From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pg1-x534.google.com (mail-pg1-x534.google.com [IPv6:2607:f8b0:4864:20::534]) by sourceware.org (Postfix) with ESMTPS id F1AEC3836E78 for ; Tue, 31 May 2022 21:52:28 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org F1AEC3836E78 Received: by mail-pg1-x534.google.com with SMTP id s68so17406pgs.10 for ; Tue, 31 May 2022 14:52:28 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=x-gm-message-state:mime-version:references:in-reply-to:from:date :message-id:subject:to:cc; bh=UklC6DEIYZdkH5a+CA7wlvNsTbLU2Khb+CgknFNLeJ0=; b=UG1BuVIFs3EvUzNPlrq6xN8fHYIK3VHvuksudeh1PtLOY1tlcWxJJF16JqTWftsti2 8HZMAnqN7dAQFnIAWeS0R+amCACQF3U9g+uc+t2Vblob6T3PINzylSXIQRdNXF4pqJUG MD6F5+F8gPOMn5H9t39ZhIIBqz9aBO092sGgfsMfLmhNyCA5xvHG7DRHGcEYOszUTd0L 0WVEcRO84xNefVXPmAYDT4yqZx0i+I0MCtzhX7ala5kRCjxmLIw98/4OyMu48STo0/KR gGByVdG4l1IiNPa2PiST7fhpdVYlhKfuX9Jhelm7NcPSzD6UolE0LJbvdINb+dJjojuA OKAA== X-Gm-Message-State: AOAM5306XbleglH+tb+BLlQfEhcEfGDuMtfzj7HQF5bE5ilVFX3XJcKP VqZJJs94z0wkehCyUCJoagJSBYTkDTArDRNzq4Bwh8m3OKc= X-Google-Smtp-Source: ABdhPJyRRptqVZfSt1G1DOz7+hQ2p8lY9IehqAmRuXVYg5x3jWuAPKcFTEYwWBQHPx5irXMeqfG5XjojS/jzOdWaZzM= X-Received: by 2002:a63:5464:0:b0:3c1:4930:fbd5 with SMTP id e36-20020a635464000000b003c14930fbd5mr52261756pgm.94.1654033947909; Tue, 31 May 2022 14:52:27 -0700 (PDT) MIME-Version: 1.0 References: <20220531213502.227895-1-adhemerval.zanella@linaro.org> <20220531213502.227895-6-adhemerval.zanella@linaro.org> In-Reply-To: <20220531213502.227895-6-adhemerval.zanella@linaro.org> From: "H.J. Lu" Date: Tue, 31 May 2022 14:51:51 -0700 Message-ID: Subject: Re: [PATCH 5/5] math: Use generic sinf implementation for SSE2 version To: Adhemerval Zanella Cc: GNU C Library Content-Type: text/plain; charset="UTF-8" X-Spam-Status: No, score=-3025.4 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, FREEMAIL_FROM, GIT_PATCH_0, KAM_SHORT, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP, T_SCC_BODY_TEXT_LINE autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org X-BeenThere: libc-alpha@sourceware.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Libc-alpha mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Tue, 31 May 2022 21:52:32 -0000 ? On Tue, May 31, 2022 at 2:35 PM Adhemerval Zanella wrote: > > The generic implementation shows slight better performance > (gcc 11.2.1 on a Ryzen 9 5900X): > > * s_sincosf-sse2.S: > "sincosf": { > "workload-random": { > "duration": 3.89961e+09, > "iterations": 9.5472e+07, > "reciprocal-throughput": 40.8429, > "latency": 40.8483, > "max-throughput": 2.4484e+07, > "min-throughput": 2.44808e+07 > } > } > > * generic s_cossinf.c: > "sincosf": { > "workload-random": { > "duration": 3.71953e+09, > "iterations": 1.48512e+08, > "reciprocal-throughput": 25.0515, > "latency": 25.0391, > "max-throughput": 3.99177e+07, > "min-throughput": 3.99375e+07 > } > } > > Checked on i686-linux-gnu. > --- > sysdeps/i386/i686/fpu/multiarch/Makefile | 1 + > .../i386/i686/fpu/multiarch/libm-test-ulps | 3 + > .../i386/i686/fpu/multiarch/s_sincosf-sse2.S | 585 ------------------ > .../i386/i686/fpu/multiarch/s_sincosf-sse2.c | 3 + > 4 files changed, 7 insertions(+), 585 deletions(-) > delete mode 100644 sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.S > create mode 100644 sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.c > > diff --git a/sysdeps/i386/i686/fpu/multiarch/Makefile b/sysdeps/i386/i686/fpu/multiarch/Makefile > index c2cd5cefef..c8b0a6528a 100644 > --- a/sysdeps/i386/i686/fpu/multiarch/Makefile > +++ b/sysdeps/i386/i686/fpu/multiarch/Makefile > @@ -9,4 +9,5 @@ CFLAGS-e_logf-sse2.c = -msse2 -mfpmath=sse > CFLAGS-e_powf-sse2.c = -msse2 -mfpmath=sse > CFLAGS-s_cosf-sse2.c = -msse2 -mfpmath=sse > CFLAGS-s_sinf-sse2.c = -msse2 -mfpmath=sse > +CFLAGS-s_sincosf-sse2.c = -msse2 -mfpmath=sse > endif > diff --git a/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps b/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps > index 0357d7bd14..284e3fa275 100644 > --- a/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps > +++ b/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps > @@ -1492,16 +1492,19 @@ ldouble: 1 > > Function: "sincos_downward": > double: 1 > +float: 1 > float128: 3 > ldouble: 3 > > Function: "sincos_towardzero": > double: 1 > +float: 1 > float128: 2 > ldouble: 2 > > Function: "sincos_upward": > double: 1 > +float: 1 > float128: 3 > ldouble: 3 > > diff --git a/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.S b/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.S > deleted file mode 100644 > index 71e222b930..0000000000 > --- a/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.S > +++ /dev/null > @@ -1,585 +0,0 @@ > -/* Optimized with sse2 version of sincosf > - Copyright (C) 2012-2022 Free Software Foundation, Inc. > - This file is part of the GNU C Library. > - > - The GNU C Library is free software; you can redistribute it and/or > - modify it under the terms of the GNU Lesser General Public > - License as published by the Free Software Foundation; either > - version 2.1 of the License, or (at your option) any later version. > - > - The GNU C Library is distributed in the hope that it will be useful, > - but WITHOUT ANY WARRANTY; without even the implied warranty of > - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU > - Lesser General Public License for more details. > - > - You should have received a copy of the GNU Lesser General Public > - License along with the GNU C Library; if not, see > - . */ > - > -#include > -#include > - > -/* Short algorithm description: > - * > - * 1) if |x|==0: sin(x)=x, > - * cos(x)=1. > - * 2) if |x|<2^-27: sin(x)=x-x*DP_SMALL, raising underflow only when needed, > - * cos(x)=1-|x|. > - * 3) if |x|<2^-5 : sin(x)=x+x*x^2*DP_SIN2_0+x^5*DP_SIN2_1, > - * cos(x)=1+1*x^2*DP_COS2_0+x^5*DP_COS2_1 > - * 4) if |x|< Pi/4: sin(x)=x+x*x^2*(S0+x^2*(S1+x^2*(S2+x^2*(S3+x^2*S4)))), > - * cos(x)=1+1*x^2*(C0+x^2*(C1+x^2*(C2+x^2*(C3+x^2*C4)))). > - * 5) if |x| < 9*Pi/4: > - * 5.1) Range reduction: > - * k=trunc(|x|/(Pi/4)), j=(k+1)&0x0e, n=k+1, t=|x|-j*Pi/4. > - * 5.2) Reconstruction: > - * sign_sin = sign(x) * (-1.0)^(( n >>2)&1) > - * sign_cos = (-1.0)^(((n+2)>>2)&1) > - * poly_sin = ((((S4*t^2 + S3)*t^2 + S2)*t^2 + S1)*t^2 + S0)*t^2*t+t > - * poly_cos = ((((C4*t^2 + C3)*t^2 + C2)*t^2 + C1)*t^2 + C0)*t^2*s+s > - * if(n&2 != 0) { > - * using cos(t) and sin(t) polynomials for |t| - * cos(x) = poly_sin * sign_cos > - * sin(x) = poly_cos * sign_sin > - * } else { > - * sin(x) = poly_sin * sign_sin > - * cos(x) = poly_cos * sign_cos > - * } > - * 6) if |x| < 2^23, large args: > - * 6.1) Range reduction: > - * k=trunc(|x|/(Pi/4)), j=(k+1)&0xfffffffe, n=k+1, t=|x|-j*Pi/4 > - * 6.2) Reconstruction same as (5.2). > - * 7) if |x| >= 2^23, very large args: > - * 7.1) Range reduction: > - * k=trunc(|x|/(Pi/4)), j=(k+1)&0xfffffffe, n=k+1, t=|x|-j*Pi/4. > - * 7.2) Reconstruction same as (5.2). > - * 8) if x is Inf, return x-x, and set errno=EDOM. > - * 9) if x is NaN, return x-x. > - * > - * Special cases: > - * sin/cos(+-0) = +-0/1 not raising inexact/underflow, > - * sin/cos(subnormal) raises inexact/underflow, > - * sin/cos(min_normalized) raises inexact/underflow, > - * sin/cos(normalized) raises inexact, > - * sin/cos(Inf) = NaN, raises invalid, sets errno to EDOM, > - * sin/cos(NaN) = NaN. > - */ > - > -#ifdef PIC > -# define MO1(symbol) L(symbol)##@GOTOFF(%ebx) > -# define MO2(symbol,reg2,_scale) L(symbol)##@GOTOFF(%ebx,reg2,_scale) > -# define CFI_PUSH(REG) cfi_adjust_cfa_offset(4); cfi_rel_offset(REG,0) > -# define CFI_POP(REG) cfi_adjust_cfa_offset(-4); cfi_restore(REG) > -# define PUSH(REG) pushl REG; CFI_PUSH(REG) > -# define POP(REG) popl REG; CFI_POP(REG) > -# define ENTRANCE PUSH(%ebx); LOAD_PIC_REG(bx) > -# define RETURN POP(%ebx); ret; CFI_PUSH(%ebx) > -# define ARG_X 8(%esp) > -# define ARG_SIN_PTR 12(%esp) > -# define ARG_COS_PTR 16(%esp) > -#else > -# define MO1(symbol) L(symbol) > -# define MO2(symbol,reg2,_scale) L(symbol)(,reg2,_scale) > -# define ENTRANCE > -# define RETURN ret > -# define ARG_X 4(%esp) > -# define ARG_SIN_PTR 8(%esp) > -# define ARG_COS_PTR 12(%esp) > -#endif > - > - .text > -ENTRY(__sincosf_sse2) > - /* Input: single precision x on stack at address ARG_X */ > - /* pointer to sin result on stack at address ARG_SIN_PTR */ > - /* pointer to cos result on stack at address ARG_COS_PTR */ > - > - ENTRANCE > - movl ARG_X, %eax /* Bits of x */ > - cvtss2sd ARG_X, %xmm0 /* DP x */ > - andl $0x7fffffff, %eax /* |x| */ > - > - cmpl $0x3f490fdb, %eax /* |x| - jb L(arg_less_pio4) > - > - /* Here if |x|>=Pi/4 */ > - movd %eax, %xmm3 /* SP |x| */ > - andpd MO1(DP_ABS_MASK),%xmm0 /* DP |x| */ > - movss MO1(SP_INVPIO4), %xmm2 /* SP 1/(Pi/4) */ > - > - cmpl $0x40e231d6, %eax /* |x|<9*Pi/4 ? */ > - jae L(large_args) > - > - /* Here if Pi/4<=|x|<9*Pi/4 */ > - mulss %xmm3, %xmm2 /* SP |x|/(Pi/4) */ > - movl ARG_X, %ecx /* Load x */ > - cvttss2si %xmm2, %eax /* k, number of Pi/4 in x */ > - shrl $29, %ecx /* (sign of x) << 2 */ > - addl $1, %eax /* k+1 */ > - movl $0x0e, %edx > - andl %eax, %edx /* j = (k+1)&0x0e */ > - subsd MO2(PIO4J,%edx,8), %xmm0/* t = |x| - j * Pi/4 */ > - > -L(reconstruction): > - /* Input: %eax=n, %xmm0=t, %ecx=sign(x) */ > - > - movaps %xmm0, %xmm4 /* t */ > - movhpd MO1(DP_ONES), %xmm4 /* 1|t */ > - mulsd %xmm0, %xmm0 /* y=t^2 */ > - movl $2, %edx > - unpcklpd %xmm0, %xmm0 /* y|y */ > - addl %eax, %edx /* k+2 */ > - movaps %xmm0, %xmm1 /* y|y */ > - mulpd %xmm0, %xmm0 /* z=t^4|z=t^4 */ > - > - movaps MO1(DP_SC4), %xmm2 /* S4 */ > - mulpd %xmm0, %xmm2 /* z*S4 */ > - movaps MO1(DP_SC3), %xmm3 /* S3 */ > - mulpd %xmm0, %xmm3 /* z*S3 */ > - xorl %eax, %ecx /* (sign_x ^ (k>>2))<<2 */ > - addpd MO1(DP_SC2), %xmm2 /* S2+z*S4 */ > - mulpd %xmm0, %xmm2 /* z*(S2+z*S4) */ > - shrl $2, %edx /* (k+2)>>2 */ > - addpd MO1(DP_SC1), %xmm3 /* S1+z*S3 */ > - mulpd %xmm0, %xmm3 /* z*(S1+z*S3) */ > - shrl $2, %ecx /* sign_x ^ k>>2 */ > - addpd MO1(DP_SC0), %xmm2 /* S0+z*(S2+z*S4) */ > - andl $1, %edx /* sign_cos = ((k+2)>>2)&1 */ > - mulpd %xmm1, %xmm2 /* y*(S0+z*(S2+z*S4)) */ > - andl $1, %ecx /* sign_sin = sign_x ^ ((k>>2)&1) */ > - addpd %xmm2, %xmm3 /* y*(S0+y*(S1+y*(S2+y*(S3+y*S4)))) */ > - mulpd %xmm4, %xmm3 /*t*y*(S0+y*(S1+y*(S2+y*(S3+y*S4))))*/ > - testl $2, %eax /* n&2 != 0 ? */ > - addpd %xmm4, %xmm3 /*t+t*y*(S0+y*(S1+y*(S2+y*(S3+y*S4))*/ > - jnz L(sin_result_sin_poly) > - > -/*L(sin_result_cos_poly):*/ > - /* > - * Here if > - * cos(x) = poly_sin * sign_cos > - * sin(x) = poly_cos * sign_sin > - */ > - movsd MO2(DP_ONES,%ecx,8), %xmm4/* 0|sign_sin */ > - movhpd MO2(DP_ONES,%edx,8), %xmm4/* sign_cos|sign_sin */ > - mulpd %xmm4, %xmm3 /* result_cos|result_sin */ > - movl ARG_SIN_PTR, %eax > - cvtpd2ps %xmm3, %xmm0 /* SP results */ > - movl ARG_COS_PTR, %ecx > - movss %xmm0, (%eax) /* store sin(x) from xmm0[0] */ > - shufps $1, %xmm0, %xmm0 /* move cos(x) to xmm0[0] */ > - movss %xmm0, (%ecx) /* store cos(x) */ > - RETURN > - > - .p2align 4 > -L(sin_result_sin_poly): > - /* > - * Here if > - * sin(x) = poly_sin * sign_sin > - * cos(x) = poly_cos * sign_cos > - */ > - movsd MO2(DP_ONES,%edx,8), %xmm4/* 0|sign_cos */ > - movhpd MO2(DP_ONES,%ecx,8), %xmm4/* sign_sin|sign_cos */ > - mulpd %xmm4, %xmm3 /* result_sin|result_cos */ > - movl ARG_SIN_PTR, %eax > - cvtpd2ps %xmm3, %xmm0 /* SP results */ > - movl ARG_COS_PTR, %ecx > - movss %xmm0, (%ecx) /* store cos(x) from xmm0[0] */ > - shufps $1, %xmm0, %xmm0 /* move sin(x) to xmm0[0] */ > - movss %xmm0, (%eax) /* store sin(x) */ > - RETURN > - > - .p2align 4 > -L(large_args): > - /* Here if |x|>=9*Pi/4 */ > - cmpl $0x7f800000, %eax /* x is Inf or NaN ? */ > - jae L(arg_inf_or_nan) > - > - /* Here if finite |x|>=9*Pi/4 */ > - cmpl $0x4b000000, %eax /* |x|<2^23 ? */ > - jae L(very_large_args) > - > - /* Here if 9*Pi/4<=|x|<2^23 */ > - movsd MO1(DP_INVPIO4), %xmm1 /* 1/(Pi/4) */ > - mulsd %xmm0, %xmm1 /* |x|/(Pi/4) */ > - cvttsd2si %xmm1, %eax /* k=trunc(|x|/(Pi/4)) */ > - addl $1, %eax /* k+1 */ > - movl %eax, %edx > - andl $0xfffffffe, %edx /* j=(k+1)&0xfffffffe */ > - cvtsi2sdl %edx, %xmm4 /* DP j */ > - movl ARG_X, %ecx /* Load