diff options
Diffstat (limited to 'sysdeps/ia64/fpu/e_acosl.S')
| -rw-r--r-- | sysdeps/ia64/fpu/e_acosl.S | 1094 |
1 files changed, 1094 insertions, 0 deletions
diff --git a/sysdeps/ia64/fpu/e_acosl.S b/sysdeps/ia64/fpu/e_acosl.S new file mode 100644 index 0000000000..81f56e41c8 --- /dev/null +++ b/sysdeps/ia64/fpu/e_acosl.S @@ -0,0 +1,1094 @@ +.file "acosl.s" + +// Copyright (c) 2000, 2001, Intel Corporation +// All rights reserved. +// +// Contributed 2/2/2000 by John Harrison, Ted Kubaska, Bob Norin, Shane Story, +// and Ping Tak Peter Tang of the Computational Software Lab, Intel Corporation. +// +// WARRANTY DISCLAIMER +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS +// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, +// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, +// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR +// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY +// OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING +// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS +// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +// +// Intel Corporation is the author of this code, and requests that all +// problem reports or change requests be submitted to it directly at +// http://developer.intel.com/opensource. +// +// History +//============================================================== +// 2/02/00 Initial version +// 2/07/00 Modified calculation of acos_corr to correct acosl +// 4/04/00 Unwind support added +// 8/15/00 Bundle added after call to __libm_error_support to properly +// set [the previously overwritten] GR_Parameter_RESULT. +// 12/20/00 Set denormal flag properly. +// +// API +//============================================================== +// double-extended = acosl (double-extended) +// input floating point f8 +// output floating point f8 +// +// Registers used +//============================================================== +// +// predicate registers used: +// p6 -> p12 +// +// floating-point registers used: +// f8 has input, then output +// f8 -> f15, f32 ->f99 +// +// general registers used: +// r32 -> r48 +// +// Overview of operation +//============================================================== +// There are three paths +// 1. |x| < 2^-25 ACOS_TINY +// 2. 2^-25 <= |x| < 1/4 ACOS_POLY +// 3. 1/4 <= |x| < 1 ACOS_ATAN + +#include "libm_support.h" + +// Assembly macros +//============================================================== + +// f8 is input, but acos_V must be put in f8 +// when __libm_atan2_reg is called, f8 must get V +// f9 gets U when __libm_atan2_reg is called + + +// __libm_atan2_reg returns +// f8 = Z_hi +// f10 = Z_lo +// f11 = s_lo + +acos_Z_hi = f8 +acos_Z_lo = f10 +acos_S_lo = f11 + +// When we call __libm_atan2_reg, we must save +// the following: + +acos_corr = f12 +acos_X = f13 +acos_pi_hi = f14 +acos_pi_lo = f15 + +// The rest of the assembly macros + +acos_P79 = f32 +acos_P59 = f33 +acos_P39 = f34 +acos_P19 = f35 + +acos_P810 = f36 +acos_P610 = f37 +acos_P410 = f38 +acos_P210 = f39 + +acos_A1 = f41 +acos_A2 = f42 +acos_A3 = f43 +acos_A4 = f44 +acos_A5 = f45 +acos_A6 = f46 +acos_A7 = f47 +acos_A8 = f48 +acos_A9 = f49 +acos_A10 = f50 + +acos_X2 = f51 +acos_X4 = f52 + +acos_B = f53 +acos_Bb = f54 +acos_A = f55 +acos_Aa = f56 + +acos_1mA = f57 + +acos_W = f58 +acos_Ww = f59 + +acos_y0 = f60 +acos_y1 = f61 +acos_y2 = f62 + +acos_H = f63 +acos_Hh = f64 + +acos_t1 = f65 +acos_t2 = f66 +acos_t3 = f67 +acos_t4 = f68 +acos_t5 = f69 + +acos_Pseries = f70 +acos_NORM_f8 = f71 +acos_ABS_NORM_f8 = f72 + +acos_2 = f73 +acos_P1P2 = f74 +acos_HALF = f75 +acos_U = f76 + +acos_1mB = f77 +acos_V = f78 +acos_S = f79 + +acos_BmUU = f80 +acos_BmUUpb = f81 +acos_2U = f82 +acos_1d2U = f83 + +acos_Dd = f84 + +acos_pi_by_2_hi = f85 +acos_pi_by_2_lo = f86 +acos_xmpi_by_2_lo = f87 +acos_xPmw = f88 + +acos_Uu = f89 +acos_AmVV = f90 +acos_AmVVpa = f91 + +acos_2V = f92 +acos_1d2V = f93 +acos_Vv = f94 + +acos_Vu = f95 +acos_Uv = f96 + +acos_2_Z_hi = f97 +acos_s_lo_Z_lo = f98 +acos_result_lo = f99 + +acos_Z_hi = f8 +acos_Z_lo = f10 +acos_s_lo = f11 + +acos_GR_17_ones = r33 +acos_GR_16_ones = r34 +acos_GR_signexp_f8 = r35 +acos_GR_exp = r36 +acos_GR_true_exp = r37 +acos_GR_fffe = r38 + +GR_SAVE_PFS = r43 +GR_SAVE_B0 = r39 +GR_SAVE_GP = r41 + +// r40 is address of table of coefficients +// r42 + +GR_Parameter_X = r44 +GR_Parameter_Y = r45 +GR_Parameter_RESULT = r46 +GR_Parameter_TAG = r47 + + +// 2^-40: +// A true exponent of -40 is +// : -40 + register_bias +// : -28 + ffff = ffd7 + +// A true exponent of 1 is +// : 1 + register_bias +// : 1 + ffff = 10000 + +// Data tables +//============================================================== + +#ifdef _LIBC +.rodata +#else +.data +#endif + +.align 16 + +acos_coefficients: +ASM_TYPE_DIRECTIVE(acos_coefficients,@object) +data8 0xc90fdaa22168c234, 0x00003FFF // pi_by_2_hi +data8 0xc4c6628b80dc1cd1, 0x00003FBF // pi_by_2_lo +data8 0xc90fdaa22168c234, 0x00004000 // pi_hi +data8 0xc4c6628b80dc1cd1, 0x00003FC0 // pi_lo + +data8 0xBB08911F2013961E, 0x00003FF8 // A10 +data8 0x981F1095A23A87D3, 0x00003FF8 // A9 +data8 0xBDF09C6C4177BCC6, 0x00003FF8 // A8 +data8 0xE4C3A60B049ACCEA, 0x00003FF8 // A7 +data8 0x8E2789F4E8A8F1AD, 0x00003FF9 // A6 +data8 0xB745D09B2B0E850B, 0x00003FF9 // A5 +data8 0xF8E38E3BC4C50920, 0x00003FF9 // A4 +data8 0xB6DB6DB6D89FCD81, 0x00003FFA // A3 +data8 0x99999999999AF376, 0x00003FFB // A2 +data8 0xAAAAAAAAAAAAAA71, 0x00003FFC // A1 +ASM_SIZE_DIRECTIVE(acos_coefficients) + + +.align 32 +.global acosl# +ASM_TYPE_DIRECTIVE(acosl#,@function) + +.section .text +.proc acosl# +.align 32 + + +acosl: + +// After normalizing f8, get its true exponent +{ .mfi + alloc r32 = ar.pfs,1,11,4,0 +(p0) fnorm.s1 acos_NORM_f8 = f8 +(p0) mov acos_GR_17_ones = 0x1ffff +} + +{ .mmi +(p0) mov acos_GR_16_ones = 0xffff +(p0) addl r40 = @ltoff(acos_coefficients), gp + nop.i 999 +} +;; + +// Set denormal flag on denormal input with fcmp +{ .mfi + ld8 r40 = [r40] + fcmp.eq p6,p0 = f8,f0 + nop.i 999 +} +;; + + +// Load the constants pi_by_2 and pi. +// Each is stored as hi and lo values +// Also load the coefficients for ACOS_POLY + +{ .mmi +(p0) ldfe acos_pi_by_2_hi = [r40],16 ;; +(p0) ldfe acos_pi_by_2_lo = [r40],16 + nop.i 999 ;; +} + +{ .mmi +(p0) ldfe acos_pi_hi = [r40],16 ;; +(p0) ldfe acos_pi_lo = [r40],16 + nop.i 999 ;; +} + +{ .mmi +(p0) ldfe acos_A10 = [r40],16 ;; +(p0) ldfe acos_A9 = [r40],16 + nop.i 999 ;; +} + +// Take the absolute value of f8 +{ .mmf + nop.m 999 +(p0) getf.exp acos_GR_signexp_f8 = acos_NORM_f8 +(p0) fmerge.s acos_ABS_NORM_f8 = f0, acos_NORM_f8 +} + +{ .mii +(p0) ldfe acos_A8 = [r40],16 + nop.i 999 ;; +(p0) and