diff options
Diffstat (limited to 'sysdeps/ia64/fpu/e_fmodl.S')
| -rw-r--r-- | sysdeps/ia64/fpu/e_fmodl.S | 221 |
1 files changed, 113 insertions, 108 deletions
diff --git a/sysdeps/ia64/fpu/e_fmodl.S b/sysdeps/ia64/fpu/e_fmodl.S index 85c9f6ef82..da08ae3f5c 100644 --- a/sysdeps/ia64/fpu/e_fmodl.S +++ b/sysdeps/ia64/fpu/e_fmodl.S @@ -1,11 +1,10 @@ .file "fmodl.s" -// Copyright (C) 2000, 2001, Intel Corporation + +// Copyright (c) 2000 - 2003, Intel Corporation // All rights reserved. // -// Contributed 2/2/2000 by John Harrison, Cristina Iordache, 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 @@ -21,7 +20,7 @@ // * 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 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR @@ -36,38 +35,42 @@ // // 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. +// http://www.intel.com/software/products/opensource/libraries/num.htm. // // History //==================================================================== -// 2/02/00 Initial version -// 3/02/00 New Algorithm -// 4/04/00 Unwind support added -// 8/15/00 Bundle added after call to __libm_error_support to properly +// 02/02/00 Initial version +// 03/02/00 New Algorithm +// 04/04/00 Unwind support added +// 08/15/00 Bundle added after call to __libm_error_support to properly // set [the previously overwritten] GR_Parameter_RESULT. -//11/28/00 Set FR_Y to f9 +// 11/28/00 Set FR_Y to f9 +// 03/11/02 Fixed flags for fmodl(qnan,zero) +// 05/20/02 Cleaned up namespace and sf0 syntax +// 02/10/03 Reordered header: .section, .global, .proc, .align +// 04/28/03 Fix: fmod(sNaN,0) no longer sets errno // // API //==================================================================== -// long double fmodl(long double,long double); +// long double fmodl(long double,long double); // // Overview of operation //==================================================================== // fmod(a,b)=a-i*b, -// where i is an integer such that, if b!=0, +// where i is an integer such that, if b!=0, // |i|<|a/b| and |a/b-i|<1 // // Algorithm //==================================================================== // a). if |a|<|b|, return a -// b). get quotient and reciprocal overestimates accurate to +// b). get quotient and reciprocal overestimates accurate to // 33 bits (q2,y2) // c). if the exponent difference (exponent(a)-exponent(b)) // is less than 32, truncate quotient to integer and // finish in one iteration // d). if exponent(a)-exponent(b)>=32 (q2>=2^32) // round quotient estimate to single precision (k=RN(q2)), -// calculate partial remainder (a'=a-k*b), +// calculate partial remainder (a'=a-k*b), // get