DescriptionMIPS: Enable aligned keyed stores and loads for double arrays.
Provide ~20% performance improvement in Navier-Stokes benchmark, by using
true load/store double (ldc1/sdc1) instructions.
These instructions require double-aligned memory address.
Since doubles are not generally aligned in v8, we historically have used
a workaround using a pair of 32-bit loads/stores (lwc1/swc1), which is slow.
This CL requires FixedDoubleArrays to be aligned in all cases, which is achieved
via these related CL's: https://chromiumcodereview.appspot.com/35103002,
https://chromiumcodereview.appspot.com/35133002,
https://chromiumcodereview.appspot.com/35313002.
TEST=
BUG=
Patch Set 1 #
Messages
Total messages: 1 (0 generated)
|