On Mon, 15 Dec 2025, Haochen Jiang wrote: > It is yet another year end and time to add the changes for Intel x86_64 > backend for GCC 16. Nothing special for that part.
Thanks for taking care of that (and apologies for the delay in responding). > In GCC 16 timeframe, there are quite a few changes for ISA enabled on > compiler switches backported to previous release branch, making the switch > behavior has minor difference between minor versions. What we have done > previously for SRF/GRR is to sliently change the documentation. This is not > a good way to show the diff. Thus, I spent some time to explicitly mention > the minor version changes correspondingly in the doc where it is introduced. This is a great approach for users! > + <li>Since GCC 12.5, CLDEMOTE is not enabled through the compiler switch > + <code>-march=alderlake</code>. Maybe add "any longer" at the end? > + Based on ISA extensions enabled on Alder Lake, the switch intially > further > + enables the AVX-IFMA, AVX-NE-CONVERT, AVX-VNNI-INT8 and CMPccXADD ISA > + extensions. Since GCC 13.2, ENQCMD and UINTR are further enabled. Typo, "initially". Though how about saying "this switch in addition enables ... Since GCC 13.2, ENQCMD and UINTR are enabled as well"? > + Based on Sierra Forest, the switch initially further enables the RAO-INT > + ISA extensions. Since GCC 13.2, ENQCMD and UINTR are further enabled. > + Since GCC 13.3, RAO-INT is not enabled. I suggest to append "any longer" or "any more". > + <li>Since GCC 13.5, CLDEMOTE is not enabled through the compiler > + switch <code>-march=alderlake</code>. Similar. > + <li>Since GCC 16, AMX-TRANSPOSE and USER_MSR are not enabled through > + the compiler switch <code>-march=diamondrapids</code>. CLDEMOTE is not > + enabled through the compiler switches <code>-march=alderlake</code>, > + <code>-march=arrowlake</code>, <code>-march=arrowlake-s</code>, > + <code>-march=gracemont</code>, <code>-march=lunarlake</code>, > + <code>-march=meteorlake</code>, <code>-march=pantherlake</code> and > + <code>-march=raptorlake</code>. KL and WIDEKL are not enabled through > + the compiler switches <code>-march=clearwaterforest</code> and > + <code>-march=pantherlake</code>. PREFETCHI is not enabled through the > + compiler switch <code>-march=pantherlake</code>. > + </li> Maybe "Starting with GCC 16"? > + <li><code>-mavx10.1-256</code>, <code>-mavx10.1-512</code>, and > + <code>-mevex512</code> were removed together with the warning for the > + behavior change on <code>-mavx10.1</code>. <code>-mavx10.1</code> has > + enabled AVX10.1 intrinsics with 512-bit vector support since GCC 15. Just to make sure: the warning was removed? If so, all good. > + </li> > + <li>Support for AMX-TRANSPOSE were removed in GCC 16. GCC will no longer > accept > + <code>-mamx-transpose</code>, > + </li> "was removed" Wow, that's quite a bit. Thanks for your careful work tracking all these! (Patch is okay.) Gerald