acos_GR_exp = acos_GR_signexp_f8, acos_GR_17_ones ;; +} + +// case 1: |x| < 2^-25 ==> p6 ACOS_TINY +// case 2: 2^-25 <= |x| < 2^-2 ==> p8 ACOS_POLY +// case 3: 2^-2 <= |x| < 1 ==> p9 ACOS_ATAN +// case 4: 1 <= |x| ==> p11 ACOS_ERROR_RETURN +// Admittedly |x| = 1 is not an error but this is where that case is +// handled. + +{ .mii +(p0) ldfe acos_A7 = [r40],16 +(p0) sub acos_GR_true_exp = acos_GR_exp, acos_GR_16_ones ;; +(p0) cmp.ge.unc p6, p7 = -26, acos_GR_true_exp ;; +} + +{ .mii +(p0) ldfe acos_A6 = [r40],16 +(p7) cmp.ge.unc p8, p9 = -3, acos_GR_true_exp ;; +(p9) cmp.ge.unc p10, p11 = -1, acos_GR_true_exp +} + +{ .mmi +(p0) ldfe acos_A5 = [r40],16 ;; +(p0) ldfe acos_A4 = [r40],16 + nop.i 999 ;; +} + +{ .mmi +(p0) ldfe acos_A3 = [r40],16 ;; +(p0) ldfe acos_A2 = [r40],16 + nop.i 999 ;; +} + +// ACOS_ERROR_RETURN ==> p11 is true +// case 4: |x| >= 1 +{ .mib +(p0) ldfe acos_A1 = [r40],16 + nop.i 999 +(p11) br.spnt L(ACOS_ERROR_RETURN) ;; +} + +// ACOS_TINY ==> p6 is true +// case 1: |x| < 2^-25 +{ .mfi + nop.m 999 +(p6) fms.s1 acos_xmpi_by_2_lo = acos_NORM_f8,f1, acos_pi_by_2_lo + nop.i 999 ;; +} + +{ .mfb + nop.m 999 +(p6) fms.s0 f8 = acos_pi_by_2_hi,f1, acos_xmpi_by_2_lo +(p6) br.ret.spnt b0 ;; +} + + + +// ACOS_POLY ==> p8 is true +// case 2: 2^-25 <= |x| < 2^-2 +{ .mfi + nop.m 999 +(p8) fms.s1 acos_W = acos_pi_by_2_hi, f1, acos_NORM_f8 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s1 acos_X2 = f8,f8, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fms.s1 acos_Ww = acos_pi_by_2_hi, f1, acos_W + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s1 acos_X4 = acos_X2,acos_X2, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fms.s1 acos_Ww = acos_Ww, f1, acos_NORM_f8 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P810 = acos_X4, acos_A10, acos_A8 + nop.i 999 +} + +// acos_P79 = X4*A9 + A7 +// acos_P810 = X4*A10 + A8 +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P79 = acos_X4, acos_A9, acos_A7 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s1 acos_Ww = acos_Ww, f1, acos_pi_by_2_lo + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P610 = acos_X4, acos_P810, acos_A6 + nop.i 999 +} + + +// acos_P59 = X4*(X4*A9 + A7) + A5 +// acos_P610 = X4*(X4*A10 + A8) + A6 +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P59 = acos_X4, acos_P79, acos_A5 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P410 = acos_X4, acos_P610, acos_A4 + nop.i 999 +} + +// acos_P39 = X4*(X4*(X4*A9 + A7) + A5) + A3 +// acos_P410 = X4*(X4*(X4*A10 + A8) + A6) + A4 +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P39 = acos_X4, acos_P59, acos_A3 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P210 = acos_X4, acos_P410, acos_A2 + nop.i 999 +} + +// acos_P19 = X4*(X4*(X4*(X4*A9 + A7) + A5) + A3) + A1 = P1 +// acos_P210 = X4*(X4*(X4*(X4*A10 + A8) + A6) + A4) + A2 = P2 +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P19 = acos_X4, acos_P39, acos_A1 + nop.i 999 ;; +} + +// acos_P1P2 = Xsq*P2 + P1 +// acos_P1P2 = Xsq*(Xsq*P2 + P1) +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P1P2 = acos_X2, acos_P210, acos_P19 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s1 acos_P1P2 = acos_X2, acos_P1P2, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fms.s1 acos_xPmw = acos_NORM_f8, acos_P1P2, acos_Ww + nop.i 999 ;; +} + +{ .mfb + nop.m 999 +(p8) fms.s0 