[Valgrind-developers] (valgrind): memcheck/tests/new_override regression test failing

[Julian S. <js...@ac...>]

[Dirk .. (1) thanks for fixing 'make install'.  However ...
         (2) pls read at least the last para 
             .. more fallout from recent changes?]

Mystery solved ... possibly a side effect of fixing 'make install'.


> As per subject... the bug has been introduced within the last few days, I
> think, because I didn't have it in my workspace copy until I updated and
> rebuilt just then.
>
> To see, run one of:
>
>   $PREFIX/bin/vg_regtest --all
>   $PREFIX/bin/vg_regtest memcheck
>   $PREFIX/bin/vg_regtest memcheck/tests/new_override
>
> within the top level valgrind/ directory.
>
> I don't see how any of my recent commits could have caused this, so I'm
> brazenly assuming it's not my fault until proven otherwise... :)

Sigh.  Curiously, it also fails on SuSE 8.1.  The .stderr.exp says there
should be an uninit-var error at line 25.  The .stderr.out mentions 
no such.

Reading the assembly for main, with and without -O is revealing.  

  for (i = 0; i < 2; i++) {
    if (toto[i].a) {
      j++;
    }
  // main returns now

With -O and above, gcc sees that j is dead after the loop, and converts 
it to

        movl    $0, %eax
.L15:
        incl    %eax
        cmpl    $1, %eax
        jle     .L15

ie 

  for (i = 0; i < 2; i++) {
  }

[Why it doesn't then unroll the loop and change this to "i = 2",
 I don't know.  But anyway ...]
So it's no wonder no error is reported for me.  

Rerunning the test without -O does indeed get an error reported.

And indeed I see that the test is built with -O2 -g.

Have a look at rev 1.6 of valgrind/memcheck/tests/Makefile.am.
It changes the flag handling for the tests (I guess) and I bet
that causes -O2 to be passed to the test compilations.  Which
I guess it shouldn't be.  Didn't we have a no-optimisation policy
for regressions, to try and ensure gcc didn't optimise stuff away?

Ha .. more clarity.  In memcheck/tests, it is only "mismatches" and
"new_override" which are compiled with -g -O2.  And these are 
exactly the C++ ones.

So I wonder if, in valgrind/memcheck/tests/Makefile.am,
the renaming of CXXFLAGS to AM_CXXFLAGS didn't have the desired
effect?

J