Compiler directed speculation for embedded clustered EPIC machines

Very Large Instruction Word (VLIW)/Explicitly Parallel Instruction Computing (EPIC) processors are a very attractive platform for many of today's multimedia and communications applications. In particular, clustered VLIW/EPIC machines can take aggressive advantage of the available instruction level parallelism (ILP), while maintaining high energy-delay ef- ciency. However, multicluster machines are more challenging to compile to than centralized machines. In this thesis, we propose a novel compilerdirected resource-aware ILP extraction technique, called predicated switching, that is targeted towards such multicluster VLIW/EPIC machines. The proposed technique integrates three powerful ILP extraction techniques { predication, speculation and software pipelining, in a combined framework. The three novel contributions in this dissertation are: (1) a compiler transformation, denoted Static Single Assignment - Predicated Switching (SSA-PS), that leverages required data transfers between clusters for performance gains; (2) a static speculation algorithm to decide which speci c kernel operations should actually be speculated in a region of code (hyperblock), possibly being simultaneously software pipelined, so as to maximize execution performance on the target processor; and (3) an ILP extraction ow incorporating several code generation phases critical to pro table ILP extraction by the compiler. Experimental results performed on a representative set of time critical kernels compiled for a number of target machines show that, when compared to two baseline \resource-unaware" speculation techniques (one that speculates aggressively and one that speculates conservatively), predicated switching improves performance with respect to at least one of the baselines in 65% of the cases by up to 50%. Moreover, we show that code size and register pressure are not adversely affected by our technique. Finally, we show that our ILP extraction framework combining speculation and software pipelining can effectively exploit the relative merits of both techniques.