From patchwork Wed Nov 21 18:11:02 2018 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: "H.J. Lu" X-Patchwork-Id: 30248 Received: (qmail 54875 invoked by alias); 21 Nov 2018 18:11:15 -0000 Mailing-List: contact libc-alpha-help@sourceware.org; run by ezmlm Precedence: bulk List-Id: List-Unsubscribe: List-Subscribe: List-Archive: List-Post: List-Help: , Sender: libc-alpha-owner@sourceware.org Delivered-To: mailing list libc-alpha@sourceware.org Received: (qmail 54752 invoked by uid 89); 21 Nov 2018 18:11:14 -0000 Authentication-Results: sourceware.org; auth=none X-Spam-SWARE-Status: No, score=-26.2 required=5.0 tests=BAYES_00, FREEMAIL_FROM, GIT_PATCH_0, GIT_PATCH_1, GIT_PATCH_2, GIT_PATCH_3, KAM_SHORT, SPF_SOFTFAIL autolearn=ham version=3.3.2 spammy=71, Short, Low, 72 X-HELO: mga18.intel.com From: "H.J. Lu" To: libc-alpha@sourceware.org Subject: [PATCH 3/3] x86-64: Remove s_sincosf-sse2.S Date: Wed, 21 Nov 2018 10:11:02 -0800 Message-Id: <20181121181102.27119-4-hjl.tools@gmail.com> In-Reply-To: <20181121181102.27119-1-hjl.tools@gmail.com> References: <20181121181102.27119-1-hjl.tools@gmail.com> MIME-Version: 1.0 The current s_sincosf.c is faster than s_sincosf-sse2.S. On Broadwell with FMA disabled, bench-sincosf shows: Before After Improvement max 154.032 114.517 34% min 6.25 5.609 11% mean 14.8728 12.8589 15% * sysdeps/x86_64/fpu/s_sincosf.S: Removed. * sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.S: Likewise. * sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.c: New file. --- sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.S | 2 - sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.c | 2 + sysdeps/x86_64/fpu/s_sincosf.S | 566 ------------------ 3 files changed, 2 insertions(+), 568 deletions(-) delete mode 100644 sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.S create mode 100644 sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.c delete mode 100644 sysdeps/x86_64/fpu/s_sincosf.S diff --git a/sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.S b/sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.S deleted file mode 100644 index 51d012bb12..0000000000 --- a/sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.S +++ /dev/null @@ -1,2 +0,0 @@ -#define __sincosf __sincosf_sse2 -#include diff --git a/sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.c b/sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.c new file mode 100644 index 0000000000..3c6d293803 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/s_sincosf-sse2.c @@ -0,0 +1,2 @@ +#define SINCOSF __sincosf_sse2 +#include diff --git a/sysdeps/x86_64/fpu/s_sincosf.S b/sysdeps/x86_64/fpu/s_sincosf.S deleted file mode 100644 index 2086e8ca5c..0000000000 --- a/sysdeps/x86_64/fpu/s_sincosf.S +++ /dev/null @@ -1,566 +0,0 @@ -/* Optimized sincosf function. - Copyright (C) 2012-2018 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 -#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|= 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. - */ - -# define ARG_SIN_PTR %rdi -# define ARG_COS_PTR %rsi - - .text -ENTRY(__sincosf) - /* Input: %xmm0 contains single precision argument x */ - /* %rdi points to sin result */ - /* %rsi points to cos result */ - - movd %xmm0, %eax /* Bits of x */ - movaps %xmm0, %xmm7 /* Copy of x */ - cvtss2sd %xmm0, %xmm0 /* DP x */ - movss L(SP_ABS_MASK)(%rip), %xmm3 - movl %eax, %r8d /* Copy of x bits */ - andl $0x7fffffff, %eax /* |x| */ - - cmpl $0x3f490fdb, %eax /* |x|=Pi/4 */ - andps %xmm7, %xmm3 /* SP |x| */ - andpd L(DP_ABS_MASK)(%rip),%xmm0 /* DP |x| */ - movss