x */ > - movsd MO1(DP_PIO4HI), %xmm2 /* -PIO4HI = high part of -Pi/4 */ > - shrl $29, %ecx /* (sign of x) << 2 */ > - mulsd %xmm4, %xmm2 /* -j*PIO4HI */ > - movsd MO1(DP_PIO4LO), %xmm3 /* -PIO4LO = low part of -Pi/4 */ > - addsd %xmm2, %xmm0 /* |x| - j*PIO4HI */ > - mulsd %xmm3, %xmm4 /* j*PIO4LO */ > - addsd %xmm4, %xmm0 /* t = |x| - j*PIO4HI - j*PIO4LO */ > - jmp L(reconstruction) > - > - .p2align 4 > -L(very_large_args): > - /* Here if finite |x|>=2^23 */ > - > - /* bitpos = (ix>>23) - BIAS_32 + 59; */ > - shrl $23, %eax /* eb = biased exponent of x */ > - subl $68, %eax /* bitpos=eb-0x7f+59, where 0x7f */ > - /*is exponent bias */ > - movl $28, %ecx /* %cl=28 */ > - movl %eax, %edx /* bitpos copy */ > - > - /* j = bitpos/28; */ > - div %cl /* j in register %al=%ax/%cl */ > - movapd %xmm0, %xmm3 /* |x| */ > - andl $0xff, %eax /* clear unneeded remainder from %ah*/ > - > - imull $28, %eax, %ecx /* j*28 */ > - movsd MO1(DP_HI_MASK), %xmm4 /* DP_HI_MASK */ > - movapd %xmm0, %xmm5 /* |x| */ > - mulsd -2*8+MO2(_FPI,%eax,8), %xmm3/* tmp3 = FPI[j-2]*|x| */ > - movapd %xmm0, %xmm1 /* |x| */ > - mulsd -1*8+MO2(_FPI,%eax,8), %xmm5/* tmp2 = FPI[j-1]*|x| */ > - mulsd 0*8+MO2(_FPI,%eax,8), %xmm0/* tmp0 = FPI[j]*|x| */ > - addl $19, %ecx /* j*28+19 */ > - mulsd 1*8+MO2(_FPI,%eax,8), %xmm1/* tmp1 = FPI[j+1]*|x| */ > - cmpl %ecx, %edx /* bitpos>=j*28+19 ? */ > - jl L(very_large_skip1) > - > - /* Here if bitpos>=j*28+19 */ > - andpd %xmm3, %xmm4 /* HI(tmp3) */ > - subsd %xmm4, %xmm3 /* tmp3 = tmp3 - HI(tmp3) */ > -L(very_large_skip1): > - > - movsd MO1(DP_2POW52), %xmm6 > - movapd %xmm5, %xmm2 /* tmp2 copy */ > - addsd %xmm3, %xmm5 /* tmp5 = tmp3 + tmp2 */ > - movl $1, %edx > - addsd %xmm5, %xmm6 /* tmp6 = tmp5 + 2^52 */ > - movsd 8+MO1(DP_2POW52), %xmm4 > - movd %xmm6, %eax /* k = I64_LO(tmp6); */ > - addsd %xmm6, %xmm4 /* tmp4 = tmp6 - 2^52 */ > - movl ARG_X, %ecx /* Load x */ > - comisd %xmm5, %xmm4 /* tmp4 > tmp5 ? */ > - jbe L(very_large_skip2) > - > - /* Here if tmp4 > tmp5 */ > - subl $1, %eax /* k-- */ > - addsd 8+MO1(DP_ONES), %xmm4 /* tmp4 -= 1.0 */ > -L(very_large_skip2): > - > - andl %eax, %edx /* k&1 */ > - subsd %xmm4, %xmm3 /* tmp3 -= tmp4 */ > - addsd MO2(DP_ZERONE,%edx,8), %xmm3/* t = DP_ZERONE[k&1] + tmp3 */ > - addsd %xmm2, %xmm3 /* t += tmp2 */ > - shrl $29, %ecx /* (sign of x) << 2 */ > - addsd %xmm3, %xmm0 /* t += tmp0 */ > - addl $1, %eax /* n=k+1 */ > - addsd %xmm1, %xmm0 /* t += tmp1 */ > - mulsd MO1(DP_PIO4), %xmm0 /* t *= PI04 */ > - > - jmp L(reconstruction) /* end of very_large_args peth */ > - > - .p2align 4 > -L(arg_less_pio4): > - /* Here if |x| - cmpl $0x3d000000, %eax /* |x|<2^-5 ? */ > - jl L(arg_less_2pn5) > - > - /* Here if 2^-5<=|x| - movaps %xmm0, %xmm3 /* DP x */ > - movhpd MO1(DP_ONES), %xmm3 /* DP 1|x */ > - mulsd %xmm0, %xmm0 /* DP y=x^2 */ > - unpcklpd %xmm0, %xmm0 /* DP y|y */ > - movaps %xmm0, %xmm1 /* y|y */ > - mulpd %xmm0, %xmm0 /* z=x^4|z=x^4 */ > - > - movapd MO1(DP_SC4), %xmm4 /* S4 */ > - mulpd %xmm0, %xmm4 /* z*S4 */ > - movapd MO1(DP_SC3), %xmm5 /* S3 */ > - mulpd %xmm0, %xmm5 /* z*S3 */ > - addpd MO1(DP_SC2), %xmm4 /* S2+z*S4 */ > - mulpd %xmm0, %xmm4 /* z*(S2+z*S4) */ > - addpd MO1(DP_SC1), %xmm5 /* S1+z*S3 */ > - mulpd %xmm0, %xmm5 /* z*(S1+z*S3) */ > - addpd MO1(DP_SC0), %xmm4 /* S0+z*(S2+z*S4) */ > - mulpd %xmm1, %xmm4 /* y*(S0+z*(S2+z*S4)) */ > - mulpd %xmm3, %xmm5 /* x*z*(S1+z*S3) */ > - mulpd %xmm3, %xmm4 /* x*y*(S0+z*(S2+z*S4)) */ > - addpd %xmm5, %xmm4 /*x*y*(S0+y*(S1+y*(S2+y*(S3+y*S4))))*/ > - movl ARG_SIN_PTR, %eax > - addpd %xmm4, %xmm3 /*x+x*y*(S0+y*(S1+y*(S2+y*(S3+y*S4))*/ > - movl ARG_COS_PTR, %ecx > - cvtpd2ps %xmm3, %xmm0 /* SP results */ > - movss %xmm0, (%eax) /* store sin(x) from xmm0[0] */ > - shufps $1, %xmm0, %xmm0 /* move cos(x) to xmm0[0] */ > - movss %xmm0, (%ecx) /* store cos(x) */ > - RETURN > - > - .p2align 4 > -L(arg_less_2pn5): > - /* Here if |x|<2^-5 */ > - cmpl $0x32000000, %eax /* |x|<2^-27 ? */ > - jl L(arg_less_2pn27) > - > - /* Here if 2^-27<=|x|<2^-5 */ > - movaps %xmm0, %xmm1 /* DP x */ > - movhpd MO1(DP_ONES), %xmm1 /* DP 1|x */ > - mulsd %xmm0, %xmm0 /* DP x^2 */ > - unpcklpd %xmm0, %xmm0 /* DP x^2|x^2 */ > - > - movaps MO1(DP_SINCOS2_1), %xmm3/* DP DP_SIN2_1 */ > - mulpd %xmm0, %xmm3 /* DP x^2*DP_SIN2_1 */ > - addpd MO1(DP_SINCOS2_0), %xmm3/* DP DP_SIN2_0+x^2*DP_SIN2_1 */ > - mulpd %xmm0, %xmm3 /* DP x^2*DP_SIN2_0+x^4*DP_SIN2_1 */ > - mulpd %xmm1, %xmm3 /* DP x^3*DP_SIN2_0+x^5*DP_SIN2_1 */ > - addpd %xmm1, %xmm3 /* DP x+x^3*DP_SIN2_0+x^5*DP_SIN2_1 */ > - movl ARG_SIN_PTR, %eax > - cvtpd2ps %xmm3, %xmm0 /* SP results */ > - movl ARG_COS_PTR, %ecx > - movss %xmm0, (%eax) /* store sin(x) from xmm0[0] */ > - shufps $1, %xmm0, %xmm0 /* move cos(x) to xmm0[0] */ > - movss %xmm0, (%ecx) /* store cos(x) */ > - RETURN > - > - .p2align 4 > -L(arg_less_2pn27): > - movss ARG_X, %xmm7 /* SP x */ > - cmpl $0, %eax /* x=0 ? */ > - je L(arg_zero) /* in case x=0 return sin(+-0)==+-0 */ > - /* Here if |x|<2^-27 */ > - /* > - * Special cases here: > - * sin(subnormal) raises inexact/underflow > - * sin(min_normalized) raises inexact/underflow > - * sin(normalized) raises inexact > - * cos(here)=1-|x| (raising inexact) > - */ > - movaps %xmm0, %xmm3 /* DP x */ > - mulsd MO1(DP_SMALL), %xmm0 /* DP x*DP_SMALL */ > - subsd %xmm0, %xmm3 /* DP sin result is x-x*DP_SMALL */ > - andps MO1(SP_ABS_MASK), %xmm7 /* SP |x| */ > - cvtsd2ss %xmm3, %xmm0 /* sin(x) */ > - movl ARG_SIN_PTR, %eax > - movss MO1(SP_ONE), %xmm1 /* SP 1.0 */ > - movss %xmm0, (%eax) /* sin(x) store */ > - movl ARG_COS_PTR, %ecx > - subss %xmm7, %xmm1 /* cos(x) */ > - movss %xmm1, (%ecx) /* cos(x) store */ > - RETURN > - > - .p2align 4 > -L(arg_zero): > - movss MO1(SP_ONE), %xmm0 /* 1.0 */ > - movl ARG_SIN_PTR, %eax > - movl ARG_COS_PTR, %ecx > - movss %xmm7, (%eax) /* sin(+-0)==x */ > - movss %xmm0, (%ecx) /* cos(+-0)==1 */ > - RETURN > - > - .p2align 4 > -L(arg_inf_or_nan): > - movss ARG_X, %xmm7 /* SP x */ > - /* Here if |x| is Inf or NAN */ > - jne L(skip_errno_setting) /* in case of x is NaN */ > - > - /* Here if x is Inf. Set errno to EDOM. */ > - call JUMPTARGET(__errno_location) > - movl $EDOM, (%eax) > - > - .p2align 4 > -L(skip_errno_setting): > - /* Here if |x| is Inf or NAN. Continued. */ > - subss %xmm7, %xmm7 /* x-x, result is NaN */ > - movl ARG_SIN_PTR, %eax > - movl ARG_COS_PTR, %ecx > - movss %xmm7, (%eax) > - movss %xmm7, (%ecx) > - RETURN > -END(__sincosf_sse2) > - > - .section .rodata, "a" > - .p2align 3 > -L(PIO4J): /* Table of j*Pi/4, for j=0,1,..,10 */ > - .long 0x00000000,0x00000000 > - .long 0x54442d18,0x3fe921fb > - .long 0x54442d18,0x3ff921fb > - .long 0x7f3321d2,0x4002d97c > - .long 0x54442d18,0x400921fb > - .long 0x2955385e,0x400f6a7a > - .long 0x7f3321d2,0x4012d97c > - .long 0xe9bba775,0x4015fdbb > - .long 0x54442d18,0x401921fb > - .long 0xbeccb2bb,0x401c463a > - .long 0x2955385e,0x401f6a7a > - .type L(PIO4J), @object > - ASM_SIZE_DIRECTIVE(L(PIO4J)) > - > - .p2align 3 > -L(_FPI): /* 4/Pi broken into sum of positive DP values */ > - .long 0x00000000,0x00000000 > - .long 0x6c000000,0x3ff45f30 > - .long 0x2a000000,0x3e3c9c88 > - .long 0xa8000000,0x3c54fe13 > - .long 0xd0000000,0x3aaf47d4 > - .long 0x6c000000,0x38fbb81b > - .long 0xe0000000,0x3714acc9 > - .long 0x7c000000,0x3560e410 > - .long 0x56000000,0x33bca2c7 > - .long 0xac000000,0x31fbd778 > - .long 0xe0000000,0x300b7246 > - .long 0xe8000000,0x2e5d2126 > - .long 0x48000000,0x2c970032 > - .long 0xe8000000,0x2ad77504 > - .long 0xe0000000,0x290921cf > - .long 0xb0000000,0x274deb1c > - .long 0xe0000000,0x25829a73 > - .long 0xbe000000,0x23fd1046 > - .long 0x10000000,0x2224baed > - .long 0x8e000000,0x20709d33 > - .long 0x80000000,0x1e535a2f > - .long 0x64000000,0x1cef904e > - .long 0x30000000,0x1b0d6398 > - .long 0x24000000,0x1964ce7d > - .long 0x16000000,0x17b908bf > - .type L(_FPI), @object > - ASM_SIZE_DIRECTIVE(L(_FPI)) > - > -/* Coefficients of polynomials for */ > -/* sin(x)~=x+x*x^2*(DP_SIN2_0+x^2*DP_SIN2_1) in low DP part, */ > -/* cos(x)~=1+1*x^2*(DP_COS2_0+x^2*DP_COS2_1) in high DP part, */ > -/* for |x|<2^-5. */ > - .p2align 4 > -L(DP_SINCOS2_0): > - .long 0x5543d49d,0xbfc55555 > - .long 0xff5cc6fd,0xbfdfffff > - .type L(DP_SINCOS2_0), @object > - ASM_SIZE_DIRECTIVE(L(DP_SINCOS2_0)) > - > - .p2align 4 > -L(DP_SINCOS2_1): > - .long 0x75cec8c5,0x3f8110f4 > - .long 0xb178dac5,0x3fa55514 > - .type L(DP_SINCOS2_1), @object > - ASM_SIZE_DIRECTIVE(L(DP_SINCOS2_1)) > - > - .p2align 3 > -L(DP_ZERONE): > - .long 0x00000000,0x00000000 /* 0.0 */ > - .long 0x00000000,0xbff00000 /* 1.0 */ > - .type L(DP_ZERONE), @object > - ASM_SIZE_DIRECTIVE(L(DP_ZERONE)) > - > - .p2align 3 > -L(DP_ONES): > - .long 0x00000000,0x3ff00000 /* +1.0 */ > - .long 0x00000000,0xbff00000 /* -1.0 */ > - .type L(DP_ONES), @object > - ASM_SIZE_DIRECTIVE(L(DP_ONES)) > - > -/* Coefficients of polynomials for */ > -/* sin(t)~=t+t*t^2*(S0+t^2*(S1+t^2*(S2+t^2*(S3+t^2*S4)))) in low DP part, */ > -/* cos(t)~=1+1*t^2*(C0+t^2*(C1+t^2*(C2+t^2*(C3+t^2*C4)))) in high DP part, */ > -/* for |t| - .p2align 4 > -L(DP_SC4): > - .long 0x1674b58a,0xbe5a947e > - .long 0xdd8844d7,0xbe923c97 > - .type