On Fri, Jan 27, 2017 at 12:11:05PM -0600, Aaron Sawdey wrote:
> The updated memcmp-1 testcase passes on ppc64le (p8/p9), ppc64 (p8/p9),
> ppc32 (p8), and x86_64. Bootstrap was successful on ppc64/ppc64le.
> Assuming regtest on ppc64/ppc64le passes, ok for trunk?
> + ldbrx 10,6,9
> + ldbrx 9,7,9
> + subf. 9,9,10
> + bne 0,.L8
subfc. 9,9,10
> + addi 9,4,7
> + lwbrx 10,0,9
> + addi 9,5,7
> + lwbrx 9,0,9
It would be nice if this was
li 9,7
lwbrx 10,9,4
lwbrx 9,9,5
but that is a generic problem I bet.
> + subfc 9,9,10
> + b .L9
> .L8: # convert_label
> - cntlzd 9,9
> - addi 9,9,-1
> - xori 9,9,0x3f
> + subfe 10,10,10
> + popcntd 9,9
> + rldimi 9,10,6,0
> .L9: # final_label
The code does not generate rldimi anymore, always just "or".
> + while maintaining <0 / ==0 / >0 properties. This sequence works:
> + subfc L,A,B
> + subfe H,H,H
> + popcntd L,L
> + rldimi L,H,6,0
"or" here, as well.
> --- gcc/config/rs6000/rs6000.md (revision 244952)
> +++ gcc/config/rs6000/rs6000.md (working copy)
> @@ -2068,6 +2068,35 @@
> "subfic %0,%1,%2"
> [(set_attr "type" "add")])
>
> +(define_insn_and_split "subf<mode>3_carry_dot2"
> + [(set (match_operand:CC 3 "cc_reg_operand" "=x,?y")
> + (compare:CC (minus:P (match_operand:P 2 "gpc_reg_operand" "r,r")
> + (match_operand:P 1 "gpc_reg_operand" "r,r"))
> + (const_int 0)))
> + (set (match_operand:P 0 "gpc_reg_operand" "=r,r")
> + (minus:P (match_dup 2)
> + (match_dup 1)))
> + (set (reg:P CA_REGNO)
> + (leu:P (match_dup 1)
> + (match_dup 2)))]
> + "<MODE>mode == Pmode"
> + "@
> + subfc. %0,%1,%2
> + #"
> + "&& reload_completed && cc_reg_not_cr0_operand (operands[3], CCmode)"
So far so good...
> + [(set (reg:P CA_REGNO)
> + (leu:P (match_dup 1)
> + (match_dup 2)))
> + (set (match_dup 0)
> + (minus:P (match_dup 2)
> + (match_dup 1)))
> + (set (match_dup 3)
> + (compare:CC (match_dup 0)
> + (const_int 0)))]
This needs a "parallel" around the two pieces that together are the subfc
instruction you split to. They also need to be in the correct order. So:
[(parallel [(set (match_dup 0)
(minus:P (match_dup 2)
(match_dup 1)))
(set (reg:P CA_REGNO)
(leu:P (match_dup 1)
(match_dup 2)))])
(set (match_dup 3)
(compare:CC (match_dup 0)
(const_int 0)))]
The rest looks good. With those fixes the patch is approved for trunk
(if all testing works out ;-) )
Thanks,
Segher
