On 11/07/2011 03:28 PM, Walter Lee wrote: > http://gcc.gnu.org/ml/gcc-patches/2011-10/msg02084.html
TileGx: Almost all the same comments as for TilePro. I see that you've started some macroization of the port, but havn't completed it. E.g. sub<mode>3 vs addsi3+adddi3. > (define_expand "addsf3" Clever. I do wonder if it wouldn't be better to wait until after reload to split these though. > (define_insn "*and_0x00000000FFFFFFFF" > (define_insn "*and_0xFFFFFFFF00000000" These probably should be merged into the main AND pattern. But failing that, they're only applicable to DImode and needn't be macroized. > (define_expand "clzsi2" > [(set (match_dup 2) > (ashift:DI (match_operand:SI 1 "reg_or_0_operand" "") > (const_int 32))) > (set (subreg:DI (match_operand:SI 0 "register_operand" "") 0) > (clz:DI (match_dup 2)))] > "" Generic code can handle this. Likewise ctz, popcount, parity. > (define_insn "insn_cmpexch<four_if_si>" > (define_insn "insn_exch<four_if_si>" > (define_insn "insn_fetchadd<four_if_si>" > (define_insn "insn_fetchand<four_if_si>" > (define_insn "insn_fetchor<four_if_si>" Use the atomic_ names and interfaces. > #define LIBCALL_VALUE(MODE) gen_rtx_REG (MODE, 0) > #define FUNCTION_VALUE_REGNO_P(N) ((N) < TILEGX_NUM_RETURN_REGS) > #define FUNCTION_ARG_REGNO_P(N) ((N) < TILEGX_NUM_ARG_REGS) > #define GO_IF_LEGITIMATE_ADDRESS(MODE, X, ADDR) \ > #define GO_IF_MODE_DEPENDENT_ADDRESS(ADDR, LABEL) \ All of these are now target hooks. There may be more. > /* We represent all SI values as sign-extended DI values in > registers. */ > #define TRULY_NOOP_TRUNCATION(OUTPREC, INPREC) \ > ((INPREC) <= 32 || (OUTPREC) > 32) I think you should be *very* careful before you insist on this. Do you not have a memory mode that ignores the high bits for 32-bit pointers? From the description of "memoryReadWord", it does seem like you've got modes that pre-process input addresses. MIPS uses this because, techinically, the 32-bit operation insns produce undefined results when given inputs that are not sign-extended. I see no such restriction in the TileGx manual. At least in 64-bit pointer mode, I think you should drop this and make sure that you've got appropriate sign_extend patterns for all of the "x" insns. Similar to how x86_64 does for the zero_extend patterns. r~
