-commits mailing list
llvm-branch-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
