Evaluation of Compilers’ Capability of Automatic Vectorization Based on Source Code Analysis

Table 4

Automatic vectorization speedup ratio.

Group	Type	Numbers	GCC	LLVM	ICC
Feature	Control flow	16	2|1.36	1|0.90	13|1.94
	Function	6	4|1.93	5|1.82	6|2.24
	Goto	12	1|1.99	0	7|1.76
	Indirect addressing	10	0	1|0.99	0
	Induction	10	8|2.00	4|1.5	7|1.79
	No dependency	16	12|2.15	9|2.12	11|2.96
	Loop reversal	2	2|1.89	1|1.91	1|2.03
	Reduction	16	9|8.81	0	13|3.66
	Regular dependency	9	3|1.5	2|1.91	2|1.41
	Rerolling	3	1|2.33	3|1.00	2|2.68
	Alias	5	5|2.29	5|2.44	4|1.25
	Symbolic resolution	10	10|2.45	9|2.43	7|2.29
Method	Node splitting	6	1|1.01	2|1.93	6|2.28
	Loop distribution	3	1|4.02	0	3|1.51
	Loop interchanging	6	3|2.20	1|1.07	3|1.81
	Loop peeling	4	0	0	2|2.51
	Scalar expansion	14	6|1.82	6|1.82	8|1.81
	Statement reordering	3	0	0	3|1.77