diff options
Diffstat (limited to 'sysdeps/ia64/fpu/s_logbl.S')
| -rw-r--r-- | sysdeps/ia64/fpu/s_logbl.S | 310 |
1 files changed, 146 insertions, 164 deletions
diff --git a/sysdeps/ia64/fpu/s_logbl.S b/sysdeps/ia64/fpu/s_logbl.S index 38b131f..e312c1b 100644 --- a/sysdeps/ia64/fpu/s_logbl.S +++ b/sysdeps/ia64/fpu/s_logbl.S @@ -1,10 +1,10 @@ .file "logbl.s" -// Copyright (C) 2000, 2001, Intel Corporation + +// Copyright (c) 2000 - 2003, 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. +// +// Contributed 2000 by the Intel Numerics Group, Intel Corporation // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are @@ -20,44 +20,46 @@ // * The name of Intel Corporation may not be used to endorse or promote // products derived from this software without specific prior written // permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + +// 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 +// 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 +// 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. -// +// 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. +// problem reports or change requests be submitted to it directly at +// http://www.intel.com/software/products/opensource/libraries/num.htm. // // History //============================================================== -// 2/02/00 Initial version -// 2/16/00 Modified to conform to C9X -// 3/16/00 Improved speed -// 4/04/00 Unwind support added -// 5/30/00 Fixed bug when x double-extended denormal -// 8/15/00 Bundle added after call to __libm_error_support to properly +// 02/02/00 Initial version +// 02/16/00 Modified to conform to C9X +// 03/16/00 Improved speed +// 04/04/00 Unwind support added +// 05/30/00 Fixed bug when x double-extended denormal +// 08/15/00 Bundle added after call to __libm_error_support to properly // set [the previously overwritten] GR_Parameter_RESULT. +// 05/20/02 Cleaned up namespace and sf0 syntax +// 01/20/03 Improved performance // // API //============================================================== -// long double logbl( long double x); +// long double logbl( long double x ); // // Overview of operation //============================================================== -// The logbl function extracts the exponent of x as an integer in -// floating-point format. +// The logbl function extracts the exponent of x as an integer in +// floating-point format. // logbl computes log2 of x as a long double // -// logbl is similar to ilogbl but differs in the following ways: +// logbl is similar to ilogbl but differs in the following ways: // +-inf // ilogbl: returns INT_MAX // logbl: returns +inf @@ -71,229 +73,209 @@ // // Registers used //============================================================== -// general registers used: -// ar.pfs r32 -// r33 -> r37 -// r38 -> r41 used as parameters to error path +// general registers used: +// r26 -> r38 +// r35 -> r38 used as parameters to error path // -// predicate registers used: +// predicate registers used: // p6, p7, p8 -// -// floating-point registers used: +// floating-point registers used: // f9, f10, f11 // f8, input -#include "libm_support.h" +rExpBias = r26 +rExpMask = r27 +rSignexp_x = r28 +rExp_x = r29 +rTrueExp_x = r30 +rExp_2to64 = r31 GR_SAVE_PFS = r32 -GR_SAVE_B0 = r34 -GR_SAVE_GP = r35 -GR_Parameter_X = r38 -GR_Parameter_Y = r39 -GR_Parameter_RESULT = r40 -GR_Parameter_TAG = r41 +GR_SAVE_B0 = r33 +GR_SAVE_GP = r34 -FR_X = f8 -FR_Y = f0 -FR_RESULT = f10 +GR_Parameter_X = r35 +GR_Parameter_Y = r36 +GR_Parameter_RESULT = r37 +GR_Parameter_TAG = r38 -.align 32 -.global logbl# +fExp_in_signif = f9 +fNorm_x = f10 +fFloat_Exp = f10 +f2to64 = f11 .section .text -.proc logbl# -.align 32 - +GLOBAL_LIBM_ENTRY(logbl) -logbl: - -// qnan snan inf norm unorm 0 -+ -// 0 0 0 0 1 0 11 -// 0 b -{ .mfi - alloc r32=ar.pfs,1,5,4,0 -(p0) fclass.m.unc p8,p0 = f8, 0x0b - nop.i 999 -} // X NORMAL -// r37 = exp(f8) - - 0xffff -// sig(f8) = r37 +// TrueExp_x = exp(f8) - 0xffff +// sig = TrueExp_x // f8 = convert_to_fp (sig)) { .mfi -(p0) getf.exp r35 = f8 -(p0) fnorm f10=f8 - nop.i 999 ;; + getf.exp rSignexp_x = f8 + fclass.m p8,p0 = f8, 0x0b // Test for x unorm + mov rExpBias = 0xffff // Exponent bias } - -// qnan snan inf norm unorm 0 -+ -// 1 1 1 0 0 0 11 -// e 3 -{ .mmf -(p0) mov r33 = 0xffff -(p0) mov r34 = 0x1ffff -(p0) fclass.m.unc p6,p0 = f8, 0xe3 ;; +{ .mfi + nop.m 0 + fnorm.s1 fNorm_x = f8 + mov rExpMask = 0x1ffff // Exponent mask } +;; +// Form signexp of 2^64 in case need to scale denormal { .mfb -(p0) and r36 = r35, r34 -(p0) fclass.m.unc p7,p0 = f8, 0x07 -(p8) br.cond.spnt L(LOGB_DENORM) ;; + mov rExp_2to64 = 0x1003f + fclass.m p6,p0 = f8, 0x1e3 // Test x natval, nan, inf +(p8) br.cond.spnt LOGB_DENORM // Branch if x unorm } +;; -{ .mib -(p0) sub r37 = r36, r33 - nop.i 999 -(p6) br.cond.spnt L(LOGB_NAN_INF) ;; +LOGB_COMMON: +// Return here from LOGB_DENORM +{ .mfi + and rExp_x = rSignexp_x, rExpMask // Get biased exponent + fclass.m p7,p0 = f8, 0x07 // Test x zero + nop.i 0 } +;; + +// X NAN or INFINITY, return f8 * f8 +{ .mfb + sub rTrueExp_x = rExp_x, rExpBias // Get true exponent +(p6) fma.s0 f8= f8,f8,f0 // Result if x natval, nan, inf +(p6) br.ret.spnt b0 // Exit if x natval, nan, inf +} +;; { .mib -(p0) setf.sig f9 = r37 + setf.sig fExp_in_signif = rTrueExp_x // Exponent as integer in fp nop.i 999 -(p7) br.cond.spnt L(LOGB_ZERO) ;; -} -{ .mfi - nop.m 999 -(p0) fcvt.xf f10 = f9 - nop.i 999 ;; +(p7) br.cond.spnt LOGB_ZERO } +;; +// Result can be represented in less than 24 bits, so no precision completer +// is needed. { .mfb - nop.m 999 -(p0) fnorm f8 = f10 -(p0) br.ret.sptk b0 ;; + nop.m 0 + fcvt.xf f8 = fExp_in_signif + br.ret.sptk b0 // Exit main path, 0 < |x| < inf } +;; -L(LOGB_DENORM): -// Form signexp of 2^64 in case need to scale denormal +LOGB_DENORM: +// Form 2^64 in case need to scale denormal // Check to see if double-extended denormal { .mfi -(p0) mov r38 = 0x1003f -(p0) fclass.m.unc p8,p0 = f10, 0x0b - nop.i 999 ;; + setf.exp f2to64 = rExp_2to64 + fclass.m p8,p0 = fNorm_x, 0x0b + nop.i 0 } +;; -// Form 2^64 in case need to scale denormal { .mfi -(p0) setf.exp f11 = r38 - nop.f 999 - nop.i 999 ;; + nop.m 0 + fcmp.eq.s0 p7,p0 = f8, f0 // Dummy op to set denormal flag + nop.i 0 } +;; // If