[WebAssembly] WebAssemblyDebugValueManager cannot handle DBG_VALUE_LISTs (original) (raw)

Extended Description

WebAssemblyDebugValueManager class
(https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/WebAssembly/WebAssemblyDebugValueManager.cpp)
was created when there was no DBG_VALUE_LIST instruction, so it assumes all
debug value instructions are DBG_VALUEs, each of which can have at most a single
register operand. WebAssemblyDebugValueManager groups a register def with its
corresponding DBG_VALUEs that refer to the def.

1. When a reg def moves, we can't move DBG_VALUE_LISTs like DBG_VALUEs.

With DBG_VALUE_LIST instruction this relationship does not hold anymore.
DBG_VALUE_LIST can contain multiple register operands and a complex expression
describing the relationship between those operands. So when moving an
instruction that defines a register, we can move DBG_VALUEs referring to that
def but can't safely do the same thing with DBG_VALUE_LISTs. For example, see
the pseudocode below:

%0 = ... DBG_VALUE %0, ... %1 = ... ... %2 = ADD %1, %2

When we move %0 right before %2, we can move the DBG_VALUE with it:

%1 = ... ... %0 = ... DBG_VALUE %0, ... %2 = ADD %1, %2

But if it is a DBG_VALUE_LIST, there can be multiple register operands that
aren't supposed to be moved:

%0 = ... DBG_VALUE_LIST ... %0, %1, ... %1 = ... ... %2 = ADD %1, %2

When we move %0 right before %2, we can't move the DBG_VALUE_LIST in the same
way, because it changes the meaning of the DBG_VALUE_LIST:

%1 = ... ... %0 = ... DBG_VALUE_LIST ... %0, %1, ... ;; Error. %1's value changed by new def %2 = ADD %1, %2

But currently WebAssemblyDebugValueManager moves all DBG_VALUE_LISTs around
in the same way, which should be fixed.

1. We can't copy DBG_VALUE_LISTs in a brute-force manner as we do for
   DBG_VALUEs.

A single DBG_VALUE_LIST instruction can contain any number of register
operands. In real world code I have seen a DBG_VALUE_LIST with dozens of them.

In WebAssemblyRegisterStackify, when we create a tee.local for a reg def, one
reg def becomes three reg defs, which is well illustrated in this comment:

(In the old code, there's one def 'Reg', but in the new code, there are three
with the same value: 'Reg', 'DefReg', and 'TeeReg')

And when a def becomes three defs, each corresponding DBG_VALUE is copied
twice so there will be three copies of it. But if a DBG_VALUE_LIST contains
dozens of registers and each of them gets tee'd, the DBG_VALUE_LIST can be
copied exponentially. For example, if a DBG_VALUE_LIST contains %0, %1, and
%2, after %0 is tee'd, the instruction has 3 copies. And after %1 is tee'd, each
of 3 copies becomes 3 copies, so there will be 9 copies. And after %2 is tee'd
there will be 27 copies.

This way, in real world code, I have seen a BB with ~100 instruction (including
DBG_VALUEs and DBG_VALUE_LISTs) becomes several million instructions,
effectively stalling the compilation process. It will look like
infinite-looping.

To fix both bugs, we should come up with a way so that it can handle
DBG_VALUE_LISTs correctly and efficiently. In the meantime, I am planning to
nullify those DBG_VALUE_LIST instructions to make them appear as "optimized
out", which is at least correct.