Byte-code optimization

[arakov]

• Reference inline conversion constructor : no need for the frame
• open [1,2,3,4}, x
• dcopy - x86-64, aarch64, ppc64le
• dcopydpn - x86, x86-64, aarch64, ppc64le
• int dump[4]; dump[i*2] := 3;
• dump[0] := 1
• dispatch opcode optimization for 1,2,3,4
• createnr - optimization for 1,2,4,8
• ppc64 optimization : use store multiple for alloc / open opcodes - for filling the stack with zeros
• aarch64 : use - ldr x17, [x16, #PLT_GOT+0x10] to load value without using another operation
• aarch64 : use adrp to load some 32bit addresses
• openheader / open - there should be an optimized version for n < 4 - with reserved raw stack
• for ppc64 / arm64 - there should be alternative version to deal with args more than 9 / 12 / 16 bits
• openin / extopenin - increase number of optimization - open 4,0; open 1,n; open 2,n; open 3,n; open 4,n; open 5,n; open 6,n
• fillir for i = 16
• copy 16
• new opcode - loadfp / lloadfp
• no need to have imul eax,eax,1