qemu-devel.nongnu.org archive mirror
 help / color / mirror / Atom feed
From: Richard Henderson <richard.henderson@linaro.org>
To: "Philippe Mathieu-Daudé" <philmd@linaro.org>, qemu-devel@nongnu.org
Cc: mark.cave-ayland@ilande.co.uk
Subject: Re: [PATCH 05/22] target/sparc: Inline FNEG, FABS
Date: Fri, 2 Feb 2024 14:32:25 +1000	[thread overview]
Message-ID: <0670f7d7-6518-4c72-a3b8-bd38152bcf94@linaro.org> (raw)
In-Reply-To: <d26582c2-88e4-4813-bef6-0c12faca34b5@linaro.org>

On 1/30/24 18:40, Philippe Mathieu-Daudé wrote:
> Hi Richard,
> 
> On 3/11/23 18:38, Richard Henderson wrote:
>> These are simple bit manipulation insns.
>> Begin using i128 for float128.
>> Implement FMOVq with do_qq.
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>>   target/sparc/helper.h     |  6 ----
>>   target/sparc/fop_helper.c | 34 ---------------------
>>   target/sparc/translate.c  | 62 +++++++++++++++++++--------------------
>>   3 files changed, 30 insertions(+), 72 deletions(-)
> 
> 
>> @@ -1239,13 +1235,13 @@ static void gen_op_fmovs(TCGv_i32 dst, TCGv_i32 src)
>>   static void gen_op_fnegs(TCGv_i32 dst, TCGv_i32 src)
>>   {
>>       gen_op_clear_ieee_excp_and_FTT();
>> -    gen_helper_fnegs(dst, src);
>> +    tcg_gen_xori_i32(dst, src, 1u << 31);
>>   }
>>   static void gen_op_fabss(TCGv_i32 dst, TCGv_i32 src)
>>   {
>>       gen_op_clear_ieee_excp_and_FTT();
>> -    gen_helper_fabss(dst, src);
>> +    tcg_gen_andi_i32(dst, src, ~(1u << 31));
>>   }
>>   static void gen_op_fmovd(TCGv_i64 dst, TCGv_i64 src)
>> @@ -1257,13 +1253,33 @@ static void gen_op_fmovd(TCGv_i64 dst, TCGv_i64 src)
>>   static void gen_op_fnegd(TCGv_i64 dst, TCGv_i64 src)
>>   {
>>       gen_op_clear_ieee_excp_and_FTT();
>> -    gen_helper_fnegd(dst, src);
>> +    tcg_gen_xori_i64(dst, src, 1ull << 63);
>>   }
>>   static void gen_op_fabsd(TCGv_i64 dst, TCGv_i64 src)
>>   {
>>       gen_op_clear_ieee_excp_and_FTT();
>> -    gen_helper_fabsd(dst, src);
>> +    tcg_gen_andi_i64(dst, src, ~(1ull << 63));
>> +}
>> +
>> +static void gen_op_fnegq(TCGv_i128 dst, TCGv_i128 src)
>> +{
>> +    TCGv_i64 l = tcg_temp_new_i64();
>> +    TCGv_i64 h = tcg_temp_new_i64();
>> +
>> +    tcg_gen_extr_i128_i64(l, h, src);
>> +    tcg_gen_xori_i64(h, h, 1ull << 63);
>> +    tcg_gen_concat_i64_i128(dst, l, h);
>> +}
>> +
>> +static void gen_op_fabsq(TCGv_i128 dst, TCGv_i128 src)
>> +{
>> +    TCGv_i64 l = tcg_temp_new_i64();
>> +    TCGv_i64 h = tcg_temp_new_i64();
>> +
>> +    tcg_gen_extr_i128_i64(l, h, src);
>> +    tcg_gen_andi_i64(h, h, ~(1ull << 63));
>> +    tcg_gen_concat_i64_i128(dst, l, h);
>>   }
> 
> Why not extract these as generic TCG FPU helpers?

The representation of floating-point registers varies wildly between targets.  Sparc would 
be the only one to (a) have float128 and (b) represent them in TCGv_i128.

Even considering float32, is the representation TCGv_i32 or TCGv_i64?
Should the result be nan-boxed (riscv and loongarch)?

We already provide tcg_gen_xori_i{32,64}, so, really that's enough for any target.

