------- Comment #3 from ubizjak at gmail dot com 2007-06-18 17:40 ------- (In reply to comment #2)
> We need a better explanation than this. Uros agreed to summarize the > IRC discussion to close this issue. It'd be useful if we keep that same > discussion on the source code itself. The need for separate IL passes could be explained by: We have reciprocal pass (in fact CSE recip pass) that CSEs 1.0/z from x/z, y/z, .../z. This is done by scanning function for RDIV_EXPR, where denominator (z) is the same. If 1.0/func() -> rfunc() conversion is done before recip pass, we loose the ability to scan for RDIV_EXPRs and the ability to CSE the division. By putting 1.0/func()->rfunc() after recip pass, we simply run another scan for RDIV_EXPRs with function as their argument. Note that we already CSE'd 1.0/func(), so the conversion into rfunc() is trivial. This is the reason why function recip pass needs to run after recip (aka CSE recip) pass. Next convesion is sqrt(a/b) -> rsqrt(b/a) conversion that runs after rfunc recip pass as a separate pass (so, in the function granulartiy). If this pass is run together or before rfunc pass, then rsqrt pass would convert expressions like 1.0/sqrt(a/b) into 1.0/rsqrt(b/a). There is no point for rfunc pass (that would follow) to convert this to sqrt(b/a). -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32390