1 /* Copyright (C) 2004-2014 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
4 The GNU C Library is free software; you can redistribute it and/or
5 modify it under the terms of the GNU Lesser General Public
6 License as published by the Free Software Foundation; either
7 version 2.1 of the License, or (at your option) any later version.
9 The GNU C Library is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 Lesser General Public License for more details.
14 You should have received a copy of the GNU Lesser General Public
15 License along with the GNU C Library. If not, see
16 <http://www.gnu.org/licenses/>. */
21 /* 64-bit unsigned long divide. These are not normal C functions. Argument
22 registers are t10 and t11, the result goes in t12. Only t12 and AT may be
25 Theory of operation here is that we can use the FPU divider for virtually
26 all operands that we see: all dividend values between -2**53 and 2**53-1
27 can be computed directly. Note that divisor values need not be checked
28 against that range because the rounded fp value will be close enough such
29 that the quotient is < 1, which will properly be truncated to zero when we
30 convert back to integer.
32 When the dividend is outside the range for which we can compute exact
33 results, we use the fp quotent as an estimate from which we begin refining
34 an exact integral value. This reduces the number of iterations in the
35 shift-and-subtract loop significantly.
37 The FPCR save/restore is due to the fact that the EV6 _will_ set FPCR_INE
38 for cvttq/c even without /sui being set. It will not, however, properly
39 raise the exception, so we don't have to worry about FPCR_INED being clear
40 and so dying by SIGFPE. */
45 .type __divqu, @funcnoplt
49 cfi_return_column (RA)
52 cfi_def_cfa_offset (FRAME)
55 /* Get the fp divide insn issued as quickly as possible. After
56 that's done, we have at least 22 cycles until its results are
57 ready -- all the time in the world to figure out how we're
58 going to use the results. */
65 cfi_rel_offset ($f0, 0)
66 cfi_rel_offset ($f1, 8)
67 cfi_rel_offset ($f3, 48)
70 _ITOFT2 X, $f0, 16, Y, $f1, 24
76 /* Check to see if Y was mis-converted as signed value. */
80 /* Check to see if X fit in the double as an exact value. */
84 /* If we get here, we're expecting exact results from the division.
85 Do nothing else besides convert and clean up. */
97 cfi_def_cfa_offset (0)
104 /* If we get here, X is so big that bit 63 is set, which made the
105 conversion come out negative. Fix it up lest we not even get
107 ldah AT, 0x5f80 /* 2**64 as float. */
109 cfi_rel_offset ($f2, 24)
118 /* Ok, we've now the divide issued. Continue with other checks. */
125 cfi_remember_state /* for y_is_neg */
129 /* If we get here, X is large enough that we don't expect exact
130 results, and neither X nor Y got mis-translated for the fp
131 division. Our task is to take the fp result, figure out how
132 far it's off from the correct result and compute a fixup. */
137 cfi_rel_offset (t0, 16)
138 cfi_rel_offset (t1, 24)
139 cfi_rel_offset (t2, 32)
140 cfi_rel_offset (t3, 40)
142 #define Q RV /* quotient */
143 #define R t0 /* remainder */
144 #define SY t1 /* scaled Y */
145 #define S t2 /* scalar */
146 #define QY t3 /* Q*Y */
157 cfi_rel_offset (t4, 8)
187 cfi_def_cfa_offset (0)
192 /* The quotient that we computed was too large. We need to reduce
193 it by S such that Y*S >= R. Obviously the closer we get to the
194 correct value the better, but overshooting high is ok, as we'll
195 fix that up later. */
209 /* The quotient that we computed was too small. Divide Y by the
210 current remainder (R) and add that to the existing quotient (Q).
211 The expectation, of course, is that R is much smaller than X. */
212 /* Begin with a shift-up loop. Compute S such that Y*S >= R. We
213 already have a copy of Y in SY and the value 1 in S. */
221 /* Shift-down and subtract loop. Each iteration compares our scaled
222 Y (SY) with the remainder (R); if SY <= R then X is divisible by
223 Y's scalar (S) so add it to the quotient (Q). */
239 /* If we get here, Y is so big that bit 63 is set. The results
240 from the divide will be completely wrong. Fortunately, the
241 quotient must be either 0 or 1, so just compute it directly. */
250 cfi_def_cfa_offset (0)
254 .size __divqu, .-__divqu