When running deltablue (hacked to not be rate-based, but run a similar number of iterations as a non-valgrind run) callgrind shows 6.25% of our instructions under js_Array. Doing s/new Array()/[]/ in the benchmark shows a 6% improvement in the v8 score. Type information (or even just plain JM ic tricks) should let us make this optimization pretty easily. Unfortunately, even with this predicted speedup, d8 w/ crankshaft is still 67% faster. According to callgrind, we actually execute 36% fewer insns and make 2x fewer data references; however, we make up for it by having 80 TIMES more simulated L1 cache misses and 13 TIMES more simulated last-level cache misses. 42% of this is under GC, 34% under CreateThis, and 14% under NewDenseEmptyArray, so this will need GC/obj-shrink work.