f8 = acos_W, f1, acos_xPmw +(p8) br.ret.spnt b0 ;; +} + + +// ACOS_ATAN +// case 3: 2^-2 <= |x| < 1 +// case 3: 2^-2 <= |x| < 1 ==> p9 ACOS_ATAN + +// Step 1.1: Get A,B and a,b +// A + a = 1- |X| +// B + b = 1+ |X| +// Note also that we will use acos_corr (f13) +// and acos_W + +// Step 2 +// Call __libm_atan2_reg + + +{ .mfi +(p0) mov acos_GR_fffe = 0xfffe +(p0) fma.s1 acos_B = f1,f1, acos_ABS_NORM_f8 +(p0) mov GR_SAVE_B0 = b0 ;; +} + +{ .mmf +(p0) mov GR_SAVE_GP = gp + nop.m 999 +(p0) fms.s1 acos_A = f1,f1, acos_ABS_NORM_f8 +} + +{ .mfi +(p0) setf.exp acos_HALF = acos_GR_fffe + nop.f 999 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fms.s1 acos_1mB = f1,f1, acos_B + nop.i 999 ;; +} + +// We want atan2(V,U) +// so put V in f8 and U in f9 +// but save X in acos_X + +{ .mfi + nop.m 999 +(p0) fmerge.se acos_X = f8, f8 + nop.i 999 ;; +} + +// Step 1.2: +///////////////////////// +// Get U = sqrt(B) +///////////////////////// + +{ .mfi + nop.m 999 +(p0) frsqrta.s1 acos_y0,p8 = acos_B + nop.i 999 +} + +{ .mfi + nop.m 999 +(p0) fms.s1 acos_1mA = f1,f1, acos_A + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_Bb = acos_1mB,f1, acos_ABS_NORM_f8 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_Hh = acos_HALF, acos_B, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_t1 = acos_y0, acos_y0, f0 + nop.i 999 +} + +{ .mfi + nop.m 999 +(p0) fms.s1 acos_Aa = acos_1mA,f1, acos_ABS_NORM_f8 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fnma.s1 acos_t2 = acos_t1, acos_Hh, acos_HALF + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_y1 = acos_t2, acos_y0, acos_y0 + nop.i 999 +} + + +// Step 1.2: +///////////////////////// +// Get V = sqrt(A) +///////////////////////// +{ .mfi + nop.m 999 +(p0) frsqrta.s1 acos_y0,p8 = acos_A + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_t3 = acos_y1, acos_Hh, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_t1 = acos_y0, acos_y0, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fnma.s1 acos_t4 = acos_t3, acos_y1, acos_HALF + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_y2 = acos_t4, acos_y1, acos_y1 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_S = acos_B, acos_y2, f0 + nop.i 999 +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_H = acos_y2, acos_HALF, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_t5 = acos_Hh, acos_y2, f0 + nop.i 999 +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_Hh = acos_HALF, acos_A, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fnma.s1 acos_Dd = acos_S, acos_S, acos_B + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fnma.s1 acos_t2 = acos_t1, acos_Hh, acos_HALF + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_U = acos_Dd, acos_H, acos_S + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_y1 = acos_t2, acos_y0, acos_y0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_2U = acos_U, f1, acos_U + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_t3 = acos_y1, acos_Hh, f0 + nop.i 999 +} + + +// Step 1.3: +// sqrt(A + a) = V + v +// sqrt(B + b) = U + u + +///////////////////////// +// Get u +///////////////////////// + +// acos_BmUU = B - UU +// acos_BmUUpb = (B - UU) + b + +{ .mfi + nop.m 999 +(p0) fnma.s1 acos_BmUU = acos_U, acos_U, acos_B + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fmerge.se f9 = acos_U, acos_U + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fnma.s1 acos_t4 = acos_t3, acos_y1, acos_HALF + nop.i 999 ;; +} + +// acos_1d2U = frcpa(2U) +{ .mfi + nop.m 999 +(p0) frcpa.s1 acos_1d2U,p9 = f1, acos_2U + nop.i 999 +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_BmUUpb = acos_BmUU, f1, acos_Bb + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_y2 = acos_t4, acos_y1, acos_y1 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +// acos_Uu = ((B - UU) + b) * frcpa(2U) +(p0) fma.s1 acos_Uu = acos_BmUUpb, acos_1d2U, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_S = acos_A, acos_y2, f0 + nop.i 999 +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_H = acos_y2, acos_HALF, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_t5 = acos_Hh, acos_y2, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fnma.s1 acos_Dd = acos_S, acos_S, acos_A + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_V = acos_Dd, acos_H, acos_S + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_2V = acos_V, f1, acos_V + nop.i 999 +} + +// Step 3 +///////////////////////// +// Calculate the correction, acos_corr +///////////////////////// +// acos_corr = U*v - (V*u) + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_Vu = acos_V,acos_Uu, f0 + nop.i 999 ;; +} + +///////////////////////// +// Get v +///////////////////////// +// acos_AmVV = A - VV +// acos_AmVVpa = (A - VV) + a + +{ .mfi + nop.m 999 +(p0) fnma.s1 acos_AmVV = acos_V, acos_V, acos_A + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fmerge.se f8 = acos_V, acos_V + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_AmVVpa = acos_AmVV, f1, acos_Aa + nop.i 999 ;; +} + +// acos_1d2V = frcpa(2V) +{ .mfi + nop.m 999 +(p0) frcpa.s1 acos_1d2V,p9 = f1, acos_2V + nop.i 999 ;; +} + +// acos_Vv = ((A - VV) + a) * frcpa(2V) +{ .mfi + nop.m 999 +(p0) fma.s1 acos_Vv = acos_AmVVpa, acos_1d2V, f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_Uv = acos_U,acos_Vv, f0 + nop.i 999 ;; +} + + +.endp acosl# +ASM_SIZE_DIRECTIVE(acosl#) + + +.proc __libm_callout +__libm_callout: +.prologue +{ .mfi + nop.m 0 + nop.f 0 +.save ar.pfs,GR_SAVE_PFS + mov GR_SAVE_PFS=ar.pfs +} +;; + +{ .mfi + mov GR_SAVE_GP=gp + nop.f 0 +.save b0, GR_SAVE_B0 + mov GR_SAVE_B0=b0 +} + +.body +{ .mfb + nop.m 999 +(p0) fms.s1 acos_corr = acos_Uv,f1, acos_Vu +(p0) br.call.sptk.many b0=__libm_atan2_reg# ;; +} + + +// p6 ==> X is negative +// p7 ==> x is positive +// We know that |X| >= 1/4 + +{ .mfi +(p0) mov gp = GR_SAVE_GP +(p0) fcmp.lt.unc p6,p7 = acos_X , f0 +(p0) mov b0 = GR_SAVE_B0 ;; +} + +// acos_2_Z_hi = 2 * acos_Z_hi +// acos_s_lo_Z_lo = s_lo * Z_lo + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_2_Z_hi = acos_Z_hi, f1, acos_Z_hi +(p0) mov ar.pfs = GR_SAVE_PFS +} + +{ .mfi + nop.m 999 +(p0) fma.s1 acos_s_lo_Z_lo = acos_s_lo, acos_Z_lo, f0 + nop.i 999 ;; +} + +// 2 is a constant needed later +{ .mfi + nop.m 999 +(p0) fma.s1 acos_2 = f1,f1,f1 + nop.i 999 ;; +} + +// X >= 1/4 +// acos_result_lo = 2(s_lo * Z_lo) - corr +// f8 = (2*Z_hi) + (2(s_lo * Z_lo) - corr) + +{ .mfi + nop.m 999 +(p7) fma.s1 acos_result_lo = acos_s_lo_Z_lo, acos_2, acos_corr + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p7) fma.s0 f8 = acos_2_Z_hi, f1, acos_result_lo + nop.i 999 +} + +// acos_result_lo = (pi_lo - corr) +// acos_result_lo = (pi_lo - corr) + acos_Ww +{ .mfi + nop.m 999 +(p6) fms.s1 acos_result_lo = acos_pi_lo, f1, acos_corr + nop.i 999 ;; +} + +// X <= -1/4 +// acos_W = pi_hi - 2 * Z_hi +{ .mfi + nop.m 999 +(p6) fnma.s1 acos_W = acos_2, acos_Z_hi, acos_pi_hi + nop.i 999 ;; +} + +// acos_Ww = pi_hi - W +// acos_Ww = (pi_hi - W) + (2 * Z_hi) +{ .mfi + nop.m 999 +(p6) fms.s1 acos_Ww = acos_pi_hi, f1, acos_W + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p6) fms.s1 acos_Ww = acos_Ww, f1, acos_2_Z_hi + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p6) fma.s1 acos_result_lo = acos_result_lo, f1, acos_Ww + nop.i 999 ;; +} + +// acos_Z_lo = ((pi_lo - corr) + acos_Ww) - 2 * (s_lo * Z_lo) +{ .mfi + nop.m 999 +(p6) fnma.s1 acos_Z_lo = acos_s_lo_Z_lo, acos_2, acos_result_lo + nop.i 999 ;; +} + +{ .mfb + nop.m 999 +(p6) fma.s0 f8 = acos_W, f1, acos_Z_lo +(p0) br.ret.sptk b0 ;; +} +.endp __libm_callout +ASM_SIZE_DIRECTIVE(__libm_callout) + +.proc SPECIAL +SPECIAL: +L(ACOS_NAN): +{ .mfb + nop.m 999 +(p0) fma.s0 f8 = f8,f1,f0 +(p0) br.ret.sptk b0 ;; +} + +L(ACOS_ERROR_RETURN): +// Save ar.pfs, b0, and gp; restore on exit + +// qnan snan inf norm unorm 0 -+ +// 1 1 0 0 0 0 11 = 0xc3 + +// Coming in as X = +- 1 +// What should we return? + +// If X is 1, return (sign of X)pi/2 + + +{ .mfi + nop.m 999 +(p0) fcmp.eq.unc p6,p7 = acos_ABS_NORM_f8,f1 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p6) fcmp.lt.unc p8,p9 = f8,f0 + nop.i 999 ;; +} + +{ .mfi + nop.m 999 +(p8) fma.s0 f8 = acos_pi_hi, f1, acos_pi_lo + nop.i 999 +} + +{ .mfb + nop.m 999 +(p9) fmerge.s f8 = f8,f0 +(p6) br.ret.spnt b0 ;; +} + +// If X is a NAN, leave +{ .mfi + nop.m 999 +(p0) fclass.m.unc p12,p0 = f8, 0xc3 + nop.i 999 ;; +} + +{ .mfb + nop.m 999 +(p12) fma.s0 f8 = f8,f1,f0 +(p12) br.ret.spnt b0 ;; +} + +{ .mfi +(p0) mov GR_Parameter_TAG = 57 +(p0) frcpa f10, p6 = f0, f0 +nop.i 999 +};; + +.endp SPECIAL +ASM_SIZE_DIRECTIVE(SPECIAL) + +.proc __libm_error_region +__libm_error_region: +.prologue +// (1) +{ .mfi + add GR_Parameter_Y=-32,sp // Parameter 2 value + nop.f 0 +.save ar.pfs,GR_SAVE_PFS + mov GR_SAVE_PFS=ar.pfs // Save ar.pfs +} +{ .mfi +.fframe 64 + add sp=-64,sp // Create new stack + nop.f 0 + mov GR_SAVE_GP=gp // Save gp +};; + + +// (2) +{ .mmi + stfe [GR_Parameter_Y] = f1,16 // Store Parameter 2 on stack + add GR_Parameter_X = 16,sp // Parameter 1 address +.save b0, GR_SAVE_B0 + mov GR_SAVE_B0=b0 // Save b0 +};; + +.body +// (3) +{ .mib + stfe [GR_Parameter_X] = f8 // Store Parameter 1 on stack + add GR_Parameter_RESULT = 0,GR_Parameter_Y + nop.b 0 // Parameter 3 address +} +{ .mib + stfe [GR_Parameter_Y] = f10 // Store Parameter 3 on stack + add GR_Parameter_Y = -16,GR_Parameter_Y + br.call.sptk b0=__libm_error_support# // Call error handling function +};; +{ .mmi + nop.m 0 + nop.m 0 + add GR_Parameter_RESULT = 48,sp +};; + +// (4) +{ .mmi + ldfe f8 = [GR_Parameter_RESULT] // Get return result off stack +.restore sp + add sp = 64,sp // Restore stack pointer + mov b0 = GR_SAVE_B0 // Restore return address +};; + +{ .mib + mov gp = GR_SAVE_GP // Restore gp + mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs + br.ret.sptk b0 // Return +};; + +.endp __libm_error_region +ASM_SIZE_DIRECTIVE(__libm_error_region) + +.type __libm_error_support#,@function +.global __libm_error_support# + +.type __libm_atan2_reg#,@function +.global __libm_atan2_reg# |