quotient estimate (a'*y2), and repeat from c). // // Registers used @@ -76,13 +79,9 @@ // General registers: r2,r29,r32 (ar.pfs), r33-r39 // Floating point registers: f6-f15 -#include "libm_support.h" - -.section .text - GR_SAVE_B0 = r33 GR_SAVE_PFS = r34 -GR_SAVE_GP = r35 +GR_SAVE_GP = r35 GR_SAVE_SP = r36 GR_Parameter_X = r37 @@ -95,18 +94,9 @@ FR_Y = f9 FR_RESULT = f8 +.section .text +GLOBAL_IEEE754_ENTRY(fmodl) -.proc fmodl# -.align 32 -.global fmodl# -.align 32 - -fmodl: -#ifdef _LIBC -.global __ieee754_fmodl -.type __ieee754_fmodl,@function -__ieee754_fmodl: -#endif // inputs in f8, f9 // result in f8 @@ -128,7 +118,7 @@ __ieee754_fmodl: // (1) y0 frcpa.s1 f10,p6=f6,f7 nop.i 0;; -} +} // eliminate special cases {.mmi @@ -141,7 +131,7 @@ cmp.eq p7,p10=r29,r0;; // Y +-NAN, +-inf, +-0? p7 { .mfi nop.m 999 -(p10) fclass.m p7,p10 = f9, 0xe7 +(p10) fclass.m p7,p10 = f9, 0xe7 nop.i 999;; } @@ -152,14 +142,14 @@ cmp.eq p7,p10=r29,r0;; { .mfi nop.m 999 -(p0) fclass.m.unc p9,p11 = f8, 0xe3 - nop.i 999 + fclass.m.unc p9,p11 = f8, 0xe3 + nop.i 999 } // |x| < |y|? Return x p8 { .mfi nop.m 999 -(p10) fcmp.lt.unc.s1 p8,p0 = f6,f7 +(p10) fcmp.lt.unc.s1 p8,p0 = f6,f7 nop.i 999 ;; } @@ -173,13 +163,13 @@ cmp.eq p7,p10=r29,r0;; // (3) e0 = 1 - b * y0 (p6) fnma.s1 f12=f7,f10,f1 nop.i 0;; -} +} // Y +-NAN, +-inf, +-0? p7 { .mfi nop.m 999 - // pseudo-NaN ? -(p10) fclass.nm p7,p0 = f9, 0xff + // pseudo-NaN ? +(p10) fclass.nm p7,p0 = f9, 0xff nop.i 999 } @@ -190,7 +180,7 @@ cmp.eq p7,p10=r29,r0;; { .mfi nop.m 999 -(p11) fclass.nm p9,p0 = f8, 0xff +(p11) fclass.nm p9,p0 = f8, 0xff nop.i 999;; } @@ -209,18 +199,18 @@ cmp.eq p7,p10=r29,r0;; nop.i 0 } {.bbb - (p9) br.cond.spnt L(FMOD_X_NAN_INF) - (p7) br.cond.spnt L(FMOD_Y_NAN_INF_ZERO) + (p9) br.cond.spnt FMOD_X_NAN_INF + (p7) br.cond.spnt FMOD_Y_NAN_INF_ZERO // if |x|<|y|, return (p8) br.ret.spnt b0;; } - {.mfi + {.mfi nop.m 0 // x denormal ? set D flag fnma.s0 f32=f6,f1,f6 nop.i 0 -} +} {.mfi nop.m 0 // y denormal ? set D flag @@ -234,46 +224,46 @@ cmp.eq p7,p10=r29,r0;; // (4) q1=q0+e0*q0 (p6) fma.s1 f13=f12,f13,f13 nop.i 0 -} +} { .mfi nop.m 0 // (5) e1 = e0 * e0 + 2^-34 (p6) fma.s1 f14=f12,f12,f11 nop.i 0;; -} +} {.mlx nop.m 0 movl r2=0x33a00000;; -} +} { .mfi nop.m 0 // (6) y1 = y0 + e0 * y0 (p6) fma.s1 f10=f12,f10,f10 nop.i 0;; -} +} {.mfi // set f12=1.25*2^{-24} setf.s f12=r2 // (7) q2=q1+e1*q1 (p6) fma.s1 f13=f13,f14,f13 nop.i 0;; -} +} {.mfi nop.m 0 fmerge.s f9=f8,f9 nop.i 0 -} +} { .mfi nop.m 0 // (8) y2 = y1 + e1 * y1 (p6) fma.s1 f10=f14,f10,f10 // set p6=0, p10=0 cmp.ne.and p6,p10=r0,r0;; -} +} .align 32 -L(loop64): +loop64: {.mfi