L(DP_SC4), @object > - ASM_SIZE_DIRECTIVE(L(DP_SC4)) > - > - .p2align 4 > -L(DP_SC3): > - .long 0x64e6b5b4,0x3ec71d72 > - .long 0x9ac43cc0,0x3efa00eb > - .type L(DP_SC3), @object > - ASM_SIZE_DIRECTIVE(L(DP_SC3)) > - > - .p2align 4 > -L(DP_SC2): > - .long 0x8b4bd1f9,0xbf2a019f > - .long 0x348b6874,0xbf56c16b > - .type L(DP_SC2), @object > - ASM_SIZE_DIRECTIVE(L(DP_SC2)) > - > - .p2align 4 > -L(DP_SC1): > - .long 0x10c2688b,0x3f811111 > - .long 0x545c50c7,0x3fa55555 > - .type L(DP_SC1), @object > - ASM_SIZE_DIRECTIVE(L(DP_SC1)) > - > - .p2align 4 > -L(DP_SC0): > - .long 0x55551cd9,0xbfc55555 > - .long 0xfffe98ae,0xbfdfffff > - .type L(DP_SC0), @object > - ASM_SIZE_DIRECTIVE(L(DP_SC0)) > - > - .p2align 3 > -L(DP_SMALL): > - .long 0x00000000,0x3cd00000 /* 2^(-50) */ > - .type L(DP_SMALL), @object > - ASM_SIZE_DIRECTIVE(L(DP_SMALL)) > - > - .p2align 3 > -L(DP_PIO4): > - .long 0x54442d18,0x3fe921fb /* Pi/4 */ > - .type L(DP_PIO4), @object > - ASM_SIZE_DIRECTIVE(L(DP_PIO4)) > - > - .p2align 3 > -L(DP_2POW52): > - .long 0x00000000,0x43300000 /* +2^52 */ > - .long 0x00000000,0xc3300000 /* -2^52 */ > - .type L(DP_2POW52), @object > - ASM_SIZE_DIRECTIVE(L(DP_2POW52)) > - > - .p2align 3 > -L(DP_INVPIO4): > - .long 0x6dc9c883,0x3ff45f30 /* 4/Pi */ > - .type L(DP_INVPIO4), @object > - ASM_SIZE_DIRECTIVE(L(DP_INVPIO4)) > - > - .p2align 3 > -L(DP_PIO4HI): > - .long 0x54000000,0xbfe921fb /* High part of Pi/4 */ > - .type L(DP_PIO4HI), @object > - ASM_SIZE_DIRECTIVE(L(DP_PIO4HI)) > - > - .p2align 3 > -L(DP_PIO4LO): > - .long 0x11A62633,0xbe010b46 /* Low part of Pi/4 */ > - .type L(DP_PIO4LO), @object > - ASM_SIZE_DIRECTIVE(L(DP_PIO4LO)) > - > - .p2align 2 > -L(SP_INVPIO4): > - .long 0x3fa2f983 /* 4/Pi */ > - .type L(SP_INVPIO4), @object > - ASM_SIZE_DIRECTIVE(L(SP_INVPIO4)) > - > - .p2align 4 > -L(DP_ABS_MASK): /* Mask for getting DP absolute value */ > - .long 0xffffffff,0x7fffffff > - .long 0xffffffff,0x7fffffff > - .type L(DP_ABS_MASK), @object > - ASM_SIZE_DIRECTIVE(L(DP_ABS_MASK)) > - > - .p2align 3 > -L(DP_HI_MASK): /* Mask for getting high 21 bits of DP value */ > - .long 0x00000000,0xffffffff > - .type L(DP_HI_MASK), @object > - ASM_SIZE_DIRECTIVE(L(DP_HI_MASK)) > - > - .p2align 4 > -L(SP_ABS_MASK): /* Mask for getting SP absolute value */ > - .long 0x7fffffff,0x7fffffff > - .long 0x7fffffff,0x7fffffff > - .type L(SP_ABS_MASK), @object > - ASM_SIZE_DIRECTIVE(L(SP_ABS_MASK)) t> - > - .p2align 2 > -L(SP_ONE): > - .long 0x3f800000 /* 1.0 */ > - .type L(SP_ONE), @object > - ASM_SIZE_DIRECTIVE(L(SP_ONE)) > - > -weak_alias(__sincosf, sincosf) > diff --git a/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.c b/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.c > new file mode 100644 > index 0000000000..46363fc11a > --- /dev/null > +++ b/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.c > @@ -0,0 +1,3 @@ > +#define SINCOSF __sincosf_sse2 > +#define SECTION __attribute__ ((section (".text.sse2"))) > +#include > -- > 2.34.1 > Don't you need to add SECTION to sysdeps/ieee754/flt-32/s_sincosf.c? -- H.J.