[RemoveDIs] Add documentation for IR debug records (#81156)

This patch adds minimal documentation for the IR representation of the
RemoveDIs model of debug info. This patch assumes
that the default for all cases is still debug intrinsic functions, and
so does not update all existing text to refer to debug records, but only
adds a section in the LangRef and SourceLevelDebugging documents to
explain the new format.

llvm/docs/LangRef.rst

@@ -908,7 +908,8 @@ Syntax::
 
 A function definition contains a list of basic blocks, forming the CFG (Control
 Flow Graph) for the function. Each basic block may optionally start with a label
-(giving the basic block a symbol table entry), contains a list of instructions,
+(giving the basic block a symbol table entry), contains a list of instructions
+and :ref:`debug records <debugrecords>`,
 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
 function return). If an explicit label name is not provided, a block is assigned
 an implicit numbered label, using the next value from the same counter as used
@@ -8541,7 +8542,10 @@ The LLVM instruction set consists of several different classifications
 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
 instructions <binaryops>`, :ref:`bitwise binary
 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
-:ref:`other instructions <otherops>`.
+:ref:`other instructions <otherops>`. There are also :ref:`debug records
+<debugrecords>`, which are not instructions themselves but are printed
+interleaved with instructions to describe changes in the state of the program's
+debug information at each position in the program's execution.
 
 .. _terminators:
 
@@ -12695,6 +12699,29 @@ Example:
 
       %tok = cleanuppad within %cs []
 
+.. _debugrecords:
+
+Debug Records
+-----------------------
+
+Debug records appear interleaved with instructions, but are not instructions;
+they are used only to define debug information, and have no effect on generated
+code. They are distinguished from instructions by the use of a leading `#` and
+an extra level of indentation. As an example:
+
+.. code-block:: llvm
+
+  %inst1 = op1 %a, %b
+    #dbg_value(%inst1, !10, !DIExpression(), !11)
+  %inst2 = op2 %inst1, %c
+
+These debug records are an optional replacement for
+:ref:`debug intrinsics<dbg_intrinsics>`. Debug records will be output if the
+``--write-experimental-debuginfo`` flag is passed to LLVM; it is an error for both
+records and intrinsics to appear in the same module. More information about
+debug records can be found in the `LLVM Source Level Debugging
+<SourceLevelDebugging.html#format-common-intrinsics>`_ document.
+
 .. _intrinsics:
 
 Intrinsic Functions

llvm/docs/SourceLevelDebugging.rst

@@ -167,6 +167,17 @@ conventions used by the C and C++ front-ends.
 Debug information descriptors are `specialized metadata nodes
 <LangRef.html#specialized-metadata>`_, first-class subclasses of ``Metadata``.
 
+There are two models for defining the values of source variables at different
+states of the program and tracking these values through optimization and code
+generation: :ref:`intrinsic function calls <format_common_intrinsics>`, the
+current default, and :ref:`debug records <debug_records>`, which are a new
+non-instruction-based model
+(for an explanation of how this works and why it is desirable, see the
+`RemoveDIs <RemoveDIsDebugInfo.html>`_ document). Each module must use one or
+the other; they may never be mixed within an IR module. To enable writing debug
+records instead of intrinsic calls, use the flag
+``--write-experimental-debuginfo``.
+
 .. _format_common_intrinsics:
 
 Debugger intrinsic functions
@@ -268,6 +279,61 @@ The formal LLVM-IR signature is:
 
 See :doc:`AssignmentTracking` for more info.
 
+.. _debug_records:
+
+Debug Records
+----------------------------
+
+LLVM also has an alternative to intrinsic functions, debug records, which
+function similarly but are not instructions. The basic syntax for debug records
+is:
+
+.. code-block:: llvm
+
+    #dbg_<kind>([<arg>, ]* <DILocation>)
+  ; Using the intrinsic model, the above is equivalent to:
+  call void llvm.dbg.<kind>([metadata <arg>, ]*), !dbg <DILocation>
+
+A debug intrinsic function can be converted to a debug record with the
+following steps:
+
+1. Add an extra level of indentation.
+2. Replace everything prior to the intrinsic kind (declare/value/assign) with
+   ``#dbg_``.
+3. Remove the leading ``metadata`` from the intrinsic's arguments.
+4. Transfer the ``!dbg`` attachment to be an argument, dropping the leading
+   ``!dbg``.
+
+For each kind of intrinsic function, there is an equivalent debug record.
+
+``#dbg_declare``
+^^^^^^^^^^^^^^^^
+
+.. code-block:: llvm
+
+    #dbg_declare([Value|MDNode], DILocalVariable, DIExpression, DILocation)
+
+Equivalent to the ``llvm.dbg.declare`` intrinsic.
+
+``#dbg_value``
+^^^^^^^^^^^^^^
+
+.. code-block:: llvm
+
+    #dbg_value([Value|DIArgList|MDNode], DILocalVariable, DIExpression, DILocation)
+
+Equivalent to the ``llvm.dbg.value`` intrinsic.
+
+``#dbg_assign``
+^^^^^^^^^^^^^^^
+
+.. code-block:: llvm
+
+    #dbg_assign([Value|DIArgList|MDNode], DILocalVariable, DIExpression,
+                DIAssignID, [Value|MDNode], DIExpression, DILocation)
+
+Equivalent to the ``llvm.dbg.assign`` intrinsic.
+
 Object lifetimes and scoping
 ============================