> $ git grep -wE 'float...?_(chs|abs)' target/
> target/arm/tcg/helper-a64.c:214:    a = float16_chs(a);
> target/arm/tcg/helper-a64.c:229:    a = float32_chs(a);
> target/arm/tcg/helper-a64.c:244:    a = float64_chs(a);
> target/arm/tcg/helper-a64.c:259:    a = float16_chs(a);
> target/arm/tcg/helper-a64.c:274:    a = float32_chs(a);
> target/arm/tcg/helper-a64.c:289:    a = float64_chs(a);
> target/arm/tcg/helper-a64.c:632:    float16 f0 = float16_abs(a);
> target/arm/tcg/helper-a64.c:633:    float16 f1 = float16_abs(b);
> target/arm/tcg/helper-a64.c:642:    float16 f0 = float16_abs(a);
> target/arm/tcg/helper-a64.c:643:    float16 f1 = float16_abs(b);
> target/arm/tcg/mve_helper.c:2840:    return float16_abs(float16_sub(a, b, s));
> target/arm/tcg/mve_helper.c:2845:    return float32_abs(float32_sub(a, b, s));
> target/arm/tcg/mve_helper.c:2854:    return float16_maxnum(float16_abs(a), float16_abs(b), 
> s);
> target/arm/tcg/mve_helper.c:2859:    return float32_maxnum(float32_abs(a), float32_abs(b), 
> s);
> target/arm/tcg/mve_helper.c:2864:    return float16_minnum(float16_abs(a), float16_abs(b), 
> s);
> target/arm/tcg/mve_helper.c:2869:    return float32_minnum(float32_abs(a), float32_abs(b), 
> s);
> target/arm/tcg/neon_helper.c:1513:    float32 f0 = float32_abs(make_float32(a));
> target/arm/tcg/neon_helper.c:1514:    float32 f1 = float32_abs(make_float32(b));
> target/arm/tcg/neon_helper.c:1521:    float32 f0 = float32_abs(make_float32(a));
> target/arm/tcg/neon_helper.c:1522:    float32 f1 = float32_abs(make_float32(b));
> target/arm/tcg/neon_helper.c:1529:    float64 f0 = float64_abs(make_float64(a));
> target/arm/tcg/neon_helper.c:1530:    float64 f1 = float64_abs(make_float64(b));
> target/arm/tcg/neon_helper.c:1537:    float64 f0 = float64_abs(make_float64(a));
> target/arm/tcg/neon_helper.c:1538:    float64 f1 = float64_abs(make_float64(b));
> target/arm/tcg/sve_helper.c:4227:DO_REDUCE(sve_fmaxv_h, float16, H1_2, max, 
> float16_chs(float16_infinity))
> target/arm/tcg/sve_helper.c:4228:DO_REDUCE(sve_fmaxv_s, float32, H1_4, max, 
> float32_chs(float32_infinity))
> target/arm/tcg/sve_helper.c:4229:DO_REDUCE(sve_fmaxv_d, float64, H1_8, max, 
> float64_chs(float64_infinity))
> target/arm/tcg/sve_helper.c:4345:    return float16_abs(float16_sub(a, b, s));
> target/arm/tcg/sve_helper.c:4350:    return float32_abs(float32_sub(a, b, s));
> target/arm/tcg/sve_helper.c:4355:    return float64_abs(float64_sub(a, b, s));
> target/arm/tcg/sve_helper.c:4997:            mm = float16_abs(mm);
> target/arm/tcg/sve_helper.c:5019:            mm = float32_abs(mm);
> target/arm/tcg/sve_helper.c:5045:            mm = float64_abs(mm);
> target/arm/tcg/sve_helper.c:5062:    float16 neg_real = float16_chs(neg_imag);
> target/arm/tcg/sve_helper.c:5094:    float32 neg_real = float32_chs(neg_imag);
> target/arm/tcg/sve_helper.c:5126:    float64 neg_real = float64_chs(neg_imag);
> target/arm/tcg/vec_helper.c:996:    return -float16_le(float16_abs(op2), float16_abs(op1), 
> stat);
> target/arm/tcg/vec_helper.c:1001:    return -float32_le(float32_abs(op2), 
> float32_abs(op1), stat);
> target/arm/tcg/vec_helper.c:1006:    return -float16_lt(float16_abs(op2), 
> float16_abs(op1), stat);
