1 # Debug Info Assignment Tracking
3 Assignment Tracking is an alternative technique for tracking variable location
4 debug info through optimisations in LLVM. It provides accurate variable
5 locations for assignments where a local variable (or a field of one) is the
6 LHS. In rare and complicated circumstances indirect assignments might be
7 optimized away without being tracked, but otherwise we make our best effort to
8 track all variable locations.
10 The core idea is to track more information about source assignments in order
11 and preserve enough information to be able to defer decisions about whether to
12 use non-memory locations (register, constant) or memory locations until after
13 middle end optimisations have run. This is in opposition to using
14 `#dbg_declare` and `#dbg_value`, which is to make the decision for most
15 variables early on, which can result in suboptimal variable locations that may
16 be either incorrect or incomplete.
18 A secondary goal of assignment tracking is to cause minimal additional work for
19 LLVM pass writers, and minimal disruption to LLVM in general.
23 **Status**: Enabled by default in Clang but disabled under some circumstances
24 (which can be overridden with the `forced` option, see below). `opt` will not
25 run the pass unless asked (`-passes=declare-to-assign`).
28 `-Xclang -fexperimental-assignment-tracking=<disabled|enabled|forced>`
30 When enabled Clang gets LLVM to run the pass `declare-to-assign`. The pass
31 converts conventional debug records to assignment tracking metadata and sets
32 the module flag `debug-info-assignment-tracking` to the value `i1 true`. To
33 check whether assignment tracking is enabled for a module call
34 `isAssignmentTrackingEnabled(const Module &M)` (from `llvm/IR/DebugInfo.h`).
36 ## Design and implementation
38 ### Assignment markers: `#dbg_assign`
40 `#dbg_value`, a conventional debug record, marks out a position in the
41 IR where a variable takes a particular value. Similarly, Assignment Tracking
42 marks out the position of assignments with a record called `#dbg_assign`.
44 In order to know where in IR it is appropriate to use a memory location for a
45 variable, each assignment marker must in some way refer to the store, if any
46 (or multiple!), that performs the assignment. That way, the position of the
47 store and marker can be considered together when making that choice. Another
48 important benefit of referring to the store is that we can then build a two-way
49 mapping of stores<->markers that can be used to find markers that need to be
50 updated when stores are modified.
52 An `#dbg_assign` marker that is not linked to any instruction signals that
53 the store that performed the assignment has been optimised out, and therefore
54 the memory location will not be valid for at least some part of the program.
56 Here's the `#dbg_assign` signature. `Value *` type parameters are first wrapped
60 #dbg_assign(Value *Value,
61 DIExpression *ValueExpression,
62 DILocalVariable *Variable,
65 DIExpression *AddressExpression)
68 The first three parameters look and behave like an `#dbg_value`. `ID` is a
69 reference to a store (see next section). `Address` is the destination address
70 of the store and it is modified by `AddressExpression`. An empty/undef/poison
71 address means the address component has been killed (the memory address is no
72 longer a valid location). LLVM currently encodes variable fragment information
73 in `DIExpression`s, so as an implementation quirk the `FragmentInfo` for
74 `Variable` is contained within `ValueExpression` only.
76 ### Instruction link: `DIAssignID`
78 `DIAssignID` metadata is the mechanism that is currently used to encode the
79 store<->marker link. The metadata node has no operands and all instances are
80 `distinct`; equality is checked for by comparing addresses.
82 `#dbg_assign` records use a `DIAssignID` metadata node instance as an
83 operand. This way it refers to any store-like instruction that has the same
84 `DIAssignID` attachment. E.g. For this test.cpp,
91 compiled without optimisations:
93 $ clang++ test.cpp -o test.ll -emit-llvm -S -g -O0 -Xclang -fexperimental-assignment-tracking=enabled
97 define dso_local noundef i32 @_Z3funi(i32 noundef %a) #0 !dbg !8 {
99 %a.addr = alloca i32, align 4, !DIAssignID !13
100 #dbg_assign(i1 undef, !14, !DIExpression(), !13, i32* %a.addr, !DIExpression(), !15)
101 store i32 %a, i32* %a.addr, align 4, !DIAssignID !16
102 #dbg_assign(i32 %a, !14, !DIExpression(), !16, i32* %a.addr, !DIExpression(), !15)
103 %0 = load i32, i32* %a.addr, align 4, !dbg !17
108 !13 = distinct !DIAssignID()
109 !14 = !DILocalVariable(name: "a", ...)
111 !16 = distinct !DIAssignID()
114 The first `#dbg_assign` refers to the `alloca` through `!DIAssignID !13`,
115 and the second refers to the `store` through `!DIAssignID !16`.
117 ### Store-like instructions
119 In the absence of a linked `#dbg_assign`, a store to an address that is
120 known to be the backing storage for a variable is considered to represent an
121 assignment to that variable.
123 This gives us a safe fall-back in cases where `#dbg_assign` records have
124 been deleted, the `DIAssignID` attachment on the store has been dropped, or the
125 optimiser has made a once-indirect store (not tracked with Assignment Tracking)
128 ### Middle-end: Considerations for pass-writers
130 #### Non-debug instruction updates
132 **Cloning** an instruction: nothing new to do. Cloning automatically clones a
133 `DIAssignID` attachment. Multiple instructions may have the same `DIAssignID`
134 instruction. In this case, the assignment is considered to take place in
135 multiple positions in the program.
