Re: decrement & branch isns

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello,

> The next bit of this port (on the 3.4 branch) I've got stuck on is
> recognizing decrement & branch instruction possibilities.  The actual
> instruction I have to play with looks like this bit of rtl,
> 
>   [(set (pc)
> 	(if_then_else (ne:SI (match_operand:SI 0 "register_operand" "+r,m")
> 			     (const_int 0))
> 	  	      (label_ref (match_operand 1 "" "")) (pc)))
>    (set (match_dup 0) (plus:SI (match_dup 0) (const_int -1)))]
> 
> I.e it does the test on the old value of the loop variable, and always
> decrements it in parallel. Hence on exit, the loop variable will be -1.
> This differs from the m68k decrement_and_branch_until_zero pattern, which
> tests the new value of the loop variable.
> 
> I've had some success with using a doloop_end expander, but that
> only allows single exit loops. If I compile the following test case
>   unsigned Baz (unsigned j, int *ptr)
>   {
>     unsigned ix;
> 
>     for (ix = j; ix--;)
>       if (!ptr[ix])
>         return ix;
>     return -1u;
>   }
> and then make the multiple-exit and contains calls tests in doloop.c
> conditional on HAVE_doloop_begin (which I presume will always be defined
> on that kind of hardware with really hardware iteration variables), then
> this too is optimized, but doesn't recogize that the loop control variable
> and ix are decremented in lock step.  What is generated is,
> 
>         mov       reg5,reg4  # reg4 contains ix
> .L8:
>         load      reg6,reg2  # reg2 contains (biased) ptr
>         add       reg2,reg2,4 # decrement ptr
>         branchz   reg6,.L10
>         sub       reg4,reg4,1  # here is the unneeded decrement of ix
>         branchdec reg5,.L8  # and here is the branch and dec of reg5
> 
> I also had problems with the loop unroller generating incorrect
> code. I had some success with the attached patch, which as you can
> see adjusts the delta based on whether the end-loop test biases the
> loop variable's value.  However, this wasn't successful in all cases
> but I've not investigated those further (the remaining failure
> looks like bad liveness analysis).

the patch below definitely is not the right way how to solve the problem;
you should need no changes to unrolling.  The problem might be either
in the parts that determine the number of iterations, or perhaps
doloop produces a wrong pattern (given that it does not know this type
of doloop pattern, it is well possible).

If you send me the patch, I will have a look at the problem.

Zdenek

> nathan
> 
> -- 
> Nathan Sidwell    ::   http://www.codesourcery.com   ::     CodeSourcery LLC
> nathan@codesourcery.com    ::     http://www.planetfall.pwp.blueyonder.co.uk
> 

> Index: loop-unroll.c
> ===================================================================
> RCS file: /home/icera/Repository/gnu/gcc/loop-unroll.c,v
> retrieving revision 1.1.1.1
> diff -c -3 -p -r1.1.1.1 loop-unroll.c
> *** loop-unroll.c	29 Apr 2004 11:47:30 -0000	1.1.1.1
> --- loop-unroll.c	9 Jun 2004 11:50:17 -0000
> *************** Software Foundation, 59 Temple Place - S
> *** 30,35 ****
> --- 30,36 ----
>   #include "params.h"
>   #include "output.h"
>   #include "expr.h"
> + #include "loop.h"
>   /* We need to use the macro exact_log2. */
>   #include "toplev.h"
>   
> *************** unroll_loop_runtime_iterations (struct l
> *** 871,876 ****
> --- 872,879 ----
>       { 
>         rtx count, count2, count_unroll_mod;
>         int count_unroll;
> +       int incr = 1;
> +       rtx cond;
>   
>         /* start_sequence (); */
>                     
> *************** unroll_loop_runtime_iterations (struct l
> *** 885,892 ****
>   				  count, count_unroll_mod,
>   				  0, 0, OPTAB_LIB_WIDEN);
>   
>         count2 = expand_simple_binop (GET_MODE (desc->var), PLUS,
> ! 				     count, GEN_INT (2),
>   				     0, 0, OPTAB_LIB_WIDEN);
>        
>         emit_move_insn (desc->var, count2);
> --- 888,900 ----
>   				  count, count_unroll_mod,
>   				  0, 0, OPTAB_LIB_WIDEN);
>   
> +       cond = doloop_condition_get (PATTERN (BB_END (desc->out_edge->src)));
> +       cond = XEXP (cond, 0);
> +       if (GET_CODE (cond) == PLUS)
> + 	incr -= INTVAL (XEXP (cond, 1));
> + 
>         count2 = expand_simple_binop (GET_MODE (desc->var), PLUS,
> ! 				     count, GEN_INT (incr),
>   				     0, 0, OPTAB_LIB_WIDEN);
>        
>         emit_move_insn (desc->var, count2);