816661 - BaselineCompiler: Optimize ARM-specific codegen

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect)

Description

[Jan de Mooij [:jandem]]

Our ARM-specific code should be optimized to have less x86-isms. For instance, we could use a 3-operand add instead of a move + 2-operand add.

Seems best to do this after the code has stabilized a bit more though.

Comment 1

[Cameron Kaiser [:spectre]]

For PowerPC Baseline Compiler, I'm also rewriting some of the guard code. It's not very efficient for CPUs that pay a higher branch penalty. Some of this could be beneficial to ARM as well (e.g., instead of two branchTestInt32s, you could have load temp,JSVAL_INT_32; xor scratch1,R0.typeReg,temp; xor scratch2,R1.typeReg,temp; or temp,scratch1,scratch2; branch(NonZero, failure) -- this gets a branch out of the fastpath and replaces it with fast integer unit work).

Comment 2

[Kannan Vijayan [:djvj]]

Closing this as wontfix. Codegen is a whole separate thing to discuss.