double-extended denormal add 64 to exponent bias for scaling // If double-extended denormal form x * 2^64 which is normal { .mfi -(p8) add r33 = 64, r33 -(p8) fmpy f10 = f10, f11 - nop.i 999 ;; +(p8) add rExpBias = 64, rExpBias +(p8) fmpy.s1 fNorm_x = fNorm_x, f2to64 + nop.i 0 } +;; // Logic is the same as normal path but use normalized input -{ .mmi -(p0) getf.exp r35 = f10 ;; - nop.m 999 - nop.i 999 ;; +{ .mib + getf.exp rSignexp_x = fNorm_x + nop.i 0 + br.cond.sptk LOGB_COMMON // Return to main path } +;; -{ .mmi -(p0) and r36 = r35, r34 ;; -(p0) sub r37 = r36, r33 - nop.i 999 ;; -} +LOGB_ZERO: +// Here if x zero +// f10 = -|f8| +// f9 = 1.0/f10 = -1.0/|f8| = -inf -{ .mmi -(p0) setf.sig f9 = r37 - nop.m 999 - nop.i 999 ;; -} - -{ .mfi - nop.m 999 -(p0) fcvt.xf f10 = f9 - nop.i 999 ;; +{ .mmf + alloc r32=ar.pfs,1,2,4,0 + mov GR_Parameter_TAG = 150 // Error code + fmerge.ns f10 = f0,f8 } +;; { .mfb - nop.m 999 -(p0) fnorm f8 = f10 -(p0) br.ret.sptk b0 ;; + nop.m 0 + frcpa.s0 f9,p6 = f1,f10 // Produce -inf, Z flag + br.cond.sptk __libm_error_region // Call error support } +;; -L(LOGB_NAN_INF): - -// X NAN or INFINITY, return f8 * f8 -{ .mfb - nop.m 999 -(p0) fma f8= f8,f8,f0 -(p0) br.ret.sptk b0 ;; -} +GLOBAL_LIBM_END(logbl) -L(LOGB_ZERO): -{.mfi - nop.m 0 -(p0) frcpa.s0 f10,p6 = f1,f0 - nop.i 0 -};; -{.mfi - mov GR_Parameter_TAG = 150 -(p0) fms.s1 f10 = f0,f0,f10 - nop.i 0 -};; -// X ZERO -// return -1.0/fabs(f8)=-inf, set divide-by-zero flag, call error support -.endp logbl -ASM_SIZE_DIRECTIVE(logbl) -.proc __libm_error_region -__libm_error_region: +LOCAL_LIBM_ENTRY(__libm_error_region) .prologue + { .mfi - add GR_Parameter_Y=-32,sp // Parameter 2 value + 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 + mov GR_SAVE_PFS=ar.pfs // Save ar.pfs } { .mfi .fframe 64 - add sp=-64,sp // Create new stack + add sp=-64,sp // Create new stack nop.f 0 - mov GR_SAVE_GP=gp // Save gp + mov GR_SAVE_GP=gp // Save gp };; + { .mmi - stfe [GR_Parameter_Y] = FR_Y,16 // Save Parameter 2 on stack - add GR_Parameter_X = 16,sp // Parameter 1 address + stfe [GR_Parameter_Y] = f0,16 // 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 + mov GR_SAVE_B0=b0 // Save b0 };; + .body { .mib - stfe [GR_Parameter_X] = FR_X // Store Parameter 1 on stack - add GR_Parameter_RESULT = 0,GR_Parameter_Y - nop.b 0 // Parameter 3 address + stfe [GR_Parameter_X] = f8 // STORE Parameter 1 on stack + add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address + nop.b 0 } { .mib - stfe [GR_Parameter_Y] = FR_RESULT // Store Parameter 3 on stack + stfe [GR_Parameter_Y] = f9 // Store Parameter 3 on stack add GR_Parameter_Y = -16,GR_Parameter_Y - br.call.sptk b0=__libm_error_support# // Call error handling function + br.call.sptk b0=__libm_error_support# // Call error handling function };; + { .mmi - nop.m 0 - nop.m 0 add GR_Parameter_RESULT = 48,sp + nop.m 0 + nop.i 0 };; + { .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 + br.ret.sptk b0 };; -.endp __libm_error_region -ASM_SIZE_DIRECTIVE(__libm_error_region) +LOCAL_LIBM_END(__libm_error_region) + .type __libm_error_support#,@function .global __libm_error_support# |