From patchwork Mon Mar 7 15:00:15 2022 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Sunil Pandey X-Patchwork-Id: 51663 Return-Path: X-Original-To: patchwork@sourceware.org Delivered-To: patchwork@sourceware.org Received: from server2.sourceware.org (localhost [IPv6:::1]) by sourceware.org (Postfix) with ESMTP id B6313385842C for ; Mon, 7 Mar 2022 15:35:27 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org B6313385842C DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=sourceware.org; s=default; t=1646667327; bh=duUECh6U/xFzp/3NqXPsoZeDMjIt1NPDc+mkPuQwd4U=; h=To:Subject:Date:In-Reply-To:References:List-Id:List-Unsubscribe: List-Archive:List-Post:List-Help:List-Subscribe:From:Reply-To: From; b=Dwwl2C+P+VYwhgiuivze54imeuxIOeZucHPkpN/WBaH4/+QkTVPNKIucM+9KMCza9 I10a1gGiVUkjjCK95KaP6bp2sxKcS1m1Ffl4XDvsfRZXWRBgbbotS0mTCDE3bsZrq3 VkLASt5pulfpl61WFLW3Kgb7qy8TTsmB9Uti9gCU= X-Original-To: libc-alpha@sourceware.org Delivered-To: libc-alpha@sourceware.org Received: from mga12.intel.com (mga12.intel.com [192.55.52.136]) by sourceware.org (Postfix) with ESMTPS id 563473858008 for ; Mon, 7 Mar 2022 15:02:29 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org 563473858008 X-IronPort-AV: E=McAfee;i="6200,9189,10278"; a="234364113" X-IronPort-AV: E=Sophos;i="5.90,162,1643702400"; d="scan'208";a="234364113" Received: from fmsmga001.fm.intel.com ([10.253.24.23]) by fmsmga106.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 07 Mar 2022 07:02:04 -0800 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.90,162,1643702400"; d="scan'208";a="687561708" Received: from scymds01.sc.intel.com ([10.148.94.138]) by fmsmga001.fm.intel.com with ESMTP; 07 Mar 2022 07:02:03 -0800 Received: from gskx-1.sc.intel.com (gskx-1.sc.intel.com [172.25.149.211]) by scymds01.sc.intel.com with ESMTP id 227F21ds016772; Mon, 7 Mar 2022 07:02:03 -0800 To: libc-alpha@sourceware.org Subject: [PATCH 020/126] x86_64: Fix svml_s_asinhf4_core_sse4.S code formatting Date: Mon, 7 Mar 2022 07:00:15 -0800 Message-Id: <20220307150201.10590-21-skpgkp2@gmail.com> X-Mailer: git-send-email 2.35.1 In-Reply-To: <20220307150201.10590-1-skpgkp2@gmail.com> References: <20220307150201.10590-1-skpgkp2@gmail.com> MIME-Version: 1.0 X-Spam-Status: No, score=-7.5 required=5.0 tests=BAYES_00, DKIM_ADSP_CUSTOM_MED, FORGED_GMAIL_RCVD, FREEMAIL_ENVFROM_END_DIGIT, FREEMAIL_FROM, GIT_PATCH_0, HK_RANDOM_ENVFROM, HK_RANDOM_FROM, KAM_DMARC_NONE, KAM_DMARC_STATUS, KAM_LOTSOFHASH, NML_ADSP_CUSTOM_MED, SPF_HELO_PASS, SPF_SOFTFAIL, TXREP, T_SCC_BODY_TEXT_LINE autolearn=ham autolearn_force=no version=3.4.4 X-Spam-Checker-Version: SpamAssassin 3.4.4 (2020-01-24) 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-Patchwork-Original-From: Sunil K Pandey via Libc-alpha From: Sunil Pandey Reply-To: Sunil K Pandey Errors-To: libc-alpha-bounces+patchwork=sourceware.org@sourceware.org Sender: "Libc-alpha" This commit contains following formatting changes 1. Instructions proceeded by a tab. 2. Instruction less than 8 characters in length have a tab between it and the first operand. 3. Instruction greater than 7 characters in length have a space between it and the first operand. 4. Tabs after `#define`d names and their value. 5. 8 space at the beginning of line replaced by tab. 6. Indent comments with code. 7. Remove redundent .text section. --- .../fpu/multiarch/svml_s_asinhf4_core_sse4.S | 895 +++++++++--------- 1 file changed, 447 insertions(+), 448 deletions(-) diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_asinhf4_core_sse4.S b/sysdeps/x86_64/fpu/multiarch/svml_s_asinhf4_core_sse4.S index 6b5c097e9c..26573bc6fe 100644 --- a/sysdeps/x86_64/fpu/multiarch/svml_s_asinhf4_core_sse4.S +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_asinhf4_core_sse4.S @@ -31,479 +31,478 @@ /* Offsets for data table __svml_sasinh_data_internal */ -#define SgnMask 0 -#define sOne 16 -#define sPoly 32 -#define iBrkValue 160 -#define iOffExpoMask 176 -#define sBigThreshold 192 -#define sC2 208 -#define sC3 224 -#define sHalf 240 -#define sLargestFinite 256 -#define sLittleThreshold 272 -#define sSign 288 -#define sThirtyOne 304 -#define sTopMask11 320 -#define sTopMask8 336 -#define XScale 352 -#define sLn2 368 +#define SgnMask 0 +#define sOne 16 +#define sPoly 32 +#define iBrkValue 160 +#define iOffExpoMask 176 +#define sBigThreshold 192 +#define sC2 208 +#define sC3 224 +#define sHalf 240 +#define sLargestFinite 256 +#define sLittleThreshold 272 +#define sSign 288 +#define sThirtyOne 304 +#define sTopMask11 320 +#define sTopMask8 336 +#define XScale 352 +#define sLn2 368 #include - .text - .section .text.sse4,"ax",@progbits + .section .text.sse4, "ax", @progbits ENTRY(_ZGVbN4v_asinhf_sse4) - subq $72, %rsp - cfi_def_cfa_offset(80) - movaps %xmm0, %xmm8 - -/* - * Split X into high and low parts, XHi (<= 11 bits) and XLo (<= 13 bits) - * We could use either X or |X| here, but it doesn't seem to matter - */ - movups sTopMask11+__svml_sasinh_data_internal(%rip), %xmm10 - movaps %xmm8, %xmm2 - andps %xmm8, %xmm10 - -/* - * Compute X^2 = (XHi + XLo)^2 = XHi^2 + XLo * (X + XHi) - * The two parts are shifted off by around 11 bits. So even though - * the low bit will not in general be exact, it's near enough - */ - movaps %xmm10, %xmm3 - subps %xmm10, %xmm2 - mulps %xmm10, %xmm3 - addps %xmm8, %xmm10 - -/* Load the constant 1 and a sign mask */ - movups sOne+__svml_sasinh_data_internal(%rip), %xmm7 - -/* - * Finally, express Y + W = X^2 + 1 accurately where Y has <= 8 bits. - * If |X| <= 1 then |XHi| <= 1 and so |X2Hi| <= 1, so we can treat 1 - * as the dominant component in the compensated summation. Otherwise, - * if |X| >= 1, then since X2Hi only has 22 significant bits, the basic - * addition will be exact anyway until we get to |X| >= 2^24. But by - * that time the log function is well-conditioned enough that the - * rounding error doesn't matter. Hence we can treat 1 as dominant even - * if it literally isn't. - */ - movaps %xmm7, %xmm11 - movaps %xmm7, %xmm4 - movups sTopMask8+__svml_sasinh_data_internal(%rip), %xmm12 - addps %xmm3, %xmm11 - mulps %xmm10, %xmm2 - subps %xmm11, %xmm4 - movaps %xmm12, %xmm0 - addps %xmm3, %xmm4 - -/* - * Unfortunately, we can still be in trouble if |X| <= 2^-5, since - * the absolute error 2^-(7+24)-ish in sqrt(1 + X^2) gets scaled up - * by 1/X and comes close to our threshold. Hence if |X| <= 2^-4, - * perform an alternative computation - * sqrt(1 + X^2) - 1 = X^2/2 - X^4/8 + X^6/16 - * X2 = X^2 - */ - addps %xmm2, %xmm3 - addps %xmm2, %xmm4 - andps %xmm11, %xmm0 - -/* - * Compute R = 1/sqrt(Y + W) * (1 + d) - * Force R to <= 8 significant bits. - * This means that R * Y and R^2 * Y are exactly representable. - */ - rsqrtps %xmm0, %xmm14 - subps %xmm0, %xmm11 - andps %xmm12, %xmm14 - addps %xmm11, %xmm4 - -/* - * Compute S = (Y/sqrt(Y + W)) * (1 + d) - * and T = (W/sqrt(Y + W)) * (1 + d) - * so that S + T = sqrt(Y + W) * (1 + d) - * S is exact, and the rounding error in T is OK. - */ - mulps %xmm14, %xmm0 - mulps %xmm14, %xmm4 - -/* - * Get the absolute value of the input, since we will exploit antisymmetry - * and mostly assume X >= 0 in the core computation - */ - movups SgnMask+__svml_sasinh_data_internal(%rip), %xmm6 - -/* - * Compute e = -(2 * d + d^2) - * The first FMR is exact, and the rounding error in the other is acceptable - * since d and e are ~ 2^-8 - */ - movaps %xmm14, %xmm13 - andps %xmm8, %xmm6 - -/* - * Obtain sqrt(1 + X^2) - 1 in two pieces - * sqrt(1 + X^2) - 1 - * = sqrt(Y + W) - 1 - * = (S + T) * (1 + Corr) - 1 - * = [S - 1] + [T + (S + T) * Corr] - * We need a compensated summation for the last part. We treat S - 1 - * as the larger part; it certainly is until about