Re: [RFC PATCH 2/8] riscv: Generate payload scripts

[LIU Zhiwei]

On 2020/5/12 1:40, Richard Henderson wrote:
> On 4/30/20 12:21 AM, LIU Zhiwei wrote:
> > +    # sequence of li rd, 0x1234567887654321
> > +    #
> > +    #  0:   002471b7                lui     rd,0x247
> > +    #  4:   8ad1819b                addiw   rd,rd,-1875
> > +    #  8:   00c19193                slli    rd,rd,0xc
> > +    #  c:   f1118193                addi    rd,rd,-239 # 0x246f11
> > +    # 10:   00d19193                slli    rd,rd,0xd
> > +    # 14:   d9518193                addi    rd,rd,-619
> > +    # 18:   00e19193                slli    rd,rd,0xe
> > +    # 1c:   32118193                addi    rd,rd,801
> You don't really need to use addiw.  Removing that special case would really
> simplify this.
I think I don't get it. Do you mean that the immediate will not be 64 bit?
> > +sub write_memblock_setup()
> > +{
> > +    # Write code which sets up the memory block for loads and stores.
> > +    # We set r0 to point to a block of 16K length
> > +    # of random data, aligned to the maximum desired alignment.
> > +
> > +    my $align = $MAXALIGN;
> > +    my $datalen = 16384 + $align;
> risu.h:#define MEMBLOCKLEN 8192
>
> Why are you using 16384?
It's a bug.

Once I thought I should make it bigger to support  vector in the future .
And  even that, 8K byts is also enough, as the most bytes operates in 
one instruction
is LMUL * RV_VLEN_MAX  / 8 = 512 Bytes.

Zhiwei
> Also, typo -- you're setting r10 not r0, obviously.
>
> The rest looks fine.
>
>
> r~