riscv: Format Rd of FMV.W.X with NoN-boxing

[PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Ian Jiang 4 years, 3 months ago

For FMV.W.X that moves the lower 32 bits of an integer register to a
floating-point register, Rd should encoded with NoN-boxing scheme.
Note: This applies to RV64 only.

Signed-off-by: Ian Jiang <ianjiang.ict@gmail.com>
---
 target/riscv/insn_trans/trans_rvf.inc.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/target/riscv/insn_trans/trans_rvf.inc.c b/target/riscv/insn_trans/trans_rvf.inc.c
index 172dbfa919..62b7a36567 100644
--- a/target/riscv/insn_trans/trans_rvf.inc.c
+++ b/target/riscv/insn_trans/trans_rvf.inc.c
@@ -368,6 +368,7 @@ static bool trans_fmv_w_x(DisasContext *ctx, arg_fmv_w_x *a)
 
 #if defined(TARGET_RISCV64)
     tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
+    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
 #else
     tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
 #endif
-- 
2.17.1

Re: [PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Richard Henderson 4 years, 3 months ago

On 1/21/20 9:24 PM, Ian Jiang wrote:
> For FMV.W.X that moves the lower 32 bits of an integer register to a
> floating-point register, Rd should encoded with NoN-boxing scheme.
> Note: This applies to RV64 only.
> 
> Signed-off-by: Ian Jiang <ianjiang.ict@gmail.com>
> ---
>  target/riscv/insn_trans/trans_rvf.inc.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/target/riscv/insn_trans/trans_rvf.inc.c b/target/riscv/insn_trans/trans_rvf.inc.c
> index 172dbfa919..62b7a36567 100644
> --- a/target/riscv/insn_trans/trans_rvf.inc.c
> +++ b/target/riscv/insn_trans/trans_rvf.inc.c
> @@ -368,6 +368,7 @@ static bool trans_fmv_w_x(DisasContext *ctx, arg_fmv_w_x *a)
>  
>  #if defined(TARGET_RISCV64)
>      tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
> +    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
>  #else
>      tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
>  #endif
> 

This doesn't look right.  There's nothing in the spec that says the nan-boxing
is restricted to rv64.  NaN-boxing is all about FLEN, not XLEN.


r~

Re: [PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Ian Jiang 4 years, 3 months ago

--
Ian Jiang

Richard Henderson <richard.henderson@linaro.org> 于2020年1月22日周三 下午4:53写道：
>
> On 1/21/20 9:24 PM, Ian Jiang wrote:
> > For FMV.W.X that moves the lower 32 bits of an integer register to a
> > floating-point register, Rd should encoded with NoN-boxing scheme.
> > Note: This applies to RV64 only.
> >
> > Signed-off-by: Ian Jiang <ianjiang.ict@gmail.com>
> > ---
> >  target/riscv/insn_trans/trans_rvf.inc.c | 1 +
> >  1 file changed, 1 insertion(+)
> >
> > diff --git a/target/riscv/insn_trans/trans_rvf.inc.c b/target/riscv/insn_trans/trans_rvf.inc.c
> > index 172dbfa919..62b7a36567 100644
> > --- a/target/riscv/insn_trans/trans_rvf.inc.c
> > +++ b/target/riscv/insn_trans/trans_rvf.inc.c
> > @@ -368,6 +368,7 @@ static bool trans_fmv_w_x(DisasContext *ctx, arg_fmv_w_x *a)
> >
> >  #if defined(TARGET_RISCV64)
> >      tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
> > +    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
> >  #else
> >      tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
> >  #endif
> >
>
> This doesn't look right.  There's nothing in the spec that says the nan-boxing
> is restricted to rv64.  NaN-boxing is all about FLEN, not XLEN.
>

Why the translation of FLW has a NaN-boxing?

file ./target/riscv/insn_trans/trans_rvf.inc.c
 26 static bool trans_flw(DisasContext *ctx, arg_flw *a)
 27 {
...
 34     tcg_gen_qemu_ld_i64(cpu_fpr[a->rd], t0, ctx->mem_idx, MO_TEUL);
 35     /* RISC-V requires NaN-boxing of narrower width floating point values */
 36     tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
...
 41 }