X < 2^-4, and in that - * case, the error is affordable since X dominates over sqrt(1 + X^2) - 1 - * Final sum is dTmp5 (hi) + dTmp7 (lo) - */ - movaps %xmm0, %xmm1 - -/* - * Check whether the input is finite, by checking |X| <= MaxFloat - * Otherwise set the rangemask so that the callout will get used. - * Note that this will also use the callout for NaNs since not(NaN <= MaxFloat) - */ - movaps %xmm6, %xmm9 - -/* - * The following computation can go wrong for very large X, basically - * because X^2 overflows. But for large X we have - * asinh(X) / log(2 X) - 1 =~= 1/(4 * X^2), so for X >= 2^30 - * we can just later stick X back into the log and tweak up the exponent. - * Actually we scale X by 2^-30 and tweak the exponent up by 31, - * to stay in the safe range for the later log computation. - * Compute a flag now telling us when do do this. - */ - movaps %xmm6, %xmm5 - cmpnleps sLargestFinite+__svml_sasinh_data_internal(%rip), %xmm9 - cmpltps sBigThreshold+__svml_sasinh_data_internal(%rip), %xmm5 - mulps %xmm0, %xmm13 - addps %xmm4, %xmm1 - subps %xmm7, %xmm0 - mulps %xmm4, %xmm14 - movmskps %xmm9, %edx - movaps %xmm7, %xmm9 - -/* - * Now 1 / (1 + d) - * = 1 / (1 + (sqrt(1 - e) - 1)) - * = 1 / sqrt(1 - e) - * = 1 + 1/2 * e + 3/8 * e^2 + 5/16 * e^3 + 35/128 * e^4 + ... - * So compute the first three nonconstant terms of that, so that - * we have a relative correction (1 + Corr) to apply to S etc. - * C1 = 1/2 - * C2 = 3/8 - * C3 = 5/16 - */ - movups sC3+__svml_sasinh_data_internal(%rip), %xmm15 - subps %xmm13, %xmm9 - movups sHalf+__svml_sasinh_data_internal(%rip), %xmm10 - subps %xmm14, %xmm9 - -/* sX2over2 = X^2/2 */ - mulps %xmm10, %xmm3 - mulps %xmm9, %xmm15 - -/* sX46 = -X^4/4 + X^6/8 */ - movaps %xmm3, %xmm2 - movaps %xmm3, %xmm12 - -/* - * Now do another compensated sum to add |X| + [sqrt(1 + X^2) - 1]. - * It's always safe to assume |X| is larger. - * This is the final 2-part argument to the log1p function - */ - movaps %xmm6, %xmm14 - addps sC2+__svml_sasinh_data_internal(%rip), %xmm15 - mulps %xmm9, %xmm15 - addps %xmm10, %xmm15 - mulps %xmm15, %xmm9 - mulps %xmm1, %xmm9 - -/* Now multiplex to the case X = 2^-30 * input, Xl = sL = 0 in the "big" case. */ - movups XScale+__svml_sasinh_data_internal(%rip), %xmm15 - addps %xmm9, %xmm4 - movaps %xmm4, %xmm11 - addps %xmm0, %xmm11 - subps %xmm11, %xmm0 - addps %xmm0, %xmm4 - -/* sX4over4 = X^4/4 */ - movaps %xmm3, %xmm0 - mulps %xmm3, %xmm0 - mulps %xmm0, %xmm2 - subps %xmm0, %xmm2 - -/* - * Now we feed into the log1p code, using H in place of _VARG1 and - * also adding L into Xl. - * compute 1+x as high, low parts - */ - movaps %xmm7, %xmm0 - -/* sX46over2 = -X^4/8 + x^6/16 */ - mulps %xmm2, %xmm10 - movaps %xmm7, %xmm2 - addps %xmm10, %xmm12 - subps %xmm12, %xmm3 - addps %xmm3, %xmm10 - -/* Now multiplex the two possible computations */ - movaps %xmm6, %xmm3 - cmpleps sLittleThreshold+__svml_sasinh_data_internal(%rip), %xmm3 - movaps %xmm3, %xmm13 - andps %xmm3, %xmm12 - andnps %xmm11, %xmm13 - movaps %xmm3, %xmm1 - orps %xmm12, %xmm13 - andnps %xmm4, %xmm1 - andps %xmm3, %xmm10 - movaps %xmm6, %xmm4 - orps %xmm10, %xmm1 - addps %xmm13, %xmm14 - mulps %xmm15, %xmm6 - maxps %xmm14, %xmm0 - minps %xmm14, %xmm2 - subps %xmm14, %xmm4 - movaps %xmm0, %xmm3 - addps %xmm4, %xmm13 - addps %xmm2, %xmm3 - addps %xmm13, %xmm1 - subps %xmm3, %xmm0 - movaps %xmm5, %xmm4 - andps %xmm5, %xmm3 - andnps %xmm6, %xmm4 - addps %xmm0, %xmm2 - -/* - * Now resume the main code. - * reduction: compute r,n - */ - movdqu iBrkValue+__svml_sasinh_data_internal(%rip), %xmm6 - orps %xmm3, %xmm4 - psubd %xmm6, %xmm4 - movaps %xmm7, %xmm0 - addps %xmm2, %xmm1 - movdqu iOffExpoMask+__svml_sasinh_data_internal(%rip), %xmm2 - pand %xmm4, %xmm2 - psrad $23, %xmm4 - cvtdq2ps %xmm4, %xmm3 - pslld $23, %xmm4 - andps %xmm5, %xmm1 - paddd %xmm6, %xmm2 - psubd %xmm4, %xmm0 - mulps %xmm0, %xmm1 - -/* polynomial evaluation */ - subps %xmm7, %xmm2 - movups sPoly+112+__svml_sasinh_data_internal(%rip), %xmm7 - addps %xmm2, %xmm1 - mulps %xmm1, %xmm7 - movaps %xmm5, %xmm2 - -/* Add 31 to the exponent in the "large" case to get log(2 * input) */ - movups sThirtyOne+__svml_sasinh_data_internal(%rip), %xmm0 - addps sPoly+96+__svml_sasinh_data_internal(%rip), %xmm7 - addps %xmm3, %xmm0 - mulps %xmm1, %xmm7 - andnps %xmm0, %xmm2 - andps %xmm5, %xmm3 - orps %xmm3, %xmm2 - addps sPoly+80+__svml_sasinh_data_internal(%rip), %xmm7 - -/* final reconstruction */ - mulps sLn2+__svml_sasinh_data_internal(%rip), %xmm2 - mulps %xmm1, %xmm7 - -/* Finally, reincorporate the original sign. */ - movups sSign+__svml_sasinh_data_internal(%rip), %xmm0 - andps %xmm8, %xmm0 - addps sPoly+64+__svml_sasinh_data_internal(%rip), %xmm7 - mulps %xmm1, %xmm7 - addps sPoly+48+__svml_sasinh_data_internal(%rip), %xmm7 - mulps %xmm1, %xmm7 - addps sPoly+32+__svml_sasinh_data_internal(%rip), %xmm7 - mulps %xmm1, %xmm7 - addps sPoly+16+__svml_sasinh_data_internal(%rip), %xmm7 - mulps %xmm1, %xmm7 - addps sPoly+__svml_sasinh_data_internal(%rip), %xmm7 - mulps %xmm1, %xmm7 - mulps %xmm1, %xmm7 - addps %xmm7, %xmm1 - addps %xmm2, %xmm1 - pxor %xmm1, %xmm0 - testl %edx, %edx - -/* Go to special inputs processing branch */ - jne L(SPECIAL_VALUES_BRANCH) - # LOE rbx rbp r12 r13 r14 r15 edx xmm0 xmm8 - -/* Restore registers - * and exit the function - */ + subq $72, %rsp + cfi_def_cfa_offset(80) + movaps %xmm0, %xmm8 + + /* + * Split X into high and low parts, XHi (<= 11 bits) and XLo (<= 13 bits) + * We could use either X or |X| here, but it doesn't seem to matter + */ + movups sTopMask11+__svml_sasinh_data_internal(%rip), %xmm10 + movaps %xmm8, %xmm2 + andps %xmm8, %xmm10 + + /* + * Compute X^2 = (XHi + XLo)^2 = XHi^2 + XLo * (X + XHi) + * The two parts are shifted off by around 11 bits. So even though + * the low bit will not in general be exact, it's near enough + */ + movaps %xmm10, %xmm3 + subps %xmm10, %xmm2 + mulps %xmm10, %xmm3 + addps %xmm8, %xmm10 + + /* Load the constant 1 and a sign mask */ + movups sOne+__svml_sasinh_data_internal(%rip), %xmm7 + + /* + * Finally, express Y + W = X^2 + 1 accurately where Y has <= 8 bits. + * If |X| <= 1 then |XHi| <= 1 and so |X2Hi| <= 1, so we can treat 1 + * as the dominant component in the compensated summation. Otherwise, + * if |X| >= 1, then since X2Hi only has 22 significant bits, the basic + * addition will be exact anyway until we get to |X| >= 2^24. But by + * that time the log function is well-conditioned enough that the + * rounding error doesn't matter. Hence we can treat 1 as dominant even + * if it literally isn't. + */ + movaps %xmm7, %xmm11 + movaps %xmm7, %xmm4 + movups sTopMask8+__svml_sasinh_data_internal(%rip), %xmm12 + addps %xmm3, %xmm11 + mulps %xmm10, %xmm2 + subps %xmm11, %xmm4 + movaps %xmm12, %xmm0 + addps %xmm3, %xmm4 + + /* + * Unfortunately, we can still be in trouble if |X| <= 2^-5, since + * the absolute error 2^-(7+24)-ish in sqrt(1 + X^2) gets scaled up + * by 1/X and comes close to our threshold. Hence if |X| <= 2^-4, + * perform an alternative computation + * sqrt(1 + X^2) - 1 = X^2/2 - X^4/8 + X^6/16 + * X2 = X^2 + */ + addps %xmm2, %xmm3 + addps %xmm2, %xmm4 + andps %xmm11, %xmm0 + + /* + * Compute R = 1/sqrt(Y + W) * (1 + d) + * Force R to <= 8 significant bits. + * This means that R * Y and R^2 * Y are exactly representable. + */ + rsqrtps %xmm0, %xmm14 + subps %xmm0, %xmm11 + andps %xmm12, %xmm14 + addps %xmm11, %xmm4 + + /* + * Compute S = (Y/sqrt(Y + W)) * (1 + d) + * and T = (W/sqrt(Y + W)) * (1 + d) + * so that S + T = sqrt(Y + W) * (1 + d) + * S is exact, and the rounding error in T is OK. + */ + mulps %xmm14, %xmm0 + mulps %xmm14, %xmm4 + + /* + * Get the absolute value of the input, since