The IA-64 Architecture at Work

24 Computer with a predicate value of true executes normally. If the predicate is false, the associated instruction—although issued—does not write its results to registers or memory. Research has shown predication to be effective at removing branches and at decreasing penalties from branch mispredicts.1 A simple code example with a difficult-to-predict branch illustrates how predication can remove the branch. Figure 1a shows the C code for a classic if-then-else statement. In a traditional architecture, the processor loads the data from memory, compares the value of a(i).ptr with zero, and uses the compare’s (cmp’s) result in a conditional-branch instruction. Because of the conditional branch, a traditional compiler structures this code into four basic blocks, as shown in Figure 1b. The processor must execute the instructions of all four blocks serially, and branch instructions are barriers to ILP. Predication is used to remove the difficult-to-predict branch in the first basic block. In the IA-64 architecture, compare instructions generate two predicates, as shown in Figure 1c: The IA-64 Architecture at Work