Re: [PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Richard Henderson 4 years, 3 months ago

On 1/21/20 11:53 PM, Ian Jiang wrote:
> --
> Ian Jiang
> 
> Richard Henderson <richard.henderson@linaro.org> 于2020年1月22日周三 下午4:53写道：
>>
>> On 1/21/20 9:24 PM, Ian Jiang wrote:
>>> For FMV.W.X that moves the lower 32 bits of an integer register to a
>>> floating-point register, Rd should encoded with NoN-boxing scheme.
>>> Note: This applies to RV64 only.
>>>
>>> Signed-off-by: Ian Jiang <ianjiang.ict@gmail.com>
>>> ---
>>>  target/riscv/insn_trans/trans_rvf.inc.c | 1 +
>>>  1 file changed, 1 insertion(+)
>>>
>>> diff --git a/target/riscv/insn_trans/trans_rvf.inc.c b/target/riscv/insn_trans/trans_rvf.inc.c
>>> index 172dbfa919..62b7a36567 100644
>>> --- a/target/riscv/insn_trans/trans_rvf.inc.c
>>> +++ b/target/riscv/insn_trans/trans_rvf.inc.c
>>> @@ -368,6 +368,7 @@ static bool trans_fmv_w_x(DisasContext *ctx, arg_fmv_w_x *a)
>>>
>>>  #if defined(TARGET_RISCV64)
>>>      tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
>>> +    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
>>>  #else
>>>      tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
>>>  #endif
>>>
>>
>> This doesn't look right.  There's nothing in the spec that says the nan-boxing
>> is restricted to rv64.  NaN-boxing is all about FLEN, not XLEN.
>>
> 
> Why the translation of FLW has a NaN-boxing?
> 
> file ./target/riscv/insn_trans/trans_rvf.inc.c
>  26 static bool trans_flw(DisasContext *ctx, arg_flw *a)
>  27 {
> ...
>  34     tcg_gen_qemu_ld_i64(cpu_fpr[a->rd], t0, ctx->mem_idx, MO_TEUL);
>  35     /* RISC-V requires NaN-boxing of narrower width floating point values */
>  36     tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
> ...
>  41 }

Eh?  Obviously because we're loading a 32-bit value into a FLEN=64 bit register.

(Except when we've selected a cpu without RVD, I suppose, but in that case this
nan-boxing is both harmless and invisible.)


r~

Re: [PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Ian Jiang 4 years, 3 months ago

Richard Henderson <richard.henderson@linaro.org> 于2020年1月23日周四 上午1:11写道：
>
> On 1/21/20 11:53 PM, Ian Jiang wrote:
> > --
> > Ian Jiang
> >
> > Richard Henderson <richard.henderson@linaro.org> 于2020年1月22日周三 下午4:53写道：
> >>
> >> On 1/21/20 9:24 PM, Ian Jiang wrote:
> >>> For FMV.W.X that moves the lower 32 bits of an integer register to a
> >>> floating-point register, Rd should encoded with NoN-boxing scheme.
> >>> Note: This applies to RV64 only.
> >>>
> >>> Signed-off-by: Ian Jiang <ianjiang.ict@gmail.com>
> >>> ---
> >>>  target/riscv/insn_trans/trans_rvf.inc.c | 1 +
> >>>  1 file changed, 1 insertion(+)
> >>>
> >>> diff --git a/target/riscv/insn_trans/trans_rvf.inc.c b/target/riscv/insn_trans/trans_rvf.inc.c
> >>> index 172dbfa919..62b7a36567 100644
> >>> --- a/target/riscv/insn_trans/trans_rvf.inc.c
> >>> +++ b/target/riscv/insn_trans/trans_rvf.inc.c
> >>> @@ -368,6 +368,7 @@ static bool trans_fmv_w_x(DisasContext *ctx, arg_fmv_w_x *a)
> >>>
> >>>  #if defined(TARGET_RISCV64)
> >>>      tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
> >>> +    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
> >>>  #else
> >>>      tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
> >>>  #endif
> >>>
> >>
> >> This doesn't look right.  There's nothing in the spec that says the nan-boxing
> >> is restricted to rv64.  NaN-boxing is all about FLEN, not XLEN.
> >>
> >
> > Why the translation of FLW has a NaN-boxing?
> >
> > file ./target/riscv/insn_trans/trans_rvf.inc.c
> >  26 static bool trans_flw(DisasContext *ctx, arg_flw *a)
> >  27 {
> > ...
> >  34     tcg_gen_qemu_ld_i64(cpu_fpr[a->rd], t0, ctx->mem_idx, MO_TEUL);
> >  35     /* RISC-V requires NaN-boxing of narrower width floating point values */
> >  36     tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
> > ...
> >  41 }
>
> Eh?  Obviously because we're loading a 32-bit value into a FLEN=64 bit register.
FMV.W.X is the same with FLW at this point that filling a 64 bits
float register based on 32 bits value.
Besides, the RISCV simulator Spike makes NaN-boxing for FLW, FMV.W.W,
FADD.S, FSUB.S, and others.
It might be better that they have a coincident behavior? I am not sure
about this and just want a discussion.

