On 4/1/07, David Daney <[EMAIL PROTECTED]> wrote:
I am looking at how the MIPS backend handles division. For the compiler configuration in question (mipsel-linux) division operations trap on division by zero. This is handled in mips_output_division in mips.c where we unconditionally emit a conditional trap.I would like to change it so that if the divisor is not zero, the conditional trap would be omitted. I am looking at this small test program: int divtest(int x, int y) { if (y == 0) return 45; else return x / y; } If I set a breakpoint in mips_output_division, I can print out the operands for the division operation: (gdb) p operands[2] $6 = (rtx) 0x2aaaae155aa0 (gdb) pr (reg/v:SI 5 $5 [orig:196 y ] [196]) Q1: Is it possible to get to the VRP information from this rtx? How? If the VRP information is available, it would be nice to |define_expand for the division/conditional trap and perhaps let the compiler schedule the trap. However the define_expand documentation states that the condition for the define_expand cannot depend on the data in the insn being matched. That would seem to imply that the expansion cannot be dependent on the VRP information. Q2: Is that correct, and why? Q3: Would it be better to do this at the tree level instead of rtl?
At the moment none of this is possible because VRP information is not retained after the VRP pass, nor is it carried over to RTL. The easiest possibility I see is to allow marking *_DIV_EXPR as TREE_THIS_NOTRAP (as we currently can do only for memory references) and translate this to a scheme following the rtl MEM_NOTRAP_P flag. I believe this should be possible right now apart from ensuring the flag doesn't get lost somewhere as they are not documented to be used for divisions. Richard.