> target/arm/tcg/vec_helper.c:1011:    return -float32_lt(float32_abs(op2), 
> float32_abs(op1), stat);
> target/arm/tcg/vec_helper.c:1124:    return float16_abs(float16_sub(op1, op2, stat));
> target/arm/tcg/vec_helper.c:1129:    return float32_abs(float32_sub(op1, op2, stat));
> target/arm/tcg/vec_helper.c:1304:    return float16_muladd(float16_chs(op1), op2, dest, 0, 
> stat);
> target/arm/tcg/vec_helper.c:1310:    return float32_muladd(float32_chs(op1), op2, dest, 0, 
> stat);
> target/arm/vfp_helper.c:286:    return float16_chs(a);
> target/arm/vfp_helper.c:291:    return float32_chs(a);
> target/arm/vfp_helper.c:296:    return float64_chs(a);
> target/arm/vfp_helper.c:301:    return float16_abs(a);
> target/arm/vfp_helper.c:306:    return float32_abs(a);
> target/arm/vfp_helper.c:311:    return float64_abs(a);
> target/arm/vfp_helper.c:688:    } else if (float16_abs(f16) < (1 << 8)) {
> target/arm/vfp_helper.c:738:    } else if (float32_abs(f32) < (1ULL << 21)) {
> target/arm/vfp_helper.c:1133:    if (value == float64_chs(float64_zero)) {
> target/i386/tcg/fpu_helper.c:591:    ST0 = floatx80_chs(ST0);
> target/i386/tcg/fpu_helper.c:596:    ST0 = floatx80_abs(ST0);
> target/i386/tcg/fpu_helper.c:781:        tmp = floatx80_chs(tmp);
> target/i386/tcg/fpu_helper.c:1739:        ST0 = floatx80_div(floatx80_chs(floatx80_one), 
> floatx80_zero,
> target/i386/tcg/fpu_helper.c:2104:            ST1 = floatx80_chs(ST1);
> target/i386/tcg/fpu_helper.c:2119:            ST1 = floatx80_chs(ST0);
> target/i386/tcg/fpu_helper.c:2135:            ST1 = floatx80_chs(ST1);
> target/i386/tcg/fpu_helper.c:2140:            ST1 = floatx80_chs(floatx80_zero);
> target/i386/tcg/fpu_helper.c:2276: floatx80_chs(floatx80_zero) :
> target/i386/tcg/fpu_helper.c:2285: floatx80_chs(floatx80_infinity) :
> target/m68k/fpu_helper.c:212:    res->d = floatx80_round(floatx80_abs(val->d), 
> &env->fp_status);
> target/m68k/fpu_helper.c:218:    res->d = floatx80_round(floatx80_abs(val->d), 
> &env->fp_status);
> target/m68k/fpu_helper.c:225:    res->d = floatx80_round(floatx80_abs(val->d), 
> &env->fp_status);
> target/m68k/fpu_helper.c:231:    res->d = floatx80_round(floatx80_chs(val->d), 
> &env->fp_status);
> target/m68k/fpu_helper.c:237:    res->d = floatx80_round(floatx80_chs(val->d), 
> &env->fp_status);
> target/m68k/fpu_helper.c:244:    res->d = floatx80_round(floatx80_chs(val->d), 
> &env->fp_status);
> target/m68k/fpu_helper.c:557:        quotient = floatx80_to_int32(floatx80_abs(fp_quot.d), 
> &env->fp_status);
> target/m68k/softfloat.c:2714:            fp0 = floatx80_abs(a); /* Y = |X| */
> target/m68k/softfloat.c:2734:        fp0 = floatx80_abs(a); /* Y = |X| */
> target/mips/tcg/fpu_helper.c:977:   return float64_abs(fdt0);
> target/mips/tcg/fpu_helper.c:982:    return float32_abs(fst0);
> target/mips/tcg/fpu_helper.c:990:    wt0 = float32_abs(fdt0 & 0XFFFFFFFF);
> target/mips/tcg/fpu_helper.c:991:    wth0 = float32_abs(fdt0 >> 32);
> target/mips/tcg/fpu_helper.c:997:   return float64_chs(fdt0);
> target/mips/tcg/fpu_helper.c:1002:    return float32_chs(fst0);
> target/mips/tcg/fpu_helper.c:1010:    wt0 = float32_chs(fdt0 & 0XFFFFFFFF);
> target/mips/tcg/fpu_helper.c:1011:    wth0 = float32_chs(fdt0 >> 32);
