3 // Copyright (c) 2000, 2001, Intel Corporation
4 // All rights reserved.
6 // Contributed 2/2/2000 by John Harrison, Ted Kubaska, Bob Norin, Shane Story,
7 // and Ping Tak Peter Tang of the Computational Software Lab, Intel Corporation.
11 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
12 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
13 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
14 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
15 // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
16 // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
17 // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
18 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
19 // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
20 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
21 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
23 // Intel Corporation is the author of this code, and requests that all
24 // problem reports or change requests be submitted to it directly at
25 // http://developer.intel.com/opensource.
28 //==============================================================
29 // 2/02/00 Initial version
30 // 2/16/00 Modified to conform to C9X
31 // 3/16/00 Improved speed
32 // 4/04/00 Unwind support added
33 // 5/30/00 Fixed bug when x double-extended denormal
34 // 8/15/00 Bundle added after call to __libm_error_support to properly
35 // set [the previously overwritten] GR_Parameter_RESULT.
38 //==============================================================
39 // double logb( double x);
41 // Overview of operation
42 //==============================================================
43 // The logb function extracts the exponent of x as an integer in
44 // floating-point format.
45 // logb computes log2 of x as a double
47 // logb is similar to ilogb but differs in the following ways:
49 // ilogb: returns INT_MAX
51 // Nan returns FP_LOGBNAN (which is either INT_MAX or INT_MIN)
52 // ilogb: returns INT_MAX (7fffffff)
53 // logb: returns QNAN (quietized SNAN)
54 // 0 returns FP_ILOGB0 (which is either INT_MIN or -INT_MAX)
55 // ilogb: returns -INT_MAX (80000001)
56 // logb: returns -inf, raises the divide-by-zero exception,
57 // and calls libm_error_support to set domain error
60 //==============================================================
61 // general registers used:
64 // r38 -> r41 used as parameters to error path
66 // predicate registers used:
68 // floating-point registers used:
72 #include "libm_support.h"
80 GR_Parameter_RESULT = r40
92 // qnan snan inf norm unorm 0 -+
96 alloc r32=ar.pfs,1,5,4,0
97 (p0) fclass.m.unc p8,p0 = f8, 0x0b
101 // r37 = exp(f8) - - 0xffff
103 // f8 = convert_to_fp (sig))
105 (p0) getf.exp r35 = f8
110 // qnan snan inf norm unorm 0 -+
114 (p0) mov r33 = 0xffff
115 (p0) mov r34 = 0x1ffff
116 (p0) fclass.m.unc p6,p0 = f8, 0xe3 ;;
120 (p0) and r36 = r35, r34
121 (p0) fclass.m.unc p7,p0 = f8, 0x07
122 (p8) br.cond.spnt L(LOGB_DENORM) ;;
126 (p0) sub r37 = r36, r33
128 (p6) br.cond.spnt L(LOGB_NAN_INF) ;;
132 (p0) setf.sig f9 = r37
134 (p7) br.cond.spnt L(LOGB_ZERO) ;;
139 (p0) fcvt.xf f10 = f9
145 (p0) fnorm.d f8 = f10
146 (p0) br.ret.sptk b0 ;;
150 // Form signexp of 2^64 in case need to scale denormal
151 // Check to see if double-extended denormal
153 (p0) mov r38 = 0x1003f
154 (p0) fclass.m.unc p8,p0 = f10, 0x0b
158 // Form 2^64 in case need to scale denormal
160 (p0) setf.exp f11 = r38
165 // If double-extended denormal add 64 to exponent bias for scaling
166 // If double-extended denormal form x * 2^64 which is normal
168 (p8) add r33 = 64, r33
169 (p8) fmpy f10 = f10, f11
173 // Logic is the same as normal path but use normalized input
175 (p0) getf.exp r35 = f10 ;;
181 (p0) and r36 = r35, r34 ;;
182 (p0) sub r37 = r36, r33
187 (p0) setf.sig f9 = r37
194 (p0) fcvt.xf f10 = f9
200 (p0) fnorm.d f8 = f10
201 (p0) br.ret.sptk b0 ;;
206 // X NAN or INFINITY, return f8 * f8
209 (p0) fma.d f8= f8,f8,f0
210 (p0) br.ret.sptk b0 ;;
214 ASM_SIZE_DIRECTIVE(logb)
216 // Stack operations when calling error support.
217 // (1) (2) (3) (call) (4)
218 // sp -> + psp -> + psp -> + sp -> +
220 // | | <- GR_Y R3 ->| <- GR_RESULT | -> f8
222 // | <-GR_Y Y2->| Y2 ->| <- GR_Y |
224 // | | <- GR_X X1 ->| |
226 // sp-64 -> + sp -> + sp -> + +
227 // save ar.pfs save b0 restore gp
228 // save gp restore ar.pfs
232 .proc __libm_error_region
239 // f9 = 1.0/f10 = -1.0/-|f8|
242 mov r41 = 151 // Error code
243 (p0) fmerge.s f9 = f0,f8
251 fmerge.ns f10 = f0,f9
258 add GR_Parameter_Y=-32,sp // Parameter 2 value
260 .save ar.pfs,GR_SAVE_PFS
261 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
265 add sp=-64,sp // Create new stack
267 mov GR_SAVE_GP=gp // Save gp
273 stfd [GR_Parameter_Y] = f0,16 // STORE Parameter 2 on stack
274 add GR_Parameter_X = 16,sp // Parameter 1 address
276 mov GR_SAVE_B0=b0 // Save b0
282 stfd [GR_Parameter_X] = f8 // STORE Parameter 1 on stack
283 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
287 stfd [GR_Parameter_Y] = f9 // Store Parameter 3 on stack
288 add GR_Parameter_Y = -16,GR_Parameter_Y
289 br.call.sptk b0=__libm_error_support# // Call error handling function
294 add GR_Parameter_RESULT = 48,sp
299 ldfd f8 = [GR_Parameter_RESULT] // Get return result off stack
301 add sp = 64,sp // Restore stack pointer
302 mov b0 = GR_SAVE_B0 // Restore return address
305 mov gp = GR_SAVE_GP // Restore gp
306 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
310 .endp __libm_error_region
311 ASM_SIZE_DIRECTIVE(__libm_error_region)
313 .type __libm_error_support#,@function
314 .global __libm_error_support#