we will exploit antisymmetry + * and mostly assume X >= 0 in the core computation + */ + movups SgnMask+__svml_sasinh_data_internal(%rip), %xmm6 + + /* + * Compute e = -(2 * d + d^2) + * The first FMR is exact, and the rounding error in the other is acceptable + * since d and e are ~ 2^-8 + */ + movaps %xmm14, %xmm13 + andps %xmm8, %xmm6 + + /* + * Obtain sqrt(1 + X^2) - 1 in two pieces + * sqrt(1 + X^2) - 1 + * = sqrt(Y + W) - 1 + * = (S + T) * (1 + Corr) - 1 + * = [S - 1] + [T + (S + T) * Corr] + * We need a compensated summation for the last part. We treat S - 1 + * as the larger part; it certainly is until about X < 2^-4, and in that + * case, the error is affordable since X dominates over sqrt(1 + X^2) - 1 + * Final sum is dTmp5 (hi) + dTmp7 (lo) + */ + movaps %xmm0, %xmm1 + + /* + * Check whether the input is finite, by checking |X| <= MaxFloat + * Otherwise set the rangemask so that the callout will get used. + * Note that this will also use the callout for NaNs since not(NaN <= MaxFloat) + */ + movaps %xmm6, %xmm9 + + /* + * The following computation can go wrong for very large X, basically + * because X^2 overflows. But for large X we have + * asinh(X) / log(2 X) - 1 =~= 1/(4 * X^2), so for X >= 2^30 + * we can just later stick X back into the log and tweak up the exponent. + * Actually we scale X by 2^-30 and tweak the exponent up by 31, + * to stay in the safe range for the later log computation. + * Compute a flag now telling us when do do this. + */ + movaps %xmm6, %xmm5 + cmpnleps sLargestFinite+__svml_sasinh_data_internal(%rip), %xmm9 + cmpltps sBigThreshold+__svml_sasinh_data_internal(%rip), %xmm5 + mulps %xmm0, %xmm13 + addps %xmm4, %xmm1 + subps %xmm7, %xmm0 + mulps %xmm4, %xmm14 + movmskps %xmm9, %edx + movaps %xmm7, %xmm9 + + /* + * Now 1 / (1 + d) + * = 1 / (1 + (sqrt(1 - e) - 1)) + * = 1 / sqrt(1 - e) + * = 1 + 1/2 * e + 3/8 * e^2 + 5/16 * e^3 + 35/128 * e^4 + ... + * So compute the first three nonconstant terms of that, so that + * we have a relative correction (1 + Corr) to apply to S etc. + * C1 = 1/2 + * C2 = 3/8 + * C3 = 5/16 + */ + movups sC3+__svml_sasinh_data_internal(%rip), %xmm15 + subps %xmm13, %xmm9 + movups sHalf+__svml_sasinh_data_internal(%rip), %xmm10 + subps %xmm14, %xmm9 + + /* sX2over2 = X^2/2 */ + mulps %xmm10, %xmm3 + mulps %xmm9, %xmm15 + + /* sX46 = -X^4/4 + X^6/8 */ + movaps %xmm3, %xmm2 + movaps %xmm3, %xmm12 + + /* + * Now do another compensated sum to add |X| + [sqrt(1 + X^2) - 1]. + * It's always safe to assume |X| is larger. + * This is the final 2-part argument to the log1p function + */ + movaps %xmm6, %xmm14 + addps sC2+__svml_sasinh_data_internal(%rip), %xmm15 + mulps %xmm9, %xmm15 + addps %xmm10, %xmm15 + mulps %xmm15, %xmm9 + mulps %xmm1, %xmm9 + + /* Now multiplex to the case X = 2^-30 * input, Xl = sL = 0 in the "big" case. */ + movups XScale+__svml_sasinh_data_internal(%rip), %xmm15 + addps %xmm9, %xmm4 + movaps %xmm4, %xmm11 + addps %xmm0, %xmm11 + subps %xmm11, %xmm0 + addps %xmm0, %xmm4 + + /* sX4over4 = X^4/4 */ + movaps %xmm3, %xmm0 + mulps %xmm3, %xmm0 + mulps %xmm0, %xmm2 + subps %xmm0, %xmm2 + + /* + * Now we feed into the log1p code, using H in place of _VARG1 and + * also adding L into Xl. + * compute 1+x as high, low parts + */ + movaps %xmm7, %xmm0 + + /* sX46over2 = -X^4/8 + x^6/16 */ + mulps %xmm2, %xmm10 + movaps %xmm7, %xmm2 + addps %xmm10, %xmm12 + subps %xmm12, %xmm3 + addps %xmm3, %xmm10 + + /* Now multiplex the two possible computations */ + movaps %xmm6, %xmm3 + cmpleps sLittleThreshold+__svml_sasinh_data_internal(%rip), %xmm3 + movaps %xmm3, %xmm13 + andps %xmm3, %xmm12 + andnps %xmm11, %xmm13 + movaps %xmm3, %xmm1 + orps %xmm12, %xmm13 + andnps %xmm4, %xmm1 + andps %xmm3, %xmm10 + movaps %xmm6, %xmm4 + orps %xmm10, %xmm1 + addps %xmm13, %xmm14 + mulps %xmm15, %xmm6 + maxps %xmm14, %xmm0 + minps %xmm14, %xmm2 + subps %xmm14, %xmm4 + movaps %xmm0, %xmm3 + addps %xmm4, %xmm13 + addps %xmm2, %xmm3 + addps %xmm13, %xmm1 + subps %xmm3, %xmm0 + movaps %xmm5, %xmm4 + andps %xmm5, %xmm3 + andnps %xmm6, %xmm4 + addps %xmm0, %xmm2 + + /* + * Now resume the main code. + * reduction: compute r, n + */ + movdqu iBrkValue+__svml_sasinh_data_internal(%rip), %xmm6 + orps %xmm3, %xmm4 + psubd %xmm6, %xmm4 + movaps %xmm7, %xmm0 + addps %xmm2, %xmm1 + movdqu iOffExpoMask+__svml_sasinh_data_internal(%rip), %xmm2 + pand %xmm4, %xmm2 + psrad $23, %xmm4 + cvtdq2ps %xmm4, %xmm3 + pslld $23, %xmm4 + andps %xmm5, %xmm1 + paddd %xmm6, %xmm2 + psubd %xmm4, %xmm0 + mulps %xmm0, %xmm1 + + /* polynomial evaluation */ + subps %xmm7, %xmm2 + movups sPoly+112+__svml_sasinh_data_internal(%rip), %xmm7 + addps %xmm2, %xmm1 + mulps %xmm1, %xmm7 + movaps %xmm5, %xmm2 + + /* Add 31 to the exponent in the "large" case to get log(2 * input) */ + movups sThirtyOne+__svml_sasinh_data_internal(%rip), %xmm0 + addps sPoly+96+__svml_sasinh_data_internal(%rip), %xmm7 + addps %xmm3, %xmm0 + mulps %xmm1, %xmm7 + andnps %xmm0, %xmm2 + andps %xmm5, %xmm3 + orps %xmm3, %xmm2 + addps sPoly+80+__svml_sasinh_data_internal(%rip), %xmm7 + + /* final reconstruction */ + mulps sLn2+__svml_sasinh_data_internal(%rip), %xmm2 + mulps %xmm1, %xmm7 + + /* Finally, reincorporate the original sign. */ + movups sSign+__svml_sasinh_data_internal(%rip), %xmm0 + andps %xmm8, %xmm0 + addps sPoly+64+__svml_sasinh_data_internal(%rip), %xmm7 + mulps %xmm1, %xmm7 + addps sPoly+48+__svml_sasinh_data_internal(%rip), %xmm7 + mulps %xmm1, %xmm7 + addps sPoly+32+__svml_sasinh_data_internal(%rip), %xmm7 + mulps %xmm1, %xmm7 + addps sPoly+16+__svml_sasinh_data_internal(%rip), %xmm7 + mulps %xmm1, %xmm7 + addps sPoly+__svml_sasinh_data_internal(%rip), %xmm7 + mulps %xmm1, %xmm7 + mulps %xmm1, %xmm7 + addps %xmm7, %xmm1 + addps %xmm2, %xmm1 + pxor %xmm1, %xmm0 + testl %edx, %edx + + /* Go to special inputs processing branch */ + jne L(SPECIAL_VALUES_BRANCH) + # LOE rbx rbp r12 r13 r14 r15 edx xmm0 xmm8 + + /* Restore registers + * and exit the function + */ L(EXIT): - addq $72, %rsp - cfi_def_cfa_offset(8) - ret - cfi_def_cfa_offset(80) + addq $72, %rsp + cfi_def_cfa_offset(8) + ret + cfi_def_cfa_offset(80) -/* Branch to process - * special inputs - */ + /* Branch to process + * special inputs + */ L(SPECIAL_VALUES_BRANCH): - movups %xmm8, 32(%rsp) - movups %xmm0, 48(%rsp) - # LOE rbx rbp r12 r13 r14 r15 edx - - xorl %eax, %eax - movq %r12, 16(%rsp) - cfi_offset(12, -64) - movl %eax, %r12d - movq %r13, 8(%rsp) - cfi_offset(13, -72) - movl %edx, %r13d - movq %r14, (%rsp) - cfi_offset(14, -80) - # LOE rbx rbp r15 r12d r13d - -/* Range mask - * bits check - */ + movups %xmm8, 32(%rsp) + movups %xmm0, 48(%rsp) + # LOE rbx rbp r12 r13 r14 r15 edx + + xorl %eax, %eax + movq %r12, 16(%rsp) + cfi_offset(12, -64) + movl %eax, %r12d + movq %r13, 8(%rsp) + cfi_offset(13, -72) + movl %edx, %r13d + movq %r14, (%rsp) + cfi_offset(14, -80) + # LOE rbx rbp r15 r12d r13d + + /* Range mask + * bits check + */ L(RANGEMASK_CHECK): - btl %r12d, %r13d + btl %r12d, %r13d -/* Call scalar math function */ - jc L(SCALAR_MATH_CALL) - # LOE rbx rbp r15 r12d r13d + /* Call scalar math function */ + jc L(SCALAR_MATH_CALL) + # LOE rbx rbp r15 r12d r13d -/* Special inputs - * processing loop - */ + /* Special inputs + * processing loop + */ L(SPECIAL_VALUES_LOOP): - incl %r12d - cmpl $4, %r12d - -/* Check bits in range mask */ - jl L(RANGEMASK_CHECK) - # LOE rbx rbp r15 r12d r13d - - movq 16(%rsp), %r12 - cfi_restore(12) - movq 8(%rsp), %r13 - cfi_restore(13) - movq (%rsp), %r14 - cfi_restore(14) - movups 48(%rsp), %xmm0 - -/* Go to exit */ - jmp L(EXIT) - cfi_offset(12, -64) - cfi_offset(13, -72) - cfi_offset(14, -80) - # LOE rbx rbp r12 r13 r14 r15 xmm0 - -/* Scalar math fucntion