> (Except when we've selected a cpu without RVD, I suppose, but in that case this
> nan-boxing is both harmless and invisible.)
>
>
> r~

Re: [PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Richard Henderson 4 years, 3 months ago

On 1/22/20 2:48 PM, Ian Jiang wrote:
> Richard Henderson <richard.henderson@linaro.org> 于2020年1月23日周四 上午1:11写道：
>>
>> On 1/21/20 11:53 PM, Ian Jiang wrote:
>>> --
>>> Ian Jiang
>>>
>>> Richard Henderson <richard.henderson@linaro.org> 于2020年1月22日周三 下午4:53写道：
>>>>
>>>> On 1/21/20 9:24 PM, Ian Jiang wrote:
>>>>> For FMV.W.X that moves the lower 32 bits of an integer register to a
>>>>> floating-point register, Rd should encoded with NoN-boxing scheme.
>>>>> Note: This applies to RV64 only.
>>>>>
>>>>> Signed-off-by: Ian Jiang <ianjiang.ict@gmail.com>
>>>>> ---
>>>>>  target/riscv/insn_trans/trans_rvf.inc.c | 1 +
>>>>>  1 file changed, 1 insertion(+)
>>>>>
>>>>> diff --git a/target/riscv/insn_trans/trans_rvf.inc.c b/target/riscv/insn_trans/trans_rvf.inc.c
>>>>> index 172dbfa919..62b7a36567 100644
>>>>> --- a/target/riscv/insn_trans/trans_rvf.inc.c
>>>>> +++ b/target/riscv/insn_trans/trans_rvf.inc.c
>>>>> @@ -368,6 +368,7 @@ static bool trans_fmv_w_x(DisasContext *ctx, arg_fmv_w_x *a)
>>>>>
>>>>>  #if defined(TARGET_RISCV64)
>>>>>      tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
>>>>> +    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
>>>>>  #else
>>>>>      tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
>>>>>  #endif
>>>>>
>>>>
>>>> This doesn't look right.  There's nothing in the spec that says the nan-boxing
>>>> is restricted to rv64.  NaN-boxing is all about FLEN, not XLEN.
>>>>
>>>
>>> Why the translation of FLW has a NaN-boxing?
>>>
>>> file ./target/riscv/insn_trans/trans_rvf.inc.c
>>>  26 static bool trans_flw(DisasContext *ctx, arg_flw *a)
>>>  27 {
>>> ...
>>>  34     tcg_gen_qemu_ld_i64(cpu_fpr[a->rd], t0, ctx->mem_idx, MO_TEUL);
>>>  35     /* RISC-V requires NaN-boxing of narrower width floating point values */
>>>  36     tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
>>> ...
>>>  41 }
>>
>> Eh?  Obviously because we're loading a 32-bit value into a FLEN=64 bit register.
> FMV.W.X is the same with FLW at this point that filling a 64 bits
> float register based on 32 bits value.
> Besides, the RISCV simulator Spike makes NaN-boxing for FLW, FMV.W.W,
> FADD.S, FSUB.S, and others.
> It might be better that they have a coincident behavior? I am not sure
> about this and just want a discussion.