> target/mips/tcg/fpu_helper.c:1365:    fdt2 = float64_chs(float64_sub(fdt2, float64_one,
> target/mips/tcg/fpu_helper.c:1374:    fst2 = float32_chs(float32_sub(fst2, float32_one,
> target/mips/tcg/fpu_helper.c:1389:    fstl2 = float32_chs(float32_sub(fstl2, float32_one,
> target/mips/tcg/fpu_helper.c:1391:    fsth2 = float32_chs(float32_sub(fsth2, float32_one,
> target/mips/tcg/fpu_helper.c:1401:    fdt2 = float64_chs(float64_div(fdt2, FLOAT_TWO64,
> target/mips/tcg/fpu_helper.c:1411:    fst2 = float32_chs(float32_div(fst2, FLOAT_TWO32,
> target/mips/tcg/fpu_helper.c:1428:    fstl2 = float32_chs(float32_div(fstl2, FLOAT_TWO32,
> target/mips/tcg/fpu_helper.c:1430:    fsth2 = float32_chs(float32_div(fsth2, FLOAT_TWO32,
> target/mips/tcg/fpu_helper.c:1633:    fst0 = float64_chs(fst0);
> target/mips/tcg/fpu_helper.c:1644:    fst0 = float32_chs(fst0);
> target/mips/tcg/fpu_helper.c:1662:    fstl0 = float32_chs(fstl0);
> target/mips/tcg/fpu_helper.c:1665:    fsth0 = float32_chs(fsth0);
> target/mips/tcg/fpu_helper.c:1676:    fst0 = float64_chs(fst0);
> target/mips/tcg/fpu_helper.c:1687:    fst0 = float32_chs(fst0);
> target/mips/tcg/fpu_helper.c:1705:    fstl0 = float32_chs(fstl0);
> target/mips/tcg/fpu_helper.c:1708:    fsth0 = float32_chs(fsth0);
> target/mips/tcg/fpu_helper.c:1781:    fdt0 = float64_abs(fdt0);                         \
> target/mips/tcg/fpu_helper.c:1782:    fdt1 = float64_abs(fdt1);                         \
> target/mips/tcg/fpu_helper.c:1860:    fst0 = float32_abs(fst0);                         \
> target/mips/tcg/fpu_helper.c:1861:    fst1 = float32_abs(fst1);                         \
> target/mips/tcg/fpu_helper.c:1950:    fst0 = float32_abs(fdt0 & 
> 0XFFFFFFFF);                      \
> target/mips/tcg/fpu_helper.c:1951:    fsth0 = float32_abs(fdt0 >> 32); 
>                           \
> target/mips/tcg/fpu_helper.c:1952:    fst1 = float32_abs(fdt1 & 
> 0XFFFFFFFF);                      \
> target/mips/tcg/fpu_helper.c:1953:    fsth1 = float32_abs(fdt1 >> 32); 
>                           \
> target/ppc/fpu_helper.c:44:        return float32_chs(a);
> target/ppc/int_helper.c:694:            float32 bneg = float32_chs(b->f32[i]);
> target/s390x/tcg/vec_fpu_helper.c:922:                a = float32_abs(a);
> target/s390x/tcg/vec_fpu_helper.c:923:                b = float32_abs(b);
> target/s390x/tcg/vec_fpu_helper.c:984:                a = float64_abs(a);
> target/s390x/tcg/vec_fpu_helper.c:985:                b = float64_abs(b);
> target/s390x/tcg/vec_fpu_helper.c:1042:            a = float128_abs(a);
> target/s390x/tcg/vec_fpu_helper.c:1043:            b = float128_abs(b);
> target/sparc/fop_helper.c:119:    return float32_chs(src);
> target/sparc/fop_helper.c:125:    return float64_chs(src);
> target/sparc/fop_helper.c:130:    QT0 = float128_chs(QT1);
> target/sparc/fop_helper.c:234:    return float32_abs(src);
> target/sparc/fop_helper.c:240:    return float64_abs(src);
> target/sparc/fop_helper.c:245:    QT0 = float128_abs(QT1);
> target/xtensa/fpu_helper.c:126:    return float64_abs(v);
> target/xtensa/fpu_helper.c:131:    return float32_abs(v);
> target/xtensa/fpu_helper.c:136:    return float64_chs(v);
> target/xtensa/fpu_helper.c:141:    return float32_chs(v);

