[polly] Add diagnostic remark for ReportVariantBasePtr

[Tobias Grosser]

On 26/06/2014 22:12, Andreas Simbuerger wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
>
>
> On 06/26/2014 09:25 PM, Tobias Grosser wrote:
>> On 26/06/2014 21:16, Andreas Simbürger wrote:
>>> What was the exact output?
>>
>> polly-clang /tmp/test.c  -c -mllvm -polly -O3  -Rpass-missed=polly
>> -mllvm -polly-detect-track-failures -mllvm
>> -polly-detect-keep-going -gcolumn-info /tmp/test.c:6:9: remark: The
>> following errors keep this region from being a Scop.
>> [-Rpass-missed=polly-detect] for (int i=0; i<32; i++) ^
>> /tmp/test.c:7:6: remark: The base address of this array is not
>> invariant inside the loop [-Rpass-missed=polly-detect] A->b[i] =
>> A->b[i+p]; ^ /tmp/test.c:7:6: remark: The base address of this
>> array is not invariant inside the loop
>> [-Rpass-missed=polly-detect]
>>
>>
>>> I would expect the source location reported for the write and
>>> then just the message with the line/column info but without
>>> source code.
>>
>> Both remarks give the very same location. I would have expected
>> the first one referring to the first memory address, the second one
>> to the second memory address. (I would also expect to see the
>> source line twice. Once with the caret at the first read, once at
>> the second).
>>
>
> That would be nice. However, if you have a look at the attached *.ll
> file (your example translated with the command-line you posted), you
> will see that the dbg metadata that is attached to all instructions
> inside the loop body is:
>
> !26 = metadata !{i32 7, i32 5, metadata !18, null}.
>
> So no matter what instruction I use inside the body, I will only be
> able to report line 7 column 5 :o(.
> The fact that both errors have the same DebugLoc leads to to effect
> that you don't get the same source location printed twice in a row (I
> guess).
>
> How can we increase the precision of the column info?

This clang hack helps:

diff --git a/lib/CodeGen/CGExpr.cpp b/lib/CodeGen/CGExpr.cpp
index 62ecc73..53c63f4 100644
--- a/lib/CodeGen/CGExpr.cpp
+++ b/lib/CodeGen/CGExpr.cpp
@@ -771,6 +771,9 @@ LValue CodeGenFunction::EmitCheckedLValue(const Expr 
*E, TypeCheckKi
  /// length type, this is not possible.
  ///
  LValue CodeGenFunction::EmitLValue(const Expr *E) {
+  if (CGDebugInfo *DI = getDebugInfo()) {
+  DI->EmitLocation(Builder, E->getLocEnd(), true);
+  }
    switch (E->getStmtClass()) {
    default: return EmitUnsupportedLValue(E, "l-value expression");


/tmp/test.c:7:16: remark: The base address of this array is not 
invariant inside the loop [-Rpass-missed=polly-detect]
      A->b[i] = A->b[i+p];
                ^
/tmp/test.c:7:6: remark: The base address of this array is not invariant 
inside the loop [-Rpass-missed=polly-detect]
      A->b[i] = A->b[i+p];
      ^
The patch is probably to aggressive, but it may be a start. Could we 
possibly emit column info for all load instructions? That would be good 
enough for Polly at least, no? We probably need to understand the 
overhead and if other people could also use this information.

Tobias