So, I wonder if we're talking past one another.

In reply to the patch at the beginning of this thread, I was suggesting that
the correct solution is

 #if defined(TARGET_RISCV64)
     tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
 #else
     tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
 #endif
+    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd],
+                    0xffffffff00000000ULL);

However, after the previous follow-up I wonder if it might be better to split
out the nan-boxing to a helper:

/*
 * RISC-V requires NaN-boxing of narrower width floating
 * point values.  This applies when a 32-bit value is
 * assigned to a 64-bit FP register.  Thus this does not
 * apply when the RVD extension is not present.
 */
static void gen_nanbox_fpr(DisasContext *ctx, int regno)
{
    if (has_ext(ctx, RVD)) {
        tcg_gen_ori_i64(cpu_fpr[regno], cpu_fpr[regno],
                        MAKE_64BIT_MASK(32, 32));
    }
}

which could elide this when only RVF is enabled, and thus
the upper 32-bits of the register are inaccessible.

This helper would then be propagated to the existing uses within translate.c
and insn_trans/trans_rvf.inc.c.

As another cleanup, the ifdef above may be replaced with

    tcg_gen_extu_tl_i64(cpu_fpr[a->rd], t0);

which will handle both RISCV64 and RISCV32.


r~

Re: [PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Ian Jiang 4 years, 3 months ago

Richard Henderson <richard.henderson@linaro.org> 于2020年1月23日周四 上午11:07写道：
>
> On 1/22/20 2:48 PM, Ian Jiang wrote:
> > Richard Henderson <richard.henderson@linaro.org> 于2020年1月23日周四 上午1:11写道：
> >>
> >> On 1/21/20 11:53 PM, Ian Jiang wrote:
> >>> --
> >>> Ian Jiang
> >>>
> >>> Richard Henderson <richard.henderson@linaro.org> 于2020年1月22日周三 下午4:53写道：
> >>>>
> >>>> On 1/21/20 9:24 PM, Ian Jiang wrote:
> >>>>> For FMV.W.X that moves the lower 32 bits of an integer register to a
> >>>>> floating-point register, Rd should encoded with NoN-boxing scheme.
> >>>>> Note: This applies to RV64 only.
> >>>>>
> >>>>> Signed-off-by: Ian Jiang <ianjiang.ict@gmail.com>
> >>>>> ---
> >>>>>  target/riscv/insn_trans/trans_rvf.inc.c | 1 +
> >>>>>  1 file changed, 1 insertion(+)
> >>>>>
> >>>>> diff --git a/target/riscv/insn_trans/trans_rvf.inc.c b/target/riscv/insn_trans/trans_rvf.inc.c
> >>>>> index 172dbfa919..62b7a36567 100644
> >>>>> --- a/target/riscv/insn_trans/trans_rvf.inc.c
> >>>>> +++ b/target/riscv/insn_trans/trans_rvf.inc.c
> >>>>> @@ -368,6 +368,7 @@ static bool trans_fmv_w_x(DisasContext *ctx, arg_fmv_w_x *a)
> >>>>>
> >>>>>  #if defined(TARGET_RISCV64)
> >>>>>      tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
> >>>>> +    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
> >>>>>  #else
> >>>>>      tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
> >>>>>  #endif
> >>>>>
> >>>>
> >>>> This doesn't look right.  There's nothing in the spec that says the nan-boxing
> >>>> is restricted to rv64.  NaN-boxing is all about FLEN, not XLEN.
> >>>>
> >>>
> >>> Why the translation of FLW has a NaN-boxing?
> >>>
> >>> file ./target/riscv/insn_trans/trans_rvf.inc.c
> >>>  26 static bool trans_flw(DisasContext *ctx, arg_flw *a)
> >>>  27 {
> >>> ...
> >>>  34     tcg_gen_qemu_ld_i64(cpu_fpr[a->rd], t0, ctx->mem_idx, MO_TEUL);
> >>>  35     /* RISC-V requires NaN-boxing of narrower width floating point values */
> >>>  36     tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd], 0xffffffff00000000ULL);
> >>> ...
> >>>  41 }
> >>
> >> Eh?  Obviously because we're loading a 32-bit value into a FLEN=64 bit register.
> > FMV.W.X is the same with FLW at this point that filling a 64 bits
> > float register based on 32 bits value.
> > Besides, the RISCV simulator Spike makes NaN-boxing for FLW, FMV.W.W,
> > FADD.S, FSUB.S, and others.
> > It might be better that they have a coincident behavior? I am not sure
> > about this and just want a discussion.
>
> So, I wonder if we're talking past one another.
>
> In reply to the patch at the beginning of this thread, I was suggesting that
> the correct solution is
>
>  #if defined(TARGET_RISCV64)
>      tcg_gen_mov_i64(cpu_fpr[a->rd], t0);
>  #else
>      tcg_gen_extu_i32_i64(cpu_fpr[a->rd], t0);
>  #endif
> +    tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd],
> +                    0xffffffff00000000ULL);
>
> However, after the previous follow-up I wonder if it might be better to split
> out the nan-boxing to a helper:
>
> /*
>  * RISC-V requires NaN-boxing of narrower width floating
>  * point values.  This applies when a 32-bit value is
>  * assigned to a 64-bit FP register.  Thus this does not
>  * apply when the RVD extension is not present.
>  */
> static void gen_nanbox_fpr(DisasContext *ctx, int regno)
> {
>     if (has_ext(ctx, RVD)) {
>         tcg_gen_ori_i64(cpu_fpr[regno], cpu_fpr[regno],
>                         MAKE_64BIT_MASK(32, 32));
>     }
> }
>
> which could elide this when only RVF is enabled, and thus
> the upper 32-bits of the register are inaccessible.
>
> This helper would then be propagated to the existing uses within translate.c
> and insn_trans/trans_rvf.inc.c.
>
> As another cleanup, the ifdef above may be replaced with
>
>     tcg_gen_extu_tl_i64(cpu_fpr[a->rd], t0);
>
> which will handle both RISCV64 and RISCV32.

