Re: [Qemu-devel] [PATCH v2] Add AVX, AVX-512, MPX support to x86_cpu_dump_state

[Doug Gale]

On Thu, Dec 7, 2017 at 5:37 PM, Richard Henderson <
address@hidden> wrote:

> On 12/02/2017 10:35 PM, Doug Gale wrote:
> > Signed-off-by: Doug Gale <address@hidden>
> > ---
> > Fix MSB LSB showing when SSE is disabled
> >  target/i386/helper.c | 95 ++++++++++++++++++++++++++++++
> +++++++++++++++-------
> >  1 file changed, 83 insertions(+), 12 deletions(-)
> >
> > diff --git a/target/i386/helper.c b/target/i386/helper.c
> > index f63eb3d3f4..03812b6e87 100644
> > --- a/target/i386/helper.c
> > +++ b/target/i386/helper.c
> > @@ -543,6 +543,7 @@ void x86_cpu_dump_state(CPUState *cs, FILE *f,
> fprintf_function cpu_fprintf,
> >          }
> >      }
> >      cpu_fprintf(f, "EFER=%016" PRIx64 "\n", env->efer);
> > +    cpu_fprintf(f, "XCR0=%016" PRIx64 "\n", env->xcr0);
> >      if (flags & CPU_DUMP_FPU) {
> >          int fptag;
> >          fptag = 0;
> > @@ -565,21 +566,91 @@ void x86_cpu_dump_state(CPUState *cs, FILE *f,
> fprintf_function cpu_fprintf,
> >              else
> >                  cpu_fprintf(f, " ");
> >          }
> > -        if (env->hflags & HF_CS64_MASK)
> > -            nb = 16;
> > -        else
> > +
> > +        if (env->hflags & HF_CS64_MASK) {
> > +            if (env->xcr0 & XSTATE_Hi16_ZMM_MASK) {
> > +                /* AVX-512 32 registers enabled */
> > +                nb = 32;
> > +            } else {
> > +                /* 64-bit mode, 16 registers */
> > +                nb = 16;
> > +            }
> > +        } else {
> > +            /* 32 bit mode, 8 registers */
> >              nb = 8;
> > -        for(i=0;i<nb;i++) {
> > -            cpu_fprintf(f, "XMM%02d=%08x%08x%08x%08x",
> > -                        i,
> > -                        env->xmm_regs[i].ZMM_L(3),
> > -                        env->xmm_regs[i].ZMM_L(2),
> > -                        env->xmm_regs[i].ZMM_L(1),
> > -                        env->xmm_regs[i].ZMM_L(0));
> > -            if ((i & 1) == 1)
> > +        }
> > +
> > +        /* sse register width in units of 64 bits */
> > +        int zmm_width;
> > +        char zmm_name;
> > +        if (env->xcr0 & XSTATE_ZMM_Hi256_MASK) {
> > +            /* 512-bit "ZMM" - AVX-512 registers enabled */
> > +            zmm_width = 8;
> > +            zmm_name = 'Z';
> > +        } else if (env->xcr0 & XSTATE_YMM_MASK) {
> > +            /* 256-bit "YMM" - AVX enabled */
> > +            zmm_width = 4;
> > +            zmm_name = 'Y';
> > +        } else if (env->cr[4] & CR4_OSFXSR_MASK) {
> > +            /* 128-bit "XMM" - SSE enabled */
> > +            zmm_width = 2;
> > +            zmm_name = 'X';
> > +        } else {
> > +            /* SSE not enabled */
> > +            zmm_width = 0;
> > +            zmm_name = 0;
>
> You should exit the function here ...
>
> > +        }
> > +
> > +        if (zmm_width > 0) {
> > +            cpu_fprintf(f, "      MSB%*sLSB\n",
> > +                        -(zmm_width * 16 + zmm_width - 7), "");
> > +        }
> > +
> > +        for (i = 0; i < nb; i++) {
> > +            if (zmm_width == 0) {
> > +                cpu_fprintf(f, "SSE not enabled\n");
> > +                break;
> > +            }
>
> ... rather than performing this test in a loop.
> Or indeed, printing anything at all.
>
> I assume this is in order to examine registers in the monitor under kvm?
>
>
> r~
>

​
Yes, to examine the registers in the monitor and improve the output of
things like -d int. I ran into an issue with
​ guest​
code and I wanted to check XCR0 to see if AVX was enabled, and I didn't see
any AVX registers
​, which was very misleading​
.
​ I couldn't check XCR0 to see if it was enabled either.​
I investigated and found that AVX was enabled, and the monitor output was
false. I added XCR0 and the entire AVX
​/AVX-512​
and MPX state to the
​qemu ​
output.

Why would it exit the function there? Are you sure the other states I
further down won't be enabled? It prints the MPX state further down. I put
the test in the loop so I wouldn't have to indent that whole `for` loop
another level, to improve readability. Is the performance of that loop
really relevant? The fprintf's will take at least thousands of times more
CPU time than that predictable branch.​