call - * to process special input - */ + incl %r12d + cmpl $4, %r12d + + /* Check bits in range mask */ + jl L(RANGEMASK_CHECK) + # LOE rbx rbp r15 r12d r13d + + movq 16(%rsp), %r12 + cfi_restore(12) + movq 8(%rsp), %r13 + cfi_restore(13) + movq (%rsp), %r14 + cfi_restore(14) + movups 48(%rsp), %xmm0 + + /* Go to exit */ + jmp L(EXIT) + cfi_offset(12, -64) + cfi_offset(13, -72) + cfi_offset(14, -80) + # LOE rbx rbp r12 r13 r14 r15 xmm0 + + /* Scalar math fucntion call + * to process special input + */ L(SCALAR_MATH_CALL): - movl %r12d, %r14d - movss 32(%rsp,%r14,4), %xmm0 - call asinhf@PLT - # LOE rbx rbp r14 r15 r12d r13d xmm0 + movl %r12d, %r14d + movss 32(%rsp, %r14, 4), %xmm0 + call asinhf@PLT + # LOE rbx rbp r14 r15 r12d r13d xmm0 - movss %xmm0, 48(%rsp,%r14,4) + movss %xmm0, 48(%rsp, %r14, 4) -/* Process special inputs in loop */ - jmp L(SPECIAL_VALUES_LOOP) - # LOE rbx rbp r15 r12d r13d + /* Process special inputs in loop */ + jmp L(SPECIAL_VALUES_LOOP) + # LOE rbx rbp r15 r12d r13d END(_ZGVbN4v_asinhf_sse4) - .section .rodata, "a" - .align 16 + .section .rodata, "a" + .align 16 #ifdef __svml_sasinh_data_internal_typedef typedef unsigned int VUINT32; typedef struct { - __declspec(align(16)) VUINT32 SgnMask[4][1]; - __declspec(align(16)) VUINT32 sOne[4][1]; - __declspec(align(16)) VUINT32 sPoly[8][4][1]; - __declspec(align(16)) VUINT32 iBrkValue[4][1]; - __declspec(align(16)) VUINT32 iOffExpoMask[4][1]; - __declspec(align(16)) VUINT32 sBigThreshold[4][1]; - __declspec(align(16)) VUINT32 sC2[4][1]; - __declspec(align(16)) VUINT32 sC3[4][1]; - __declspec(align(16)) VUINT32 sHalf[4][1]; - __declspec(align(16)) VUINT32 sLargestFinite[4][1]; - __declspec(align(16)) VUINT32 sLittleThreshold[4][1]; - __declspec(align(16)) VUINT32 sSign[4][1]; - __declspec(align(16)) VUINT32 sThirtyOne[4][1]; - __declspec(align(16)) VUINT32 sTopMask11[4][1]; - __declspec(align(16)) VUINT32 sTopMask8[4][1]; - __declspec(align(16)) VUINT32 XScale[4][1]; - __declspec(align(16)) VUINT32 sLn2[4][1]; + __declspec(align(16)) VUINT32 SgnMask[4][1]; + __declspec(align(16)) VUINT32 sOne[4][1]; + __declspec(align(16)) VUINT32 sPoly[8][4][1]; + __declspec(align(16)) VUINT32 iBrkValue[4][1]; + __declspec(align(16)) VUINT32 iOffExpoMask[4][1]; + __declspec(align(16)) VUINT32 sBigThreshold[4][1]; + __declspec(align(16)) VUINT32 sC2[4][1]; + __declspec(align(16)) VUINT32 sC3[4][1]; + __declspec(align(16)) VUINT32 sHalf[4][1]; + __declspec(align(16)) VUINT32 sLargestFinite[4][1]; + __declspec(align(16)) VUINT32 sLittleThreshold[4][1]; + __declspec(align(16)) VUINT32 sSign[4][1]; + __declspec(align(16)) VUINT32 sThirtyOne[4][1]; + __declspec(align(16)) VUINT32 sTopMask11[4][1]; + __declspec(align(16)) VUINT32 sTopMask8[4][1]; + __declspec(align(16)) VUINT32 XScale[4][1]; + __declspec(align(16)) VUINT32 sLn2[4][1]; } __svml_sasinh_data_internal; #endif __svml_sasinh_data_internal: - /*== SgnMask ==*/ - .long 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff - /*== sOne = SP 1.0 ==*/ - .align 16 - .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 - /*== sPoly[] = SP polynomial ==*/ - .align 16 - .long 0xbf000000, 0xbf000000, 0xbf000000, 0xbf000000 /* -5.0000000000000000000000000e-01 P0 */ - .long 0x3eaaaa94, 0x3eaaaa94, 0x3eaaaa94, 0x3eaaaa94 /* 3.3333265781402587890625000e-01 P1 */ - .long 0xbe80058e, 0xbe80058e, 0xbe80058e, 0xbe80058e /* -2.5004237890243530273437500e-01 P2 */ - .long 0x3e4ce190, 0x3e4ce190, 0x3e4ce190, 0x3e4ce190 /* 2.0007920265197753906250000e-01 P3 */ - .long 0xbe28ad37, 0xbe28ad37, 0xbe28ad37, 0xbe28ad37 /* -1.6472326219081878662109375e-01 P4 */ - .long 0x3e0fcb12, 0x3e0fcb12, 0x3e0fcb12, 0x3e0fcb12 /* 1.4042308926582336425781250e-01 P5 */ - .long 0xbe1ad9e3, 0xbe1ad9e3, 0xbe1ad9e3, 0xbe1ad9e3 /* -1.5122179687023162841796875e-01 P6 */ - .long 0x3e0d84ed, 0x3e0d84ed, 0x3e0d84ed, 