137 **Moving** a non-debug instruction: nothing new to do. Instructions linked to a
138 `#dbg_assign` have their initial IR position marked by the position of the
141 **Deleting** a non-debug instruction: nothing new to do. Simple DSE does not
142 require any change; it’s safe to delete an instruction with a `DIAssignID`
143 attachment. A `#dbg_assign` that uses a `DIAssignID` that is not attached
144 to any instruction indicates that the memory location isn’t valid.
146 **Merging** stores: In many cases no change is required as `DIAssignID`
147 attachments are automatically merged if `combineMetadata` is called. One way or
148 another, the `DIAssignID` attachments must be merged such that new store
149 becomes linked to all the `#dbg_assign` records that the merged stores
150 were linked to. This can be achieved simply by calling a helper function
151 `Instruction::mergeDIAssignID`.
153 **Inlining** stores: As stores are inlined we generate `#dbg_assign`
154 records and `DIAssignID` attachments as if the stores represent source
155 assignments, just like the in frontend. This isn’t perfect, as stores may have
156 been moved, modified or deleted before inlining, but it does at least keep the
157 information about the variable correct within the non-inlined scope.
159 **Splitting** stores: SROA and passes that split stores treat `#dbg_assign`
160 records similarly to `#dbg_declare` records. Clone the
161 `#dbg_assign` records linked to the store, update the FragmentInfo in
162 the `ValueExpression`, and give the split stores (and cloned records) new
163 `DIAssignID` attachments each. In other words, treat the split stores as
164 separate assignments. For partial DSE (e.g. shortening a memset), we do the
165 same except that `#dbg_assign` for the dead fragment gets an `Undef`
168 **Promoting** allocas and store/loads: `#dbg_assign` records implicitly
169 describe joined values in memory locations at CFG joins, but this is not
170 necessarily the case after promoting (or partially promoting) the
171 variable. Passes that promote variables are responsible for inserting
172 `#dbg_assign` records after the resultant PHIs generated during
173 promotion. `mem2reg` already has to do this (with `#dbg_value`) for
174 `#dbg_declare`s. Where a store has no linked record, the store is
175 assumed to represent an assignment for variables stored at the destination
178 #### Debug record updates
180 **Moving** a debug record: avoid moving `#dbg_assign` records where
181 possible, as they represent a source-level assignment, whose position in the
182 program should not be affected by optimization passes.
184 **Deleting** a debug record: Nothing new to do. Just like for conventional
185 debug records, unless it is unreachable, it’s almost always incorrect to
186 delete a `#dbg_assign` record.
188 ### Lowering `#dbg_assign` to MIR
190 To begin with only SelectionDAG ISel will be supported. `#dbg_assign`
191 records are lowered to MIR `DBG_INSTR_REF` instructions. Before this happens
192 we need to decide where it is appropriate to use memory locations and where we
193 must use a non-memory location (or no location) for each variable. In order to
194 make those decisions we run a standard fixed-point dataflow analysis that makes
195 the choice at each instruction, iteratively joining the results for each block.
199 Outstanding improvements:
201 * As mentioned in test llvm/test/DebugInfo/assignment-tracking/X86/diamond-3.ll,
202 the analysis should treat escaping calls like untagged stores.
204 * The system expects locals to be backed by a local alloca. This isn't always
205 the case - sometimes a pointer to storage is passed into a function
206 (e.g. sret, byval). We need to be able to handle those cases. See
207 llvm/test/DebugInfo/Generic/assignment-tracking/track-assignments.ll and
208 clang/test/CodeGen/assignment-tracking/assignment-tracking.cpp for examples.
210 * `trackAssignments` doesn't yet work for variables that have their
211 `#dbg_declare` location modified by a `DIExpression`, e.g. when the
212 address of the variable is itself stored in an `alloca` with the
213 `#dbg_declare` using `DIExpression(DW_OP_deref)`. See `indirectReturn` in
214 llvm/test/DebugInfo/Generic/assignment-tracking/track-assignments.ll and in
215 clang/test/CodeGen/assignment-tracking/assignment-tracking.cpp for an
218 * In order to solve the first bullet-point we need to be able to specify that a
219 memory location is available without using a `DIAssignID`. This is because
220 the storage address is not computed by an instruction (it's an argument
221 value) and therefore we have nowhere to put the metadata attachment. To solve
222 this we probably need another marker record to denote "the variable's
223 stack home is X address" - similar to `#dbg_declare` except that it needs
224 to compose with `#dbg_assign` records such that the stack home address
225 is only selected as a location for the variable when the `#dbg_assign`
226 records agree it should be.
228 * Given the above (a special "the stack home is X" record), and the fact
229 that we can only track assignments with fixed offsets and sizes, I think we
230 can probably get rid of the address and address-expression part, since it
231 will always be computable with the info we have.
233 * Assignment tracking is disabled by default for LTO and thinLTO builds, and
234 if LLDB debugger tuning has been specified. We should remove these
235 restrictions. See EmitAssemblyHelper::RunOptimizationPipeline in
236 clang/lib/CodeGen/BackendUtil.cpp.