https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97515
--- Comment #7 from Andrew Macleod <amacleod at redhat dot com> --- Created attachment 49607 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49607&action=edit tweak testcase The fix for PR 93781 changed the order of some evaluations in the testcase. the problem is that we have [-1,-1] >> VARYING This use to overflow in the cross product code and just return VARYING. now that its being masked to [-1,-1] >> [0,31] so in the end, the rshift cross product code decides the result is [-1, -1] instead of varying. THis ends up changing the calculations, and we realized certain other parts of the program are unreachable sooner.. we eventually figure out all the same things, but the problem now is that the first statement evaluated: xe_3 = PHI <0(2), xe_11(5)> which we now dont determine xe_11 is 0 right off the bat like we use to. THe final IL does look like: <bb 2> : goto <bb 6>; [INV] <bb 3> : ui_8 = ~xe_3; if (ui_8 == 0) goto <bb 4>; [INV] else goto <bb 5>; [INV] <bb 4> : <bb 5> : # ui_4 = PHI <ui_8(3), -1(4)> <bb 6> : # xe_3 = PHI <0(2), 0(5)> if (xe_3 <= 0) goto <bb 3>; [INV] else goto <bb 7>; [INV] <bb 7> : # xe_5 = PHI <xe_3(6)> return xe_5; The problem is we now don't fold away the PHI: # ui_4 = PHI <ui_8(3), -1(4)> because although the code looks obvious that ui_8 must be 0 on the edge 3->5, ranger has determined that xe_3 must be 0 feeding into this block, which means ui_8 must be -1, which means the TRUE edge can never be taken, and therefore ui_8 is actually undefined on this edge. THe stmt fold routine only looks for constant values in phi arguements, and thus cannot fold away the PHI anymore since there isnt a constant value on the edge. All the same things will fold away here, but the undefined bit is preventing it from happening within EVRP. THe next pass of CCP does turn this into: <bb 2> [local count: 10631108]: <bb 3> [local count: 1073741824]: goto <bb 3>; [100.00%] SO for now at least, I'll just change the testcase to confirm its down to a single goto by this pass. Ideally we'd be able to detect the undefined during statement folding and allow it to match with any other constants.. much like we do with undefined local values.
