The code generator gets two new methods, a_mul_reg_reg_pair and g_div_const_reg_reg. The first one is basically 32x32 to 64 bits multiplication (or any other size, with result having twice the size of arguments), which must be implemented for every target. The second one actually does the job, its default implementation taken from powerpc64 and is sufficiently good for all three-address targets.
+ Enabled optimized division for MIPS target, target-specific changes are under 30 lines.
git-svn-id: trunk@27904 -
* Deallocate GP save temp in epilogue to avoid warnings when compiled with -dEXTDEBUG
* g_concatcopy: don't check alignment, this allows single byte or word locations to be copied with 2 instructions. Larger unaligned references are supposed to be handled in g_concatcopy_unaligned instead.
git-svn-id: trunk@25147 -
+ Implemented subset of "spill replace" functionality, replacing moves from/to spilled registers with loads/stores to spill locations. This helps to reduce amount of instructions.
git-svn-id: trunk@24900 -
+ Implemented overflow checking for unsigned 32-bit addition and subtraction.
* Use optimize_op_const instead of custom optimizations.
* Change AND/OR/XOR into ANDI/ORI/XORI if they use immediate operands, and use correct range for these immediate operands, must be 0..65535 unlike -32768..32767 for arithmetic operations.
* Don't treat AND/OR/XOR as macros, no longer necessary.
* Don't treat BEQ/BNE as macros either.
git-svn-id: trunk@24445 -
* TCGMIPS.a_load_reg_reg: generate CPU instructions instead of macros
* TCGMIPS.a_cmp_const_reg_label: load constant using a_load_const_reg instead of LI macro (it may also end up with LI, but tries to optimize when possible).
- removed unused variables.
git-svn-id: trunk@24418 -
- Removed TMIPSProcInfo.needs_frame_pointer, it duplicates pi_needs_stackframe flag.
- Don't save RA for leaf procedures.
* GP is caller-saved register, it does not need saving/restoring as nonvolatile registers. Handle its restore position as a "normal" temp, whose offset is known before secondpass. This allows to handle PIC calls without involving GAS macro processing.
git-svn-id: trunk@23885 -
+ Favor 'weak' parameter in a_call_name of both thlcgmips and TCGMIPS.
* make_simple_ref and a_loadaddr_ref_reg: require input references to be 'raw' (no refaddr=addr_pic, etc) and do not use GP as base/index. If it's not true, raise an internal error. When addr_pic_call16 or so needs to be generated, it must be done without calling the mentioned methods.
* thlcgmips.a_call_name: generate PIC sequence for procedures declared as 'external', instead of 'cdecl', this is more correct because ABI has only one calling convention.
- make_simple_ref_fpu removed, there's no reason to handle references to floating-point data in different way.
- a_loadaddr_ref_cgpara override also removed, generic method does the job just well.
- thlcgmips.a_call_ref and a_call_reg overrides removed because indirect calls are now always done using $t9.
git-svn-id: trunk@23698 -
* Abort if attempting to emit a call without pi_do_call flag set, because the outgoing parameter area determines layout of temps and therefore cannot be created/changed during pass 2.
* Bypass g_concatcopy_move if procedure is a leaf one, due to the restriction mentioned above.
git-svn-id: trunk@23644 -
* Parameters passed in registers are no longer forced to stack, therefore assembler routines using 'lw' instruction to load such parameters must have it changed to 'move' (tb0072.pp).
git-svn-id: trunk@23627 -
* Also changed parameter management to allocate a single LOC_REFERENCE location for any parameter (or part of a parameter) passed on stack. Such locations can now be handled by generic code. As a consequence, TGMIPS.a_load_const_cgpara and a_load_ref_cgpara methods are no longer necessary (moreover they were ignoring parameter shifting).
* a_loadfpu_ref_cgpara patched to handle 64-bit location, which are now possible.
* This change eliminates code generation madness where each word of a record was copied individually, thus passing by value a 32K-word sized record was bloating into 64K instructions, multiply this number by 2 for larger records due to limited immediate offsets, and by another 2 for global record in PIC mode :/
git-svn-id: trunk@23575 -
- removed a_jmp_cond, it's not virtual and not applicable to this CPU.
* a_loadfpu_reg_cgpara: use direct register moves for simple destinations.
* g_concatcopy: don't take address of source/destination for small amounts of data if possible, for single 32-bit moves this reduces number of generated instructions from 4 to 2.
* g_intf_wrapper: 'Self' is in R4 (a0), not R2. Fixes test/tinterface1.pp.
* mips/cpupara.pas: for functions with result returned in parameter, pass its address in a0 only if result is a record. ABI does not specify behavior for types except records/unions. At the same time, Pascal code relies on the fact that results like strings/interfaces does not change their locations, i.e. "function foo(<self>): IInterface" can be invoked as "procedure foo(<self>, out obj)". This fixes test/tdel1.pp and some Variant-related tests.
git-svn-id: trunk@23377 -
that the type of the parameters can be determined automatically
o added compilerproc declarations for all helpers called in the compiler
via their assembler name, so we can look up the corresponding procdef
git-svn-id: trunk@23325 -
