[Lldb-commits] [lldb] [lldb][AArch64] Add release notes and documentation for SME (PR #66767)

David Spickett via lldb-commits lldb-commits at lists.llvm.org
Fri Oct 6 06:40:35 PDT 2023


================
@@ -0,0 +1,190 @@
+Using LLDB On AArch64 Linux
+===========================
+
+This page explains the details of debugging certain AArch64 extensions using
+LLDB. If something is not mentioned here, it likely works as you would expect.
+
+This is not a replacement for ptrace and Linux Kernel documentation. This covers
+how LLDB has chosen to use those things and how that effects your experience as
+a user.
+
+Scalable Vector Extension (SVE)
+-------------------------------
+
+See `here <https://developer.arm.com/Architectures/Scalable%20Vector%20Extensions>`__
+to learn about the extension and `here <https://kernel.org/doc/html/latest/arch/arm64/sve.html>`__
+for the Linux Kernel's handling of it.
+
+In LLDB you will be able to see the following new registers:
+
+* ``z0-z31`` vector registers, each one has size equal to the vector length.
+* ``p0-p15`` predicate registers, each one containing 1 bit per byte in the vector
+  length. Making each one vector length / 8 sized.
+* ``ffr`` the first fault register, same size as a predicate register.
+* ``vg``, the vector length in "granules". Each granule is 8 bytes.
+
+.. code-block::
+
+       Scalable Vector Extension Registers:
+             vg = 0x0000000000000002
+             z0 = {0x01 0x01 0x01 0x01 0x01 0x01 0x01 0x01 <...> }
+           <...>
+             p0 = {0xff 0xff}
+           <...>
+            ffr = {0xff 0xff}
+
+The example above has a vector length of 16 bytes. Within LLDB you will always
+see "vg" as in the ``vg`` register, which is 2 in this case (8*2 = 16).
+Elsewhere you may see "vq" which is the vector length in quadwords (16 bytes)
+elsewhere. Where you see "vl", it is in bytes.
+
+Changing the Vector Length
+..........................
+
+While you can count the size of a P or Z register, it is intended that ``vg`` be
+used to find the current vector length.
+
+vg can be written. Writing the current vector length changes nothing. If you
+increase the vector length, the registers will likely be reset to 0. If you
+decrease it, LLDB will truncate the Z registers but everything else will be reset
+to 0.
+
+Generally you should not assume that SVE state after changing the vector length
+is in any way the same as it was previously. If you need to do it, do it before
+a function's first use of SVE.
+
+Z Register Presentation
+.......................
+
+LLDB makes no attempt to predict how an SVE Z register will be used. Even if the
+next SVE instruction (which may some distance away) would use, for example, 32
+bit elements, LLDB prints ``z0`` as single bytes.
+
+If you know what format you are going to use, give a format option::
+
+  (lldb) register read z0 -f uint32_t[]
+      z0 = {0x01010101 0x01010101 0x01010101 0x01010101}
+
+FPSIMD and SVE Modes
+....................
+
+Prior to the debugee's first use of SVE, it is in what the Linux Kernel terms
+SIMD mode. Only the FPU is being used. In this state LLDB will still show the
+SVE registers however the values are simply the FPU values zero extended up to
+the vector length.
+
+On first access to SVE, the process goes into SVE mode. Now the Z values are
+in the real Z registers.
+
+You can also trigger this with LLDB by writing to an SVE register. Note that
+there is no way to undo this change from within LLDB. However, the debugee
+itself could do something to end up back in SIMD mode.
+
+Expression evaluation
+.....................
+
+If you evaluate an expression, all SVE state is saved prior to, and restored
+after the expression has been evaluated. Including the register values and
+vector length.
+
+Scalable Matrix Extension (SME)
+-------------------------------
+
+See `here <https://community.arm.com/arm-community-blogs/b/architectures-and-processors-blog/posts/scalable-matrix-extension-armv9-a-architecture>`__
+to learn about the extension and `here <https://kernel.org/doc/html/latest/arch/arm64/sme.html>`__
+for the Linux Kernel's handling of it.
+
+SME adds a "Streaming Mode" to SVE. This mode has its own vector length.
+
+In LLDB you will see the following new registers:
+
+* ``tpidr2``, an extra per thread pointer reserved for use by the SME ABI.
+  This is not scalable, just pointer sized aka 64 bit.
+* ``z0-z31`` streaming SVE registers. These have the same names as the
+  non-streaming registers and therefore you will only see the active set in
+  LLDB. You cannot read or write the inactive mode's registers. Their size
+  is the same as the streaming vector length.
+* ``za`` the Array Storage register. The "Matrix" part of "Scalable Matrix
+  Extension". This is a square made up of rows of length equal to the streaming
+  vector length (svl). Meaning that the total size is svl * svl.
+* ``svg`` the vector length in granules. This acts the same as ``vg`` for SVE.
+  Except that where ``vg`` shows the length for the active mode, ``svg`` will
+  always show the streaming vector length, even in non-streaming mode. This
+  register is read only.
----------------
DavidSpickett wrote:

I've noted explicitly in a few more places that the streaming and non-streaming vector length are not tied together.

Then I've moved the comment about the contents of svg and vg into its own note after the bullet points. So that it can be properly explained.

https://github.com/llvm/llvm-project/pull/66767


More information about the lldb-commits mailing list