0x3e0d84ed /* 1.3820238411426544189453125e-01 P7 */ - /*== iBrkValue = SP 2/3 ==*/ - .align 16 - .long 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab - /*== iOffExpoMask = SP significand mask ==*/ - .align 16 - .long 0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff - /*== sBigThreshold ==*/ - .align 16 - .long 0x4E800000, 0x4E800000, 0x4E800000, 0x4E800000 - /*== sC2 ==*/ - .align 16 - .long 0x3EC00000, 0x3EC00000, 0x3EC00000, 0x3EC00000 - /*== sC3 ==*/ - .align 16 - .long 0x3EA00000, 0x3EA00000, 0x3EA00000, 0x3EA00000 - /*== sHalf ==*/ - .align 16 - .long 0x3F000000, 0x3F000000, 0x3F000000, 0x3F000000 - /*== sLargestFinite ==*/ - .align 16 - .long 0x7F7FFFFF, 0x7F7FFFFF, 0x7F7FFFFF, 0x7F7FFFFF - /*== sLittleThreshold ==*/ - .align 16 - .long 0x3D800000, 0x3D800000, 0x3D800000, 0x3D800000 - /*== sSign ==*/ - .align 16 - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 - /*== sThirtyOne ==*/ - .align 16 - .long 0x41F80000, 0x41F80000, 0x41F80000, 0x41F80000 - /*== sTopMask11 ==*/ - .align 16 - .long 0xFFFFE000, 0xFFFFE000, 0xFFFFE000, 0xFFFFE000 - /*== sTopMask8 ==*/ - .align 16 - .long 0xFFFF0000, 0xFFFF0000, 0xFFFF0000, 0xFFFF0000 - /*== XScale ==*/ - .align 16 - .long 0x30800000, 0x30800000, 0x30800000, 0x30800000 - /*== sLn2 = SP ln(2) ==*/ - .align 16 - .long 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218 - .align 16 - .type __svml_sasinh_data_internal,@object - .size __svml_sasinh_data_internal,.-__svml_sasinh_data_internal + /* SgnMask */ + .long 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff + /* sOne = SP 1.0 */ + .align 16 + .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 + /* sPoly[] = SP polynomial */ + .align 16 + .long 0xbf000000, 0xbf000000, 0xbf000000, 0xbf000000 /* -5.0000000000000000000000000e-01 P0 */ + .long 0x3eaaaa94, 0x3eaaaa94, 0x3eaaaa94, 0x3eaaaa94 /* 3.3333265781402587890625000e-01 P1 */ + .long 0xbe80058e, 0xbe80058e, 0xbe80058e, 0xbe80058e /* -2.5004237890243530273437500e-01 P2 */ + .long 0x3e4ce190, 0x3e4ce190, 0x3e4ce190, 0x3e4ce190 /* 2.0007920265197753906250000e-01 P3 */ + .long 0xbe28ad37, 0xbe28ad37, 0xbe28ad37, 0xbe28ad37 /* -1.6472326219081878662109375e-01 P4 */ + .long 0x3e0fcb12, 0x3e0fcb12, 0x3e0fcb12, 0x3e0fcb12 /* 1.4042308926582336425781250e-01 P5 */ + .long 0xbe1ad9e3, 0xbe1ad9e3, 0xbe1ad9e3, 0xbe1ad9e3 /* -1.5122179687023162841796875e-01 P6 */ + .long 0x3e0d84ed, 0x3e0d84ed, 0x3e0d84ed, 0x3e0d84ed /* 1.3820238411426544189453125e-01 P7 */ + /* iBrkValue = SP 2/3 */ + .align 16 + .long 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab + /* iOffExpoMask = SP significand mask */ + .align 16 + .long 0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff + /* sBigThreshold */ + .align 16 + .long 0x4E800000, 0x4E800000, 0x4E800000, 0x4E800000 + /* sC2 */ + .align 16 + .long 0x3EC00000, 0x3EC00000, 0x3EC00000, 0x3EC00000 + /* sC3 */ + .align 16 + .long 0x3EA00000, 0x3EA00000, 0x3EA00000, 0x3EA00000 + /* sHalf */ + .align 16 + .long 0x3F000000, 0x3F000000, 0x3F000000, 0x3F000000 + /* sLargestFinite */ + .align 16 + .long 0x7F7FFFFF, 0x7F7FFFFF, 0x7F7FFFFF, 0x7F7FFFFF + /* sLittleThreshold */ + .align 16 + .long 0x3D800000, 0x3D800000, 0x3D800000, 0x3D800000 + /* sSign */ + .align 16 + .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 + /* sThirtyOne */ + .align 16 + .long 0x41F80000, 0x41F80000, 0x41F80000, 0x41F80000 + /* sTopMask11 */ + .align 16 + .long 0xFFFFE000, 0xFFFFE000, 0xFFFFE000, 0xFFFFE000 + /* sTopMask8 */ + .align 16 + .long 0xFFFF0000, 0xFFFF0000, 0xFFFF0000, 0xFFFF0000 + /* XScale */ + .align 16 + .long 0x30800000, 0x30800000, 0x30800000, 0x30800000 + /* sLn2 = SP ln(2) */ + .align 16 + .long 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218 + .align 16 + .type __svml_sasinh_data_internal, @object + .size __svml_sasinh_data_internal, .-__svml_sasinh_data_internal