L(SP_INVPIO4)(%rip), %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 %r8d, %ecx /* Load x */ - cvttss2si %xmm2, %eax /* k, number of Pi/4 in x */ - lea L(PIO4J)(%rip), %r9 - shrl $29, %ecx /* (sign of x) << 2 */ - addl $1, %eax /* k+1 */ - movl $0x0e, %edx - andl %eax, %edx /* j = (k+1)&0x0e */ - subsd (%r9,%rdx,8), %xmm0 /* t = |x| - j * Pi/4 */ - -L(reconstruction): - /* Input: %eax=n, %xmm0=t, %ecx=sign(x) */ - - movaps %xmm0, %xmm4 /* t */ - movhpd L(DP_ONES)(%rip), %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 L(DP_SC4)(%rip), %xmm2 /* S4 */ - mulpd %xmm0, %xmm2 /* z*S4 */ - movaps L(DP_SC3)(%rip), %xmm3 /* S3 */ - mulpd %xmm0, %xmm3 /* z*S3 */ - xorl %eax, %ecx /* (sign_x ^ (k>>2))<<2 */ - addpd L(DP_SC2)(%rip), %xmm2 /* S2+z*S4 */ - mulpd %xmm0, %xmm2 /* z*(S2+z*S4) */ - shrl $2, %edx /* (k+2)>>2 */ - addpd L(DP_SC1)(%rip), %xmm3 /* S1+z*S3 */ - mulpd %xmm0, %xmm3 /* z*(S1+z*S3) */ - shrl $2, %ecx /* sign_x ^ k>>2 */ - addpd L(DP_SC0)(%rip), %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)))) */ - lea L(DP_ONES)(%rip), %r9 - 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 (%r9,%rcx,8), %xmm4 /* 0|sign_sin */ - movhpd (%r9,%rdx,8), %xmm4 /* sign_cos|sign_sin */ - mulpd %xmm4, %xmm3 /* result_cos|result_sin */ - cvtpd2ps %xmm3, %xmm0 /* SP results */ - movss %xmm0, (ARG_SIN_PTR) /* store sin(x) from xmm0[0] */ - shufps $1, %xmm0, %xmm0 /* move cos(x) to xmm0[0] */ - movss %xmm0, (ARG_COS_PTR) /* store cos(x) */ - ret - - .p2align 4 -L(sin_result_sin_poly): - /* - * Here if - * sin(x) = poly_sin * sign_sin - * cos(x) = poly_cos * sign_cos - */ - movsd (%r9,%rdx,8), %xmm4 /* 0|sign_cos */ - movhpd (%r9,%rcx,8), %xmm4 /* sign_sin|sign_cos */ - mulpd %xmm4, %xmm3 /* result_sin|result_cos */ - cvtpd2ps %xmm3, %xmm0 /* SP results */ - movss %xmm0, (ARG_COS_PTR) /* store cos(x) from xmm0[0] */ - shufps $1, %xmm0, %xmm0 /* move sin(x) to xmm0[0] */ - movss %xmm0, (ARG_SIN_PTR) /* store sin(x) */ - ret - - .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 L(DP_INVPIO4)(%rip), %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 %r8d, %ecx /* Load x */ - movsd L(DP_PIO4HI)(%rip), %xmm2 /* -PIO4HI = high part of -Pi/4 */ - shrl $29, %ecx /* (sign of x) << 2 */ - mulsd %xmm4, %xmm2 /* -j*PIO4HI */ - movsd L(DP_PIO4LO)(%rip), %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 */ - lea L(_FPI)(%rip), %r9 - movsd L(DP_HI_MASK)(%rip), %xmm4 /* DP_HI_MASK */ - movapd %xmm0, %xmm5 /* |x| */ - mulsd -16(%r9,%rax,8), %xmm3 /* tmp3 = FPI[j-2]*|x| */ - movapd %xmm0, %xmm1 /* |x| */ - mulsd -8(%r9,%rax,8), %xmm5 /* tmp2 = FPI[j-1]*|x| */ - mulsd (%r9,%rax,8), %xmm0 /* tmp0 = FPI[j]*|x| */ - addl $19, %ecx /* j*28+19 */ - mulsd 8(%r9,%rax,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 L(DP_2POW52)(%rip), %xmm6 - movapd %xmm5, %xmm2 /* tmp2 copy */ - addsd %xmm3, %xmm5 /* tmp5 = tmp3 + tmp2 */ - movl $1, %edx - addsd %xmm5, %xmm6 /* tmp6 = tmp5 + 2^52 */ - movsd 8+L(DP_2POW52)(%rip), %xmm4 - movd %xmm6, %eax /* k = I64_LO(tmp6); */ - addsd %xmm6, %xmm4 /* tmp4 = tmp6 - 2^52 */ - movl %r8d, %ecx /* Load x */ - comisd %xmm5, %xmm4 /* tmp4 > tmp5 ? */ - jbe L(very_large_skip2) - - /* Here if tmp4 > tmp5 */ - subl $1, %eax /* k-- */ - addsd 8+L(DP_ONES)(%rip), %xmm4 /* tmp4 -= 1.0 */ -L(very_large_skip2): - - andl %eax, %edx /* k&1 */ - lea L(DP_ZERONE)(%rip), %r9 - subsd %xmm4, %xmm3 /* tmp3 -= tmp4 */ - addsd (%r9,%rdx,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 L(DP_PIO4)(%rip), %xmm0 /* t *= PI04 */ - - jmp L(reconstruction) /* end of very_large_args peth */ - - .p2align 4 -L(arg_less_pio4): - /* Here if |x|