Re: new-ra weirdness

[Daniel Berlin <dberlin at dberlin dot org>]

On Tuesday, July 15, 2003, at 7:06 PM, <tm_gccmail@mail.kloo.net> wrote:

>    host: i386-linux
>  target: sh-elf
> version: CVS head today
>
> new-ra is performing unspeakable acts on some of my SH4 testcases.
> It seems to be spilling the FPU registers to integer 
> registers....through
> memory. There is no fast way to move the FPU registers to the integer
> registers on the SH4, so it shouldn't be doing this at all.
>
> Also, it seems to be generating lots of useless loads and stores.
> It loads a DFmode value into two integer pairs, then does nothing
> and stores it back.

As I pointed out to Toshi on IRC, memory_move_cost is not defined on 
SH, and thus, gets the default value of "4".
In contrast, the register move cost on SH ranges from 4 to 20.
The costs are relative, so a memory_move_cost of 4 and a register move 
cost of 20 means the memory move is 5 times cheaper.

So it thinks it's *cheaper* to move this stuff around in memory, then 
it would be to move it around in registers.

Given the numbers right now, I think new-ra is doing the absolute right 
thing here, because it's done the cheapest thing it knows how to do, 
which is move this stuff through memory.

If the MEMORY_MOVE_COST is corrected to be more expensive than the 
REGISTER_MOVE_COST and it *still* does this for no apparent reason (IE 
it moves through memory when it could move through registers), *then* 
something is borked.

> Here's some samples from quatern.i from stress-1.17 compiled with
> -O2 -m4 -fnew-ra:
>
>
> _Normal_Calc_z3:
> ...
>         mov.l   r3,@-r15
>         mov.l   r2,@-r15
> ...
>         fmov.s  @r15+,fr2
>         fmov.s  @r15+,fr3
> ...
>         mov.l   r3,@-r15
>         mov.l   r2,@-r15
> ...
>         fmov.s  @r15+,fr6
>         fmov.s  @r15+,fr7
> ...
>         mov.l   r3,@-r15
>         mov.l   r2,@-r15
>         fmov.s  @r15+,fr14
>         fmov.s  @r15+,fr15
> ...
>         fmov.s  fr7,@-r15
>         fmov.s  fr6,@-r15
>         mov.l   @r15+,r10
>         mov.l   @r15+,r11
> ...
>         fmov.s  fr7,@-r15
>         fmov.s  fr6,@-r15
>         mov.l   @r15+,r2
>         mov.l   @r15+,r3
> ...
>         fmov.s  fr3,@-r15	<- do 16 memory load/stores to execute one
>         fmov.s  fr2,@-r15	<- floating-point multiply
>         mov.l   @r15+,r10
>         mov.l   @r15+,r11
>         mov.l   r11,@-r15
>         mov.l   r10,@-r15
>         fmov.s  @r15+,fr14
>         fmov.s  @r15+,fr15
>         mov.l   r3,@-r15
>         mov.l   r2,@-r15
>         fmov.s  @r15+,fr2
>         fmov.s  @r15+,fr3
>         fmul    dr2,dr14
>         fmov.s  fr15,@-r15
>         fmov.s  fr14,@-r15
>         mov.l   @r15+,r10
>         mov.l   @r15+,r11
> ,,,
>
>         mov.l   @r15+,r2	<- load r2/r3...
>         mov.l   @r15+,r3
>         fmov.s  @r0,fr3
>         mov     #8,r0
>         mov.l   r3,@-r15	<- save back r2/r3 unchanged!!!
>         mov.l   r2,@-r15
> ...
>         mov.l   @r15+,r2	<- load r2/r3...
>         mov.l   @r15+,r3
>         fmov.s  @r0,fr7
>         mov     #24,r0
>         mov.l   r3,@-r15	<- save back r2/r3 unchanged!!!
>         mov.l   r2,@-r15
> ...
>         mov.l   @r15+,r2	<- load r2/r3...
>         mov.l   @r15+,r3
>         mov.l   r3,@-r15	<- save back r2/r3 unchanged!!!
>         mov.l   r2,@-r15
> ...
>         mov.l   @r15+,r10
>         mov.l   @r15+,r11
>         mov.l   r11,@-r15
>         mov.l   r10,@-r15
> ...
>         mov.l   @r15+,r2	<- load r2/r3
>         mov.l   @r15+,r3
>         fmov.s  @r0,fr3
>         mov     #48,r0
>         mov.l   r3,@-r15	<- save back r2/r3 unchanged!!!
>         mov.l   r2,@-r15
> ...
>         mov.l   @r15+,r2
>         mov.l   @r15+,r3
>         fmov.s  @r0,fr3
>         mov     #48,r0
>         mov.l   r3,@-r15
>         mov.l   r2,@-r15
> ...
>         mov.l   @r15+,r2
>         mov.l   @r15+,r3
>         fmov.s  @r0,fr7
>         mov     #72,r0
>         mov.l   r3,@-r15
>         mov.l   r2,@-r15
> ...
>
> Have any idea what's going on?
>
> Toshi