nop.m 0 // compare q2, 2^32 @@ -305,7 +295,7 @@ L(loop64): // normalize truncated quotient (p8) fcvt.xf f13=f11 nop.i 0;; -} +} { .mfi nop.m 0 // calculate remainder (assuming f13=RZ(Q)) @@ -314,7 +304,7 @@ L(loop64): } {.mfi nop.m 0 - // also if exponent>32, round quotient to single precision + // also if exponent>32, round quotient to single precision // and subtract 1 ulp: q=q-q*(1.25*2^{-24}) (p7) fnma.s.s1 f11=f13,f12,f13 nop.i 0;; @@ -357,7 +347,7 @@ L(loop64): .pred.rel "mutex",p6,p10 {.mfb nop.m 0 - // add b to estimated remainder (to cover the case when the quotient was overestimated) + // add b to estimated remainder (to cover the case when the quotient was overestimated) // also set correct sign by using f9=|b|*sgn(a), f12=sgn(a) (p6) fma.s0 f8=f11,f12,f9 nop.b 0 @@ -378,43 +368,59 @@ L(loop64): nop.m 0 // if f14 was RZ(Q), set remainder to f14 (p9) mov f6=f14 - br.cond.sptk L(loop64);; + br.cond.sptk loop64;; } -L(FMOD_X_NAN_INF): +FMOD_X_NAN_INF: // Y zero ? -{.mfi +{.mfi + nop.m 0 + fclass.m p10,p0=f8,0xc3 // Test x=nan + nop.i 0 +} +{.mfi nop.m 0 fma.s1 f10=f9,f1,f0 nop.i 0;; } + +{.mfi + nop.m 0 + fma.s0 f8=f8,f1,f0 + nop.i 0 +} {.mfi + nop.m 0 +(p10) fclass.m p10,p0=f9,0x07 // Test x=nan, and y=zero + nop.i 0;; +} +{.mfb nop.m 0 fcmp.eq.unc.s1 p11,p0=f10,f0 - nop.i 0;; +(p10) br.ret.spnt b0;; // Exit with result=x if x=nan and y=zero } {.mib nop.m 0 nop.i 0 // if Y zero - (p11) br.cond.spnt L(FMOD_Y_ZERO);; + (p11) br.cond.spnt FMOD_Y_ZERO;; } // X infinity? Return QNAN indefinite { .mfi - // set p7 t0 0 - cmp.ne p7,p0=r0,r0 -(p0) fclass.m.unc p8,p9 = f8, 0x23 - nop.i 999;; + // set p7 t0 0 + cmp.ne p7,p0=r0,r0 + fclass.m.unc p8,p9 = f8, 0x23 + nop.i 999;; } // Y NaN ? {.mfi nop.m 999 (p8) fclass.m p9,p8=f9,0xc3 - nop.i 0;; + nop.i 0;; } // Y not pseudo-zero ? (r29 holds significand) {.mii @@ -423,63 +429,63 @@ L(FMOD_X_NAN_INF): nop.i 0;; } {.mfi - nop.m 999 -(p8) frcpa.s0 f8,p0 = f8,f8 + nop.m 999 +(p8) frcpa.s0 f8,p0 = f8,f8 nop.i 0 -} +} { .mfi nop.m 999 - // also set Denormal flag if necessary + // also set Denormal flag if necessary (p7) fnma.s0 f9=f9,f1,f9 nop.i 999 ;; } { .mfb nop.m 999 -(p8) fma.s0 f8=f8,f1,f0 - nop.b 999 ;; +(p8) fma.s0 f8=f8,f1,f0 + nop.b 999 ;; } { .mfb nop.m 999 -(p9) frcpa.s0 f8,p7=f8,f9 - br.ret.sptk b0 ;; +(p9) frcpa.s0 f8,p7=f8,f9 + br.ret.sptk b0 ;; } -L(FMOD_Y_NAN_INF_ZERO): +FMOD_Y_NAN_INF_ZERO: // Y INF { .mfi nop.m 999 -(p0) fclass.m.unc p7,p0 = f9, 0x23 + fclass.m.unc p7,p0 = f9, 0x23 nop.i 999 ;; } { .mfb nop.m 999 -(p7) fma f8=f8,f1,f0 -(p7) br.ret.spnt b0 ;; +(p7) fma.s0 f8=f8,f1,f0 +(p7) br.ret.spnt b0 ;; } // Y NAN? { .mfi