[PATCH] D37078: [DebugInfo] Handle memory locations in LiveDebugValues

Reid Kleckner via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Thu Aug 24 10:14:25 PDT 2017


rnk added inline comments.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp:5173
+    // node map.
+    // FIXME: Handle byval/inalloca params as above.
+    if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
----------------
rnk wrote:
> aprantl wrote:
> > I think this should be a separate commit?
> Ugh, this doesn't actually do the right thing. Here's my test that comes here:
> ```
> void useptr(int *);
> void use_alloca(int cond) {
>   int x = 0;
>   int *px = &x;
>   if (cond)
>     useptr(px);
> }
> ```
> 
> After optimization, `px` will be described by a dbg.value of x's alloca. This code in InstrEmitter.cpp incorrectly builds an "indirect in memory" DBG_VALUE for px:
> ```
>   if (SD->getKind() == SDDbgValue::FRAMEIX) {
>     // Stack address; this needs to be lowered in target-dependent fashion.
>     // EmitTargetCodeForFrameDebugValue is responsible for allocation.
>     return BuildMI(*MF, DL, TII->get(TargetOpcode::DBG_VALUE))
>         .addFrameIndex(SD->getFrameIx())
>         .addImm(0)
>         .addMetadata(Var)
>         .addMetadata(Expr);
>   }
> ```
> 
> I guess I'll need to fix that to look for DW_OP_deref.
Alternatively, you could say I misunderstood `DAG.getFrameIndexDbgValue`.


https://reviews.llvm.org/D37078





More information about the llvm-commits mailing list