With only a few exceptions, most of of these results are part of a larger out-of-line 
operation.


r~



  reply	other threads:[~2024-02-02  4:32 UTC|newest]

Thread overview: 35+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-11-03 17:38 [PATCH 00/22] target/sparc: floating-point cleanup Richard Henderson
2023-11-03 17:38 ` [PATCH 01/22] target/sparc: Use tcg_gen_qemu_{ld, st}_i128 for ASI_M_BCOPY Richard Henderson
2023-11-03 17:38 ` [PATCH 02/22] target/sparc: Use tcg_gen_qemu_{ld, st}_i128 for ASI_M_BFILL Richard Henderson
2023-11-03 17:38 ` [PATCH 03/22] target/sparc: Remove gen_dest_fpr_F Richard Henderson
2024-01-30  7:54   ` Philippe Mathieu-Daudé
2023-11-03 17:38 ` [PATCH 04/22] target/sparc: Introduce gen_{load,store}_fpr_Q Richard Henderson
2024-01-30  7:56   ` Philippe Mathieu-Daudé
2023-11-03 17:38 ` [PATCH 05/22] target/sparc: Inline FNEG, FABS Richard Henderson
2024-01-30  8:04   ` Philippe Mathieu-Daudé
2024-01-30  8:40   ` Philippe Mathieu-Daudé
2024-02-02  4:32     ` Richard Henderson [this message]
2023-11-03 17:38 ` [PATCH 06/22] target/sparc: Use i128 for FSQRTq Richard Henderson
2023-11-03 17:38 ` [PATCH 07/22] target/sparc: Use i128 for FADDq, FSUBq, FMULq, FDIVq Richard Henderson
2023-11-03 17:38 ` [PATCH 08/22] target/sparc: Use i128 for FqTOs, FqTOi Richard Henderson
2023-11-03 17:38 ` [PATCH 09/22] target/sparc: Use i128 for FqTOd, FqTOx Richard Henderson
2023-11-03 17:38 ` [PATCH 10/22] target/sparc: Use i128 for FCMPq, FCMPEq Richard Henderson
2023-11-03 17:38 ` [PATCH 11/22] target/sparc: Use i128 for FsTOq, FiTOq Richard Henderson
2023-11-03 17:38 ` [PATCH 12/22] target/sparc: Use i128 for FdTOq, FxTOq Richard Henderson
2023-11-03 17:38 ` [PATCH 13/22] target/sparc: Use i128 for Fdmulq Richard Henderson
2023-11-03 17:38 ` [PATCH 14/22] target/sparc: Remove qt0, qt1 temporaries Richard Henderson
2023-11-03 17:38 ` [PATCH 15/22] target/sparc: Introduce cpu_get_fsr, cpu_put_fsr Richard Henderson
2024-01-30  8:10   ` Philippe Mathieu-Daudé
2023-11-03 17:38 ` [PATCH 16/22] target/split: Split ver from env->fsr Richard Henderson
2024-01-30  8:11   ` Philippe Mathieu-Daudé
2024-01-30 12:56     ` BALATON Zoltan
2023-11-03 17:38 ` [PATCH 17/22] target/sparc: Clear cexc and ftt in do_check_ieee_exceptions Richard Henderson
2023-11-03 17:38 ` [PATCH 18/22] target/sparc: Merge check_ieee_exceptions with FPop helpers Richard Henderson
2023-11-03 17:38 ` [PATCH 19/22] target/sparc: Split cexc and ftt from env->fsr Richard Henderson
2023-11-03 17:38 ` [PATCH 20/22] target/sparc: Remove cpu_fsr Richard Henderson
2023-11-03 17:38 ` [PATCH 21/22] target/sparc: Split fcc out of env->fsr Richard Henderson
2023-11-03 17:38 ` [PATCH 22/22] target/sparc: Remove FSR_FTT_NMASK, FSR_FTT_CEXC_NMASK Richard Henderson
2024-01-28  6:49 ` [PATCH 00/22] target/sparc: floating-point cleanup Richard Henderson
2024-01-29 21:50   ` Mark Cave-Ayland
2024-01-31 21:49   ` Mark Cave-Ayland
2024-02-02  5:36     ` Richard Henderson

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=0670f7d7-6518-4c72-a3b8-bd38152bcf94@linaro.org \
    --to=richard.henderson@linaro.org \
    --cc=mark.cave-ayland@ilande.co.uk \
    --cc=philmd@linaro.org \
    --cc=qemu-devel@nongnu.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).