111
|
1 /* Machine description for AArch64 architecture.
|
145
|
2 Copyright (C) 2009-2020 Free Software Foundation, Inc.
|
111
|
3 Contributed by ARM Ltd.
|
|
4
|
|
5 This file is part of GCC.
|
|
6
|
|
7 GCC is free software; you can redistribute it and/or modify it
|
|
8 under the terms of the GNU General Public License as published by
|
|
9 the Free Software Foundation; either version 3, or (at your option)
|
|
10 any later version.
|
|
11
|
|
12 GCC is distributed in the hope that it will be useful, but
|
|
13 WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
15 General Public License for more details.
|
|
16
|
|
17 You should have received a copy of the GNU General Public License
|
|
18 along with GCC; see the file COPYING3. If not see
|
|
19 <http://www.gnu.org/licenses/>. */
|
|
20
|
145
|
21 /* Important note about Carry generation in AArch64.
|
|
22
|
|
23 Unlike some architectures, the C flag generated by a subtract
|
|
24 operation, or a simple compare operation is set to 1 if the result
|
|
25 does not overflow in an unsigned sense. That is, if there is no
|
|
26 borrow needed from a higher word. That means that overflow from
|
|
27 addition will set C, but overflow from a subtraction will clear C.
|
|
28 We use CC_Cmode to represent detection of overflow from addition as
|
|
29 CCmode is used for 'normal' compare (subtraction) operations. For
|
|
30 ADC, the representation becomes more complex still, since we cannot
|
|
31 use the normal idiom of comparing the result to one of the input
|
|
32 operands; instead we use CC_ADCmode to represent this case. */
|
111
|
33 CC_MODE (CCFP);
|
|
34 CC_MODE (CCFPE);
|
|
35 CC_MODE (CC_SWP);
|
145
|
36 CC_MODE (CC_NZC); /* Only N, Z and C bits of condition flags are valid.
|
|
37 (Used with SVE predicate tests.) */
|
111
|
38 CC_MODE (CC_NZ); /* Only N and Z bits of condition flags are valid. */
|
|
39 CC_MODE (CC_Z); /* Only Z bit of condition flags is valid. */
|
145
|
40 CC_MODE (CC_C); /* C represents unsigned overflow of a simple addition. */
|
|
41 CC_MODE (CC_ADC); /* Unsigned overflow from an ADC (add with carry). */
|
131
|
42 CC_MODE (CC_V); /* Only V bit of condition flags is valid. */
|
111
|
43
|
|
44 /* Half-precision floating point for __fp16. */
|
|
45 FLOAT_MODE (HF, 2, 0);
|
|
46 ADJUST_FLOAT_FORMAT (HF, &ieee_half_format);
|
|
47
|
|
48 /* Vector modes. */
|
131
|
49
|
|
50 VECTOR_BOOL_MODE (VNx16BI, 16, 2);
|
|
51 VECTOR_BOOL_MODE (VNx8BI, 8, 2);
|
|
52 VECTOR_BOOL_MODE (VNx4BI, 4, 2);
|
|
53 VECTOR_BOOL_MODE (VNx2BI, 2, 2);
|
|
54
|
|
55 ADJUST_NUNITS (VNx16BI, aarch64_sve_vg * 8);
|
|
56 ADJUST_NUNITS (VNx8BI, aarch64_sve_vg * 4);
|
|
57 ADJUST_NUNITS (VNx4BI, aarch64_sve_vg * 2);
|
|
58 ADJUST_NUNITS (VNx2BI, aarch64_sve_vg);
|
|
59
|
|
60 ADJUST_ALIGNMENT (VNx16BI, 2);
|
|
61 ADJUST_ALIGNMENT (VNx8BI, 2);
|
|
62 ADJUST_ALIGNMENT (VNx4BI, 2);
|
|
63 ADJUST_ALIGNMENT (VNx2BI, 2);
|
|
64
|
145
|
65 /* Bfloat16 modes. */
|
|
66 FLOAT_MODE (BF, 2, 0);
|
|
67 ADJUST_FLOAT_FORMAT (BF, &arm_bfloat_half_format);
|
|
68
|
111
|
69 VECTOR_MODES (INT, 8); /* V8QI V4HI V2SI. */
|
|
70 VECTOR_MODES (INT, 16); /* V16QI V8HI V4SI V2DI. */
|
|
71 VECTOR_MODES (FLOAT, 8); /* V2SF. */
|
|
72 VECTOR_MODES (FLOAT, 16); /* V4SF V2DF. */
|
|
73 VECTOR_MODE (FLOAT, DF, 1); /* V1DF. */
|
131
|
74 VECTOR_MODE (FLOAT, HF, 2); /* V2HF. */
|
111
|
75
|
|
76 /* Oct Int: 256-bit integer mode needed for 32-byte vector arguments. */
|
|
