4 // Copyright (c) 2000 - 2003, Intel Corporation
5 // All rights reserved.
8 // Redistribution and use in source and binary forms, with or without
9 // modification, are permitted provided that the following conditions are
12 // * Redistributions of source code must retain the above copyright
13 // notice, this list of conditions and the following disclaimer.
15 // * Redistributions in binary form must reproduce the above copyright
16 // notice, this list of conditions and the following disclaimer in the
17 // documentation and/or other materials provided with the distribution.
19 // * The name of Intel Corporation may not be used to endorse or promote
20 // products derived from this software without specific prior written
23 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
26 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
27 // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
28 // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
29 // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
30 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
31 // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
32 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
33 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35 // Intel Corporation is the author of this code, and requests that all
36 // problem reports or change requests be submitted to it directly at
37 // http://www.intel.com/software/products/opensource/libraries/num.htm.
40 //==============================================================
41 // 02/02/00 Initial version
42 // 01/26/01 ldexpf completely reworked and now standalone version
43 // 01/04/02 Added handling for int 32 or 64 bits
44 // 05/20/02 Cleaned up namespace and sf0 syntax
45 // 02/10/03 Reordered header: .section, .global, .proc, .align
46 // 08/04/03 Improved performance
49 //==============================================================
50 // float __libm_ldexpf (float x, int n, int int_type)
51 // input floating point f8 and int n (r33), int int_type (r34)
52 // output floating point f8
54 // int_type = 0 if int is 32 bits
55 // int_type = 1 if int is 64 bits
57 // Returns x* 2**n using an fma and detects overflow
62 // Compute biased exponent of result exp_Result = N + exp_X
64 // exp_Result > 0x1007e -> Certain overflow
65 // exp_Result = 0x1007e -> Possible overflow
66 // 0x0ff81 <= exp_Result < 0x1007e -> No over/underflow (main path)
67 // 0x0ff81 - 23 <= exp_Result < 0x0ff81 -> Possible underflow
68 // exp_Result < 0x0ff81 - 23 -> Certain underflow
99 GR_Parameter_RESULT = r37
103 GLOBAL_LIBM_ENTRY(__libm_ldexpf)
106 // Is x NAN, INF, ZERO, +-?
107 // Build the exponent Bias
110 getf.exp GR_signexp_X = FR_Floating_X // Get signexp of x
111 fclass.m p6,p0 = FR_Floating_X, 0xe7 // @snan | @qnan | @inf | @zero
112 mov GR_Bias = 0x0ffff
116 // Is integer type 32 bits?
119 mov GR_Big = 35000 // If N this big then certain overflow
120 fnorm.s1 FR_Norm_X = FR_Floating_X
121 cmp.eq p8,p9 = r34,r0
125 // Sign extend N if int is 32 bits
127 (p9) mov GR_N_as_int = r33 // Copy N if int is 64 bits
128 fclass.m p9,p0 = FR_Floating_X, 0x0b // Test for x=unorm
129 (p8) sxt4 GR_N_as_int = r33 // Sign extend N if int is 32 bits
132 mov GR_NBig = -35000 // If N this small then certain underflow
134 mov GR_max_exp = 0x1007e // Exponent of maximum float
138 // Create biased exponent for 2**N
140 add GR_N_Biased = GR_Bias,GR_N_as_int
142 cmp.ge p7, p0 = GR_N_as_int, GR_Big // Certain overflow?
145 cmp.le p8, p0 = GR_N_as_int, GR_NBig // Certain underflow?
146 mov GR_min_exp = 0x0ff81 // Exponent of minimum float
147 (p9) br.cond.spnt LDEXPF_UNORM // Branch if x=unorm
152 // Main path continues. Also return here from x=unorm path.
