This document describes how debug info works at the SIL level and how to correctly update debug info in SIL optimization passes. This document is inspired by its LLVM analog, How to Update Debug Info: A Guide for LLVM Pass Authors, which is recommended reading, since all of the concepts discussed there also apply to SIL.
Contrary to LLVM IR, SIL makes source locations and lexical scopes mandatory on
all instructions. SIL transformations should follow the LLVM guide for when to
merge drop and copy locations, since all the same considerations apply. Helpers
like SILBuilderWithScope
make it easy to copy source locations when expanding
SIL instructions.
Warning
Don't use SILBuilderWithScope
when replacing a single instruction of type
AllocStackInst
or DebugValueInst
. These meta instructions are skipped,
so the wrong scope will be inferred.
Each debug_value
(and variable-carrying instruction) defines an update point
for the location of (part of) that source variable. A variable location is an
SSA value, modified by a debug expression that can transform that value,
yielding the value of that variable. Optimizations like SROA may split a source
variable into multiple smaller fragments, other optimizations such as Mem2Reg
may split a debug value describing an address into multiple debug values
describing different SSA values. Each variable (fragment) location is valid
until the end of the current basic block, or until another debug_value
describes another location for a variable fragment for the same unique variable
that overlaps with that (fragment of the) variable.
Source variables are represented by debug_value
instructions, and may also be
described in debug variable-carrying instructions (alloc_stack
, alloc_box
).
There is no semantic difference between describing a variable in an allocation
instruction directly or describing it in an debug_value
following the
allocation instruction.
This is equivalent, and should be optimized similarly:
%0 = alloc_stack $T, var, name "value", loc "a.swift":4:2, scope 1
// equivalent to:
%0 = alloc_stack $T, loc "a.swift":4:2, scope 1
debug_value %0 : $*T, var, name "value", expr op_deref, loc "a.swift":4:2, scope 1
Note
In the future, we may want to remove the debug variable from the alloc_stack
to only use the second form, in order to simplify SIL. Additionally, we could
then move the debug_value
instruction to the point where the variable is
initialized to avoid showing ununitialized memory in the debugger. This would
be a change in SILGen, which should not affect the optimizer.
For now, the DebugVarCarryingInst
type can be used to handle both cases.
Variables are uniquely identified via their debug scope, their location, and their name.
The debug scope, is the range in which the variable is declared and available. More information about debug scopes is available on the Swift blog For arguments, this will be the function's scope, otherwise, this will be a subscope within a function. When a function is inlined, a new scope is created, including information about the inlined function, and in which function it was inlined (inlined_at).
The location of the variable is the source location where the variable was declared.
If the location and scope of a debug variable isn't set, it will use the scope
and location of the instruction, which is correct in most cases. However, if a
debug_value
describes a modification of a variable, the instruction should
have the location of the update point, and the variable must keep the location
of the variable declaration:
%0 = integer_literal $Int, 2
debug_value %0 : $Int, var, name "a", loc "a.swift":2:5, scope 2
%2 = integer_literal $Int, 3
debug_value %2 : $Int, var, (name "a", loc "a.swift":2:5, scope 2), loc "a.swift":3:3, scope 2
For this code:
var a = 2
a = 3
By default the type of the variable will be the object type of the SSA value. If this is not the correct type, a type must be attached to the debug variable to override it.
Example:
debug_value %0 : $*T, let, name "address", type $UnsafeRawPointer
The variable will usually have an associated expression yielding the correct type.
Note
As there are no pointers in Swift, the type should never be an address type.
A variable can have an associated expression if the value needs computation. This can be for dereferencing a pointer, arithmetic, or for splitting structs. An expression is a sequence of operations to be executed left to right. Debug expressions get lowered into LLVM DIExpressions which get lowered into DWARF expressions.
A variable's expression may include an op_deref
, usually at the beginning, in
which case the SSA value is a pointer that must be dereferenced to access the
value of the variable.
In this example, the value returned by the alloc_stack
is an address that must
be dereferenced.
%0 = alloc_stack $T
debug_value %0 : $*T, var, name "value", expr op_deref
SILGen should always use SILBuilder::emitDebugDescription
to create debug
values, which will automatically add an op_deref depending on the type of the
SSA value. As there are no pointers in Swift, this will always do the right
thing. In SIL passes, use SILBuilder::createDebugValue
to create debug values,
or SILBuilder::createDebugValueAddr
to add an op_deref.
Warning
At the optimizer level, Swift Unsafe*Pointer
types can be simplified
to address types. As such, a debug_value
with an address type without an
op_deref
can be valid. SIL passes must not assume that op_deref
and
address types correlate.
Even if op_deref
is usually at the beginning, it doesn't have to be:
debug_value %0 : $*UInt8, let, name "hello", expr op_constu:3:op_plus:op_deref
This will add 3
to the pointer contained in %0
, then dereference the result.
If a variable is partially updated, a fragment can be used to specify that this update refers to an element of an aggregate type.
Tip
When using fragments, always specify the type of the variable, as it will be different from the SSA value.