77 INT_MODE (OI, 32);
|
|
78
|
|
79 /* Opaque integer modes for 3 or 4 Neon q-registers / 6 or 8 Neon d-registers
|
|
80 (2 d-regs = 1 q-reg = TImode). */
|
|
81 INT_MODE (CI, 48);
|
|
82 INT_MODE (XI, 64);
|
|
83
|
131
|
84 /* Define SVE modes for NVECS vectors. VB, VH, VS and VD are the prefixes
|
|
85 for 8-bit, 16-bit, 32-bit and 64-bit elements respectively. It isn't
|
|
86 strictly necessary to set the alignment here, since the default would
|
|
87 be clamped to BIGGEST_ALIGNMENT anyhow, but it seems clearer. */
|
|
88 #define SVE_MODES(NVECS, VB, VH, VS, VD) \
|
145
|
89 VECTOR_MODES_WITH_PREFIX (VNx, INT, 16 * NVECS, 0); \
|
|
90 VECTOR_MODES_WITH_PREFIX (VNx, FLOAT, 16 * NVECS, 0); \
|
131
|
91 \
|
|
92 ADJUST_NUNITS (VB##QI, aarch64_sve_vg * NVECS * 8); \
|
|
93 ADJUST_NUNITS (VH##HI, aarch64_sve_vg * NVECS * 4); \
|
|
94 ADJUST_NUNITS (VS##SI, aarch64_sve_vg * NVECS * 2); \
|
|
95 ADJUST_NUNITS (VD##DI, aarch64_sve_vg * NVECS); \
|
145
|
96 ADJUST_NUNITS (VH##BF, aarch64_sve_vg * NVECS * 4); \
|
131
|
97 ADJUST_NUNITS (VH##HF, aarch64_sve_vg * NVECS * 4); \
|
|
98 ADJUST_NUNITS (VS##SF, aarch64_sve_vg * NVECS * 2); \
|
|
99 ADJUST_NUNITS (VD##DF, aarch64_sve_vg * NVECS); \
|
|
100 \
|
|
101 ADJUST_ALIGNMENT (VB##QI, 16); \
|
|
102 ADJUST_ALIGNMENT (VH##HI, 16); \
|
|
103 ADJUST_ALIGNMENT (VS##SI, 16); \
|
|
104 ADJUST_ALIGNMENT (VD##DI, 16); \
|
145
|
105 ADJUST_ALIGNMENT (VH##BF, 16); \
|
131
|
106 ADJUST_ALIGNMENT (VH##HF, 16); \
|
|
107 ADJUST_ALIGNMENT (VS##SF, 16); \
|
|
108 ADJUST_ALIGNMENT (VD##DF, 16);
|
|
109
|
|
110 /* Give SVE vectors the names normally used for 256-bit vectors.
|
|
111 The actual number depends on command-line flags. */
|
|
112 SVE_MODES (1, VNx16, VNx8, VNx4, VNx2)
|
|
113 SVE_MODES (2, VNx32, VNx16, VNx8, VNx4)
|
|
114 SVE_MODES (3, VNx48, VNx24, VNx12, VNx6)
|
|
115 SVE_MODES (4, VNx64, VNx32, VNx16, VNx8)
|
|
116
|
145
|
117 /* Partial SVE vectors:
|
|
118
|
|
119 VNx2QI VNx4QI VNx8QI
|
|
120 VNx2HI VNx4HI
|
|
121 VNx2SI
|
|
122
|
|
123 In memory they occupy contiguous locations, in the same way as fixed-length
|
|
124 vectors. E.g. VNx8QImode is half the size of VNx16QImode.
|
|
125
|
|
126 Passing 1 as the final argument ensures that the modes come after all
|
|
127 other modes in the GET_MODE_WIDER chain, so that we never pick them
|
|
128 in preference to a full vector mode. */
|
|
129 VECTOR_MODES_WITH_PREFIX (VNx, INT, 2, 1);
|
|
130 VECTOR_MODES_WITH_PREFIX (VNx, INT, 4, 1);
|
|
131 VECTOR_MODES_WITH_PREFIX (VNx, INT, 8, 1);
|
|
132 VECTOR_MODES_WITH_PREFIX (VNx, FLOAT, 4, 1);
|
|
133 VECTOR_MODES_WITH_PREFIX (VNx, FLOAT, 8, 1);
|
|
134
|
|
135 ADJUST_NUNITS (VNx2QI, aarch64_sve_vg);
|
|
136 ADJUST_NUNITS (VNx2HI, aarch64_sve_vg);
|
|
137 ADJUST_NUNITS (VNx2SI, aarch64_sve_vg);
|
|
138 ADJUST_NUNITS (VNx2HF, aarch64_sve_vg);
|
|
139 ADJUST_NUNITS (VNx2SF, aarch64_sve_vg);
|
|
140
|
|
141 ADJUST_NUNITS (VNx4QI, aarch64_sve_vg * 2);
|
|
142 ADJUST_NUNITS (VNx4HI, aarch64_sve_vg * 2);
|
|
143 ADJUST_NUNITS (VNx4HF, aarch64_sve_vg * 2);
|
|
144
|
|
145 ADJUST_NUNITS (VNx8QI, aarch64_sve_vg * 4);
|
|
146
|
|
147 ADJUST_ALIGNMENT (VNx2QI, 1);
|
|
148 ADJUST_ALIGNMENT (VNx4QI, 1);
|
|
149 ADJUST_ALIGNMENT (VNx8QI, 1);
|
|
150
|
|
151 ADJUST_ALIGNMENT (VNx2HI, 2);
|
|
152 ADJUST_ALIGNMENT (VNx4HI, 2);
|
|
153 ADJUST_ALIGNMENT (VNx2HF, 2);
|
|
154 ADJUST_ALIGNMENT (VNx4HF, 2);
|
|
155
|
|
156 ADJUST_ALIGNMENT (VNx2SI, 4);
|
|
157 ADJUST_ALIGNMENT (VNx2SF, 