1 Notes on the Mips target (3/4/2012) 2 ----------------------------------- 3 4Testing 5 6The initial implementation of Mips support in the compiler is untested on 7actual hardware, and as such should be expected to have many bugs. However, 8the vast majority of code for Mips support is either shared with other 9tested targets, or was taken from the functional Mips JIT compiler. The 10expectation is that when it is first tried out on actual hardware lots of 11small bugs will be flushed out, but it should not take long to get it 12solidly running. The following areas are considered most likely to have 13problems that need to be addressed: 14 15 o Endianness. Focus was on little-endian support, and if a big-endian 16 target is desired, you should pay particular attention to the 17 code generation for switch tables, fill array data, 64-bit 18 data handling and the register usage conventions. 19 20 o The memory model. Verify that GenMemoryBarrier() generates the 21 appropriate flavor of sync. 22 23Register promotion 24 25The resource masks in the LIR structure are 64-bits wide, which is enough 26room to fully describe def/use info for Arm and x86 instructions. However, 27the larger number of MIPS core and float registers render this too small. 28Currently, the workaround for this limitation is to avoid using floating 29point registers 16-31. These are the callee-save registers, which therefore 30means that no floating point promotion is allowed. Among the solution are: 31 o Expand the def/use mask (which, unfortunately, is a significant change) 32 o The Arm target uses 52 of the 64 bits, so we could support float 33 registers 16-27 without much effort. 34 o We could likely assign the 4 non-register bits (kDalvikReg, kLiteral, 35 kHeapRef & kMustNotAlias) to positions occuped by MIPS registers that 36 don't need def/use bits because they are never modified by code 37 subject to scheduling: r_K0, r_K1, r_SP, r_ZERO, r_S1 (rSELF). 38 39Branch delay slots 40 41Little to no attempt was made to fill branch delay slots. Branch 42instructions in the encoding map are given a length of 8 bytes to include 43an implicit NOP. It should not be too difficult to provide a slot-filling 44pass following successful assembly, but thought should be given to the 45design. Branches are currently treated as scheduling barriers. One 46simple solution would be to copy the instruction at branch targets to the 47slot and adjust the displacement. However, given that code expansion is 48already a problem it would be preferable to use a more sophisticated 49scheduling solution. 50 51Code expansion 52 53Code expansion for the MIPS target is significantly higher than we see 54for Arm and x86. It might make sense to replace the inline code generation 55for some of the more verbose Dalik byte codes with subroutine calls to 56shared helper functions. 57 58