Good point.
But I am not clear where to call this new helper gen_nanbox_fpr(). Is
there a position that could affect all floating-point instructions? So
that we don't have to modify so many translating functions. Please
give more details.

Re: [PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Richard Henderson 4 years, 3 months ago

On 1/22/20 6:05 PM, Ian Jiang wrote:
> But I am not clear where to call this new helper gen_nanbox_fpr(). Is
> there a position that could affect all floating-point instructions? So
> that we don't have to modify so many translating functions. Please
> give more details.

No, this will have to be called for each instruction individually.

That said, all of the insns that use helper functions, such as fsqrt_s, should
be doing the nan-boxing within the helper function.  Thus you'll want a
different helper function for use within fpu_helper.c.

r~

Re: [PATCH] riscv: Format Rd of FMV.W.X with NoN-boxing

Posted by Ian Jiang 4 years, 3 months ago

Got it.
As the first step, I just summit a new patch:
[PATCH] riscv: Add helper to make NaN-boxing for FP register
I'd like to carry out other fixes after this patch is reviewed.

Richard Henderson <richard.henderson@linaro.org> 于2020年1月24日周五 上午2:53写道：
>
> On 1/22/20 6:05 PM, Ian Jiang wrote:
> > But I am not clear where to call this new helper gen_nanbox_fpr(). Is
> > there a position that could affect all floating-point instructions? So
> > that we don't have to modify so many translating functions. Please
> > give more details.
>
> No, this will have to be called for each instruction individually.
>
> That said, all of the insns that use helper functions, such as fsqrt_s, should
> be doing the nan-boxing within the helper function.  Thus you'll want a
> different helper function for use within fpu_helper.c.
>
>
> r~