When SROA is splitting a struct or tuple, it will also split the debug values, and add a fragment to specify which field is being updated.
struct Pair { var a, b: Int }
alloc_stack $Pair, var, name "pair"
// -->
alloc_stack $Int, var, name "pair", type $Pair, expr op_fragment:#Pair.a
alloc_stack $Int, var, name "pair", type $Pair, expr op_fragment:#Pair.b
// -->
alloc_stack $Builtin.Int64, var, name "pair", type $Pair, expr op_fragment:#Pair.a:op_fragment:#Int._value
alloc_stack $Builtin.Int64, var, name "pair", type $Pair, expr op_fragment:#Pair.b:op_fragment:#Int._value
Here, Pair is a struct containing two Ints, so each alloc_stack
will receive a
fragment with the field it is describing. Int, in Swift, is itself a struct
containing one Builtin.Int64 (on 64 bits systems), so it can itself be SROA'ed.
Fragments can be chained to describe this.
Tuple fragments use a different syntax, but work similarly:
alloc_stack $(Int, Int), var, name "pair"
// -->
alloc_stack $Int, var, name "pair", type $(Int, Int), expr op_tuple_fragment:$(Int, Int):0
alloc_stack $Int, var, name "pair", type $(Int, Int), expr op_tuple_fragment:$(Int, Int):1
// -->
alloc_stack $Builtin.Int64, var, name "pair", type $(Int, Int), expr op_tuple_fragment:$(Int, Int):0:op_fragment:#Int._value
alloc_stack $Builtin.Int64, var, name "pair", type $(Int, Int), expr op_tuple_fragment:$(Int, Int):1:op_fragment:#Int._value
Tuple fragments and struct fragments can be mixed freely, however, they must all be at the end of the expression. That is because the fragment operator can be seen as returning a struct containing a single element, with the rest undefined, and, except fragments, no debug expression operator take a struct as input.
Note
When multiple fragments are present, they are evaluated in the reverse way — from the field within the variable first, to the SSA's type at the end
An expression can add or subtract a constant offset to a value. To do so, an
op_constu
or op_consts
can be used to push a constant integer to the stack,
respectively unsigned and signed. Then, the op_plus
and op_minus
operators
can be used to sum or subtract the two values on the stack.
debug_value %0 : $Builtin.Int64, var, name "previous", type $Int, expr op_consts:1:op_minus:op_fragment:#Int._value
debug_value %0 : $Builtin.Int64, var, name "next", type $Int, expr op_consts:1:op_plus:op_fragment:#Int._value
Caution
This currently doesn't work if a fragment is present.
If a debug_value
is describing a constant, such as in let x = 1
, and the
value is optimized out, we can keep it, using a constant expression, and no SSA
value.
debug_value undef : $Int, let, name "x", expr op_consts:1:op_fragment:#Int._value
If the value of the variable cannot be recovered as the value is entirely optimized away, an undef debug value should still be kept:
debug_value undef : $Int, let, name "x"
Additionally, if a previous debug_value
exists for the variable, a debug value
of undef invalidates the previous value, in case the value of the variable isn't
known anymore:
debug_value %0 : $Int, var, name "x" // var x = a
...
debug_value undef : $Int, var, name "x" // x = <optimized out>
Combined with fragments, some parts of the variable can be undefined and some not:
... // pair = ?
debug_value %0 : $Int, var, name "pair", type $Pair, expr op_fragment:#Pair.a // pair.a = x
debug_value %0 : $Int, var, name "pair", type $Pair, expr op_fragment:#Pair.b // pair.b = x
... // pair = (x, x)
debug_value undef : $Pair, var, name "pair", expr op_fragment:#Pair.a // pair.a = <optimized out>
... // pair = (?, x)
debug_value undef : $Pair, var, name "pair" // pair = <optimized out>
... // pair = ?
debug_value %1 : $Int, var, name "pair", type $Pair, expr op_fragment:#Pair.a // pair.a = y
... // pair = (y, ?)
A debug_value
must always describe a correct value for that source variable
at that source location. If a value is only correct on some paths through that
instruction, it must be replaced with undef
. Debug info never speculates.
Optimization passes may never drop a variable entirely. If a variable is
entirely optimized away, an undef
debug value should still be kept. The only
exception is when the variable is in an unreachable function or scope, where it
can be removed with the rest of the instructions.
When a SIL instruction is deleted, call salvageDebugInfo
. It will try to
capture the effect of the deleted instruction in a debug expression, so the
location can be preserved.
Alternatively, you can use an InstructionDeleter
, which will automatically
call salvageDebugInfo
.
If the debug info cannot be salvaged by salvageDebugInfo
, and the pass has a
special knowledge of the value, the pass can directly replace the debug value
with the known value.
If an instruction is being replaced by another, use replaceAllUsesWith
. It
will also update debug values to use the new instruction.
Tip
To detect when a pass drops a variable, you can use the
-Xllvm -sil-stats-lost-variables
to print when a variable is lost by a pass.
More information about this option is available in
Optimizer Counter Analysis