154 .pred.rel "mutex",p7,p8
156 setf.exp FR_Two_N = GR_N_Biased
158 (p7) mov GR_N_as_int = GR_Big // Limit max N
161 (p8) mov GR_N_as_int = GR_NBig // Limit min N
163 (p8) cmp.eq p7,p0 = r0,r0 // Set p7 if |N| big
168 // Create biased exponent for 2**N for N big
172 (p7) add GR_N_Biased = GR_Bias,GR_N_as_int
174 cmp.eq.or p6,p0 = r33,r0
177 mov GR_pos_ov_limit = 0x1007f // Exponent for positive overflow
179 mov GR_exp_mask = 0x1ffff // Exponent mask
184 // Create 2**N for N big
185 // Return x when N = 0 or X = Nan, Inf, Zero
188 (p7) setf.exp FR_Two_N = GR_N_Biased
190 mov GR_min_den_exp = 0x0ff81 - 23 // Exponent of min denorm float
193 and GR_exp_X = GR_exp_mask, GR_signexp_X
194 (p6) fma.s.s0 FR_Result = FR_Floating_X, f1, f0
200 // Raise Denormal operand flag with compare
201 // Compute biased result exponent
204 add GR_exp_Result = GR_exp_X, GR_N_as_int
205 fcmp.ge.s0 p0,p11 = FR_Floating_X,f0
206 mov GR_neg_ov_limit = 0x3007f // Exponent for negative overflow
211 // Do final operation
214 cmp.lt p7,p6 = GR_exp_Result, GR_max_exp // Test no overflow
215 fma.s.s0 FR_Result = FR_Two_N,FR_Norm_X,f0
216 cmp.lt p9,p0 = GR_exp_Result, GR_min_den_exp // Test sure underflow
221 (p9) br.cond.spnt LDEXPF_UNDERFLOW // Branch if certain underflow
226 (p6) cmp.gt.unc p6,p8 = GR_exp_Result, GR_max_exp // Test sure overflow
227 (p7) cmp.ge.unc p7,p9 = GR_exp_Result, GR_min_exp // Test no over/underflow
228 (p7) br.ret.sptk b0 // Return from main path
233 (p6) br.cond.spnt LDEXPF_OVERFLOW // Branch if certain overflow
234 (p8) br.cond.spnt LDEXPF_POSSIBLE_OVERFLOW // Branch if possible overflow
235 (p9) br.cond.spnt LDEXPF_POSSIBLE_UNDERFLOW // Branch if possible underflow
239 // Here if possible underflow.
240 // Resulting exponent: 0x0ff81-23 <= exp_Result < 0x0ff81
241 LDEXPF_POSSIBLE_UNDERFLOW:
243 // Here if possible overflow.
244 // Resulting exponent: 0x1007e = exp_Result
245 LDEXPF_POSSIBLE_OVERFLOW:
247 // Set up necessary status fields
249 // S0 user supplied status
250 // S2 user supplied status + WRE + TD (Overflows)
251 // S3 user supplied status + FZ + TD (Underflows)
266 // Do final operation with s2 and s3
269 setf.exp FR_NBig = GR_neg_ov_limit
270 fma.s.s3 FR_Result3 = FR_Two_N,FR_Norm_X,f0
274 setf.exp FR_Big = GR_pos_ov_limit
275 fma.s.s2 FR_Result2 = FR_Two_N,FR_Norm_X,f0
280 // Check for overflow or underflow.
297 // Is the result zero?
301 fclass.m p6, p0 = FR_Result3, 0x007
306 fcmp.ge.s1 p7, p8 = FR_Result2 , FR_Big
312 // Detect masked underflow - Tiny + Inexact Only
316 (p6) fcmp.neq.unc.s1 p6, p0 = FR_Result , FR_Result2
322 // Is result bigger the allowed range?
323 // Branch out for underflow
327 (p8) fcmp.le.unc.s1 p9, p10 = FR_Result2 , FR_NBig
328 (p6) br.cond.spnt LDEXPF_UNDERFLOW
333 // Branch out for overflow
336 (p7) br.cond.spnt LDEXPF_OVERFLOW
337 (p9) br.cond.spnt LDEXPF_OVERFLOW
338 br.ret.sptk b0 // Return from main path.
342 // Here if result overflows
345 alloc r32=ar.pfs,3,0,4,0
346 addl GR_Tag = 148, r0 // Set error tag for overflow
347 br.cond.sptk __libm_error_region // Call error support for overflow
351 // Here if result underflows
354 alloc r32=ar.pfs,3,0,4,0
355 addl GR_Tag = 149, r0 // Set error tag for underflow
356 br.cond.sptk __libm_error_region // Call error support for underflow
363 getf.exp GR_signexp_X = FR_Norm_X // Get signexp of normalized x
365 br.cond.sptk LDEXPF_COMMON // Return to main path
370 GLOBAL_LIBM_END(__libm_ldexpf)
371 LOCAL_LIBM_ENTRY(__libm_error_region)
374 // Get stack address of N
378 add GR_Parameter_Y=-32,sp
380 .save ar.pfs,GR_SAVE_PFS
381 mov GR_SAVE_PFS=ar.pfs
394 // Store N on stack in correct position
395 // Locate the address of x on stack
398 st8 [GR_Parameter_Y] = GR_N_as_int,16
399 add GR_Parameter_X = 16,sp
405 // Store x on the stack.
406 // Get address for result on stack.
410 stfs [GR_Parameter_X] = FR_Norm_X
411 add GR_Parameter_RESULT = 0,GR_Parameter_Y
415 stfs [GR_Parameter_Y] = FR_Result
416 add GR_Parameter_Y = -16,GR_Parameter_Y
417 br.call.sptk b0=__libm_error_support#
421 // Get location of result on stack
424 add GR_Parameter_RESULT = 48,sp
430 // Get the new result
433 ldfs FR_Result = [GR_Parameter_RESULT]
440 // Restore gp, ar.pfs and return
444 mov ar.pfs = GR_SAVE_PFS
448 LOCAL_LIBM_END(__libm_error_region)
450 .type __libm_error_support#,@function
451 .global __libm_error_support#