I do not buy some of HAVEGE, specifically the claim made here "tens of thousands of unpredictable bits can be gathered per operating system call in average", and the methodology used to support that claim, as found here.
Entropy gathering is described by this pseudocode:
(Emphasis using bold is mine)
The data gathered in
The entropy at the output of that PRNG has been evaluated using an experimental test (not designed specifically against this PRNG). That's a cardinal mistake: any half-decent PRNG will pass this test, even if fed very little actual entropy in its input, which makes its output highly predictable.
While this does not constitute a practical break of HAVEGE, it shows that the justification of its entropy-gathering part is flawed. My recommendation is thus not to use HAVEGE.
Update: the above is advice about HAVEGE, as in the original title and some of the text of the original question, based on references from 2003. In it's successor HAVEGED (notice the additional D) of the 201x, there was some effort made towards a sound methodology, in particular checking the entropy at input of the PRNG, leading to the conclusion: "Most samples failed all tests!". It is acknowledged that theoretical analysis of the source of entropy to build a model suitable to support a demonstration of security "were not sufficient to constitute such a model". This at least is quite believable.