[Bug rtl-optimization/44141] Redundant loads and stores generated for AMD bdver1 target

[venkataramanan.kumar at amd dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44141

--- Comment #15 from Venkataramanan <venkataramanan.kumar at amd dot com> 
2012-04-01 07:55:24 UTC ---
Hi Uros,

I had a look at reload pass.

I have an RTL sequence that look like this.

(insn 32 31 33 2 (set (subreg:V4SF (reg:V2DF 284) 0) <== psuedo reguster 
        (unspec:V4SF [
                (mem/c:V4SF (plus:DI (reg/f:DI 20 frame)
                        (const_int -568 [0xfffffffffffffdc8])) [3
MEM[(real(kind=8)[26] *)&dsroo + 120B]+0 S16 A64])
            ] UNSPEC_MOVU)) test.f90:16 1106 {*sse_movups}
     (nil))

(insn 33 32 371 2 (set (reg:V2DF 285) <= pseudo register 
        (div:V2DF (reg:V2DF 284) <== pseudo register
            (mem/c:V2DF (plus:DI (reg/f:DI 20 frame)
                    (const_int -576 [0xfffffffffffffdc0])) [3
MEM[(real(kind=8)[26] *)&dsroo + 112B]+0 S16 A128]))) test.f90:16 1146
{sse2_divv2df3}
     (nil))

Now reload examines the first RTL. 

(insn 32 31 33 2 (set (subreg:V4SF (reg:V2DF 284) 0) <== psuedo reguster 
        (unspec:V4SF [
                (mem/c:V4SF (plus:DI (reg/f:DI 20 frame)
                        (const_int -568 [0xfffffffffffffdc8])) [3
MEM[(real(kind=8)[26] *)&dsroo + 120B]+0 S16 A64])
            ] UNSPEC_MOVU)) test.f90:16 1106 {*sse_movups}
     (nil))

This did not get an Hard register and so at reload it chooses xmm0 and
generates an output reload as follows.


(insn 393 0 0 (set (reg:V2DF 284)
                                                                             
(reg:V2DF 21 xmm0)) -1

There is possiblity of reload inheritenace and we can avoid input reload for
the next RTL insn 33

(insn 33 32 371 2 (set (reg:V2DF 285)
        (div:V2DF (reg:V2DF 284)
            (mem/c:V2DF (plus:DI (reg/f:DI 20 frame)
                    (const_int -576 [0xfffffffffffffdc0])) [3
MEM[(real(kind=8)[26] *)&dsroo + 112B]+0 S16 A128]))) test.f90:16 1146
{sse2_divv2df3}
     (nil))

But It is not happening and input reload gets generated again before this RTL
as follows:

(insn 395 0 0 (set (reg:V2DF 21 xmm0)
                                                                               
 (reg:V2DF 284)

Also another outpout reload gets emitted after the insn 33 for its output
reload as  

(insn 394 0 0 (set (reg:V2DF 285)
                                                                             
(reg:V2DF 21 xmm0)) -1. But I am not sure If this computation prevented input
reload inheritence in the insn 33 .

I suspect emit_reload_insns is not preserving output reloads in insn32 for
further inheritence insn33. 

Please povide your opinion.