https://github.com/krzysz00 updated https://github.com/llvm/llvm-project/pull/231302
>From e10e2b68139765cdaf1918ef7db0df39ce8c0ba3 Mon Sep 17 00:00:00 2001 From: Krzysztof Drewniak <[email protected]> Date: Fri, 9 Oct 2026 16:29:02 +0000 Subject: [PATCH] [AMDGPU][docs] Document the <4 x i32> buffer intrinsic deprecation Flag the deprecation of <4 x i32> buffer intrinsics in the release notes, flagging the migration path and the s.buffer.load subtleties. AI disclosure: Claude did draft some of this, but I deleted basically all of it and the words are mine Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> --- llvm/docs/AMDGPUUsage.rst | 25 ++++++++++++++++++++++++- llvm/docs/ReleaseNotes.md | 31 +++++++++++++++++++++++++++++++ 2 files changed, 55 insertions(+), 1 deletion(-) diff --git a/llvm/docs/AMDGPUUsage.rst b/llvm/docs/AMDGPUUsage.rst index 34dc835438378..27f002e4a40ec 100644 --- a/llvm/docs/AMDGPUUsage.rst +++ b/llvm/docs/AMDGPUUsage.rst @@ -1355,6 +1355,14 @@ supported for the ``amdgcn`` target. to make incorrect transformations (such as promotion to ``s_buffer_load``, which requires such componentwise alignment). + The buffer intrinsics that take their resource as a ``<4 x i32>`` instead of + a ``ptr addrspace(8)`` are deprecated and will be removed in a future LLVM + release. Add ``ptr`` after the + ``raw``/``struct`` component of the intrinsic name for the new version. + However, ``llvm.amdgcn.s.buffer.load`` becomes + ``llvm.amdgcn.ptr.s.buffer.load`` and needs to be annotated with ``!invariant.load`` + to preserve the old merging behavior. This migration was already done in clang. + **Buffer Strided Pointer** The buffer index pointer is an experimental address space. It represents a 128-bit buffer descriptor and a 32-bit offset, like the **Buffer Fat @@ -2231,6 +2239,17 @@ The AMDGPU backend implements the following LLVM IR intrinsics. * Local pointer to the LDS barrier data. + llvm.amdgcn.s.buffer.load Deprecated. Takes its buffer resource as a ``<4 x i32>``. Use + ``llvm.amdgcn.ptr.s.buffer.load``, which takes a ``ptr addrspace(8)`` + produced by ``llvm.amdgcn.make.buffer.rsrc``. + The implicit assumptions encoded by the fact that this intrinsic was tagged as not reading memory are replaced by needing to tag ``llvm.amdgcn.ptr.s.buffer.load`` with ``!invariant.load`` metadata. + + llvm.amdgcn.{raw|struct}.*buffer.* Deprecated. The ``buffer``, ``tbuffer`` and ``atomic.buffer.load`` + intrinsics that take their buffer resource as a ``<4 x i32>`` are + replaced by the variants with ``ptr`` inserted after the ``raw`` or + ``struct`` component of the name, which take a ``ptr addrspace(8)`` + produced by ``llvm.amdgcn.make.buffer.rsrc``. Codegen warns for their use. + ============================================== ========================================================== .. TODO:: @@ -2406,10 +2425,14 @@ registers. See :ref:`amdgpu-dma-operations` for full documentation. llvm.amdgcn.load[.async].to.lds llvm.amdgcn.global.load[.async].lds - llvm.amdgcn.{raw|struct}[.ptr].buffer.load[.async].lds + llvm.amdgcn.{raw|struct}.ptr.buffer.load[.async].lds llvm.amdgcn.{global|cluster}.load.async.to.lds.b{8,32,64,128} llvm.amdgcn.global.store.async.from.lds.b{8,32,64,128} +The ``llvm.amdgcn.{raw|struct}.buffer.load[.async].lds`` spellings, which take +their buffer resource as a ``<4 x i32>`` rather than a ``ptr addrspace(8)``, +are deprecated. + Tensor Intrinsics ~~~~~~~~~~~~~~~~~ diff --git a/llvm/docs/ReleaseNotes.md b/llvm/docs/ReleaseNotes.md index cec21851f0d76..c7366a6af8230 100644 --- a/llvm/docs/ReleaseNotes.md +++ b/llvm/docs/ReleaseNotes.md @@ -284,6 +284,37 @@ Makes programs 10x faster by doing Special New Thing. * `llvm.amdgcn.icmp` * `llvm.amdgcn.fcmp` +* The AMDGPU buffer intrinsics that take their buffer resource (rsrc/V#) + argument as a `<4 x i32>` are deprecated and will be removed in a future + LLVM release. Codegen now warns for their use once per module. + + They all have `.ptr.` equivalents that take a `ptr addrspace(8)`, produced by + `llvm.amdgcn.make.buffer.rsrc`, loaded directly from memory, or, if + necessary, converted from an existing descriptor with + `bitcast <4 x i32> %rsrc to i128` followed by `inttoptr i128 %i to ptr + addrspace(8)`. To migrate, insert `ptr` after the `raw`/`struct` component of + the intrinsic name; `llvm.amdgcn.s.buffer.load` is the sole exception and + becomes `llvm.amdgcn.ptr.s.buffer.load`. + + ```llvm + ; Before + %v = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %off, i32 0, i32 0) + + ; After + %p = call ptr addrspace(8) @llvm.amdgcn.make.buffer.rsrc.p8.p1.i32(ptr addrspace(1) %base, i16 0, i32 %numrecords, i32 %flags) + %v = call float @llvm.amdgcn.raw.ptr.buffer.load.f32(ptr addrspace(8) %p, i32 %off, i32 0, i32 0) + ``` + + This migration enables future flexibility in resource size and usage and + allows for more accurate analysis of memory usage. + + Note that when emitting `llvm.amdgcn.ptr.s.buffer.load`, you must tag the intrinsic with `!invariant.load` metadata in order to recover the merging and loop hoisting behavior that `llvm.amdgcn.s.buffer.load` had because it lied about 'not reading memory'. Explicitly tagging these buffer loads as invariant will teach LLVM's general infrastructure about the property that users of the old ``s.buffer.load`` intrinsic were already asserting. + + While we do intend to autoupgrade the old intrinsics when they are removed, this upgrade will have to use `inttoptr` to convert buffer resources, losing pointer provenance. Therefore, we strongly encourage users to migrate to the new forms (which have been available for years) before this removal. + + The warning can be suppressed with the + `amdgpu-warn-deprecated-buffer-intrinsics=0` option. + ### Changes to the ARM Backend * Using the hard-float procedure call standard without floating-point registers _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
