[llvm-branch-commits] [llvm] 57381c3 - AMDGPU: Add some release notes

[Tobias Hieta via llvm-branch-commits]

Author: Matt Arsenault
Date: 2023-07-29T08:39:50+02:00
New Revision: 57381c31a1765421c47fb5837e3d09470cd368a7

URL: https://github.com/llvm/llvm-project/commit/57381c31a1765421c47fb5837e3d09470cd368a7
DIFF: https://github.com/llvm/llvm-project/commit/57381c31a1765421c47fb5837e3d09470cd368a7.diff

LOG: AMDGPU: Add some release notes

Added: 
    

Modified: 
    llvm/docs/LangRef.rst
    llvm/docs/ReleaseNotes.rst

Removed: 
    


################################################################################
diff  --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst
index eeba36e4167323..701cd6e5dbb6f7 100644
--- a/llvm/docs/LangRef.rst
+++ b/llvm/docs/LangRef.rst
@@ -14732,6 +14732,8 @@ trapping or setting ``errno``.
 When specified with the fast-math-flag 'afn', the result may be approximated
 using a less accurate calculation.
 
+.. _int_ldexp:
+
 '``llvm.ldexp.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 

diff  --git a/llvm/docs/ReleaseNotes.rst b/llvm/docs/ReleaseNotes.rst
index a1ca57e9a7db72..d1fe34063ee57f 100644
--- a/llvm/docs/ReleaseNotes.rst
+++ b/llvm/docs/ReleaseNotes.rst
@@ -149,6 +149,15 @@ Changes to the AMDGPU Backend
   improves the interaction between AMDGPU buffer operations and the LLVM memory
   model, and so the non `.ptr` intrinsics are deprecated.
 
+* SGPR spilling is now performed to virtual VGPRs. This should avoid
+  some assorted register allocation failures.
+
+* Backend now performs range merging of "amdgpu-waves-per-eu" attribute based on
+  known callers.
+
+* Certain :ref:`atomicrmw <i_atomicrmw>` operations are now optimized by
+  performing a wave reduction if the access is uniform by default.
+
 * Removed ``llvm.amdgcn.atomic.inc`` and ``llvm.amdgcn.atomic.dec``
   intrinsics. :ref:`atomicrmw <i_atomicrmw>` should be used instead
   with ``uinc_wrap`` and ``udec_wrap``.
@@ -176,6 +185,9 @@ Changes to the AMDGPU Backend
 * `llvm.sqrt.f64` is now lowered correctly. Use `llvm.amdgcn.sqrt.f64`
   for raw instruction access.
 
+* Deprecate `llvm.amdgcn.ldexp` intrinsic. :ref:`llvm.ldexp <int_ldexp>`
+  should be used instead.
+
 Changes to the ARM Backend
 --------------------------