On 1/9/20 4:51 PM, Segher Boessenkool wrote: > On Thu, Jan 09, 2020 at 01:44:59PM -0600, Peter Bergner wrote: >> The other test cases just need a slight adjustment to some of their >> counts. > > What were the changes? Or, I'll just trust you looked at it and it is > okay :-)
I did look. Most of the changes brought us closer to one vsx instruction for each builtin than what we had before. We still have some optimization across builtin functions occurring, so we still don't get a one-to-one mapping of builtin call to vsx instruction. >> I must say that the vsx-vector-6.p*.c tests are fragile! They're so big and >> reusing source operands, that the compiler can sometimes optimize several >> builtin calls together, meaning we don't see as many vsx instructions as >> we have calls to builtins. I didn't bother trying to fix that, since that >> is a lot more work! I just wanted to vent! :-) > > Splitting out separate functions in the testcase shouldn't be so much > work? Or am I too optimistic :-) > > This should make the test a good deal less prone to random changes in > output caused by the lunar cycle. Ok, let me take a stab at rewriting the tests to be more similar to the pr92923-[12].c and see how much work that is. I do agree that it would be nice not having the insn counts be so fragile. Peter
