Re: [PATCH] loop invariant motion expensive codes

[David Edelsohn <dje at watson dot ibm dot com>]

>>>>> Jeffrey A Law writes:

Jeff> First, to move the divide you must be sure the division will always
Jeff> execute in its original position.  Otherwise you can introduce a
Jeff> spurious divide by zero fault.

Jeff> We have also avoided hoisting division operations in the past as
Jeff> doing so changes the point at which a divide by zero fault will
Jeff> occur.

	This patch does not change whether the optimization can or does
hoist a FP divide.  The loop invariant motion pass tests the safety of the
motion elsewhere.  The pass has a cost model and the patch makes the cost
model consistent for all types of divides.

David