nop.m 999 -(p0) fclass.m.unc p9,p10 = f9, 0xc3 + fclass.m.unc p9,p10 = f9, 0xc3 nop.i 999 ;; } { .mfi nop.m 999 -(p10) fclass.nm p9,p0 = f9, 0xff +(p10) fclass.nm p9,p0 = f9, 0xff nop.i 999 ;; } { .mfb nop.m 999 -(p9) fma f8=f9,f1,f0 -(p9) br.ret.spnt b0 ;; +(p9) fma.s0 f8=f9,f1,f0 +(p9) br.ret.spnt b0 ;; } -L(FMOD_Y_ZERO): +FMOD_Y_ZERO: // Y zero? Must be zero at this point // because it is the only choice left. // Return QNAN indefinite @@ -487,62 +493,59 @@ L(FMOD_Y_ZERO): {.mfi nop.m 0 // set Invalid - frcpa f12,p0=f0,f0 + frcpa.s0 f12,p0=f0,f0 nop.i 0 } // X NAN? { .mfi nop.m 999 -(p0) fclass.m.unc p9,p10 = f8, 0xc3 + fclass.m.unc p9,p10 = f8, 0xc3 nop.i 999 ;; } { .mfi nop.m 999 -(p10) fclass.nm p9,p10 = f8, 0xff +(p10) fclass.nm p9,p10 = f8, 0xff nop.i 999 ;; } {.mfi nop.m 999 - (p9) frcpa f11,p7=f8,f0 + (p9) frcpa.s0 f11,p7=f8,f0 nop.i 0;; } { .mfi nop.m 999 -(p10) frcpa f11,p7 = f9,f9 -(p0) mov GR_Parameter_TAG = 120 ;; +(p10) frcpa.s0 f11,p7 = f9,f9 + mov GR_Parameter_TAG = 120 ;; } { .mfi nop.m 999 -(p0) fmerge.s f10 = f8, f8 + fmerge.s f10 = f8, f8 nop.i 999 } { .mfb nop.m 999 -(p0) fma f8=f11,f1,f0 -(p0) br.sptk __libm_error_region;; + fma.s0 f8=f11,f1,f0 + br.sptk __libm_error_region;; } -.endp fmodl -ASM_SIZE_DIRECTIVE(fmodl) -ASM_SIZE_DIRECTIVE(__ieee754_fmodl) +GLOBAL_IEEE754_END(fmodl) -.proc __libm_error_region -__libm_error_region: +LOCAL_LIBM_ENTRY(__libm_error_region) .prologue { .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 + mov GR_SAVE_PFS=ar.pfs // Save ar.pfs } { .mfi -.fframe 64 +.fframe 64 add sp=-64,sp // Create new stack nop.f 0 mov GR_SAVE_GP=gp // Save gp @@ -550,18 +553,18 @@ __libm_error_region: { .mmi stfe [GR_Parameter_Y] = FR_Y,16 // Save 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 +.save b0, GR_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] = FR_X // Store Parameter 1 on stack + add GR_Parameter_RESULT = 0,GR_Parameter_Y + nop.b 0 // Parameter 3 address } { .mib stfe [GR_Parameter_Y] = FR_RESULT // Store Parameter 3 on stack - add GR_Parameter_Y = -16,GR_Parameter_Y + add GR_Parameter_Y = -16,GR_Parameter_Y br.call.sptk b0=__libm_error_support# // Call error handling function };; { .mmi @@ -576,15 +579,17 @@ __libm_error_region: mov b0 = GR_SAVE_B0 // Restore return address };; { .mib - mov gp = GR_SAVE_GP // Restore gp + mov gp = GR_SAVE_GP // Restore gp mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs br.ret.sptk b0 // Return -};; +};; + +LOCAL_LIBM_END(__libm_error_region) -.endp __libm_error_region -ASM_SIZE_DIRECTIVE(__libm_error_region) .type __libm_error_support#,@function .global __libm_error_support# + + |