@@ -1016,6 +1016,15 @@ supported for the ``amdgcn`` target.
10161016 `ptr addrspace(7)` directly, which produces a buffer fat pointer with an initial
10171017 offset of 0 and prevents the address space cast from being rewritten away.
10181018
1019+ The ``align`` attribute on operations from buffer fat pointers is deemed to apply
1020+ to all componenents of the pointer - that is, an ``align 4`` load is expected to
1021+ both have the offset be a multiple of 4 and to have a base pointer with an
1022+ alignment of 4.
1023+
1024+ This componentwise definition of alignment is needed to allow for promotion of
1025+ aligned loads to ``s_buffer_load``, which requires that both the base pointer and
1026+ offset be appropriately aligned.
1027+
10191028**Buffer Resource**
10201029 The buffer resource pointer, in address space 8, is the newer form
10211030 for representing buffer descriptors in AMDGPU IR, replacing their
@@ -1039,6 +1048,25 @@ supported for the ``amdgcn`` target.
10391048 (bits `127:96`). The specific interpretation of these fields varies by the
10401049 target architecture and is detailed in the ISA descriptions.
10411050
1051+ When buffer resources are passed to buffer intrinsics such as
1052+ ``llvm.amdgcn.raw.ptr.buffer.load`` or
1053+ ``llvm.amdgcn.struct.ptr.buffer.store``, the ``align`` attribute on the
1054+ pointer is assumed to apply to both the offset and the base pointer value.
1055+ That is, ``align 8`` means that both the base address within the ``ptr
1056+ addrspace(8)`` and the ``offset`` argument have their three lowest bits set
1057+ to 0. If the stride of the resource is nonzero, the stride must be a multiple
1058+ of the given alignment.
1059+
1060+ In other words, the ``align`` attribute specifies the alignment of the effective
1061+ address being loaded from/stored to *and* acts as a guarantee that this is
1062+ not achieved from adding lower-alignment parts (as hardware may not always
1063+ allow for such an addition). For example, if a buffer resource has the base
1064+ address ``0xfffe`` and is accessed with a ``raw.ptr.buffer.load`` with an offset
1065+ of ``2``, the load must **not** be marked ``align 4`` (even though the
1066+ effective adddress ``0x10000`` is so aligned) as this would permit the compiler
1067+ to make incorrect transformations (such as promotion to ``s_buffer_load``,
1068+ which requires such componentwise alignment).
1069+
10421070**Buffer Strided Pointer**
10431071 The buffer index pointer is an experimental address space. It represents
10441072 a 128-bit buffer descriptor and a 32-bit offset, like the **Buffer Fat
@@ -1057,6 +1085,12 @@ supported for the ``amdgcn`` target.
10571085 index and offset values are both 0. This prevents the address space cast from
10581086 being rewritten away.
10591087
1088+ As with buffer fat pointers, alignment of a buffer strided pointer applies to
1089+ both the base pointer address and the offset. In addition, the alignment also
1090+ constrains the stride of the pointer. That is, if you do an ``align 4`` load from
1091+ a buffer strided pointer, this means that the base pointer is ``align(4)``, that
1092+ the offset is a multiple of 4 bytes, and that the stride is a multiple of 4.
1093+
10601094**Streamout Registers**
10611095 Dedicated registers used by the GS NGG Streamout Instructions. The register
10621096 file is modelled as a memory in a distinct address space because it is indexed
0 commit comments