On Mon, Dec 8, 2008 at 8:54 AM, Eric Fisher <[EMAIL PROTECTED]> wrote: > Hi, > > For if-conversion pass (pass_if_after_combine), we can see there're > some IF-THEN-ELSE cases which we try to transform. Let's say > find_if_case_1, for an example. > > (1) > if (test) goto over; // x not live > x = a; > goto label; > over: > > becomes > > x = a; > if (! test) goto label; > > One of my questions is what the benefit this convertion can attain?
It avoids a branch instruction. For the first case, the path via "x = a" is "test -> jump -> set -> jump" (where "set" is th "x = a"-instruction). This becomes "set -> test -> jump". So one jump less. The assumption is that a branch instruction is much more expensive than a register set. Which brings us to... > Also, there's a precondition, which the THEN bb must cost cheap > then BRANCH_COST. I'm not clear about the reason. ... this test to verify the abovementioned assumption. THEN_BB contains the "x = a"-instruction. So the precondition verifies that the branch instruction is expensive enough to avoid compared to executing "x = a" on a path where x is not live. So, the decision to perform the transformation or not, is a cost trade-off between executing an extra branch instruction and creating partially dead code. Gr. Steven
