2013/6/4 Aniket Deole <aniket.deole@xxxxxxxxx>: > Hello, > > We are building GCC for the Mips Simulator "SPIM". > We have previously built the same for GCC 4.6.2, and are now trying to > build it for 4.7.2. > > We are experiencing an issue related to the cbranchsi4 in GCC 4.7.2, > machine description wherein, GCC is compiled correctly and cc1 is > generated. But while compiling an external C program having > conditional statements using cc1, the compilation goes in an infinite > loop. > > We were able to rectify this infinite loop by removing an SI from the > instruction for the first match operator clause. > > But we are not sure as to why this fix is required with respect to the > different GCC versions 4.6.2 and 4.7.2. > > --------------------------------------------------------------- > GCC 4.6.2 MD: > -------------------------- > (define_insn "cbranchsi4" > [(set (pc) > (if_then_else > > (match_operator:SI 0 "comparison_operator" > > [(match_operand:SI 1 "register_operand" "") > (match_operand:SI 2 "register_operand" "")]) > (label_ref (match_operand 3 "" "")) > (pc)))] > "" > "* return conditional_insn(GET_CODE(operands[0]),operands,0); " > ) > -------------------------------------------------------------- > GCC 4.7.2 MD: > --------------------------- > (define_insn "cbranchsi4" > [(set (pc) > (if_then_else > > (match_operator 0 "comparison_operator" > > [(match_operand:SI 1 "register_operand" "") > (match_operand:SI 2 "register_operand" "")]) > (label_ref (match_operand 3 "" "")) > (pc)))] > "" > "* return conditional_insn(GET_CODE(operands[0]),operands,0); " > ) > > Thank You, > --------------------------------- > Aniket Deole > aniket.deole@xxxxxxxxx Hi, Aniket, It's hard tell why your cc1 went into infinite loop. It requires tracing gcc code by yourself. However, refer to to GCC Internal, 10.10 Comparison Operations: "If the comparison operation is being tested (e.g., the first operand of an if_then_else), the mode must be VOIDmode." IIUC, the mode of cbranch pattern is determined by the compared operands, not comparison operator. The comparison operator is supposed to be VOIDmode and many optimization/simplification phases are based on such principle. So I would guess it was just lucky that you did not suffer any compilation problems with your 4.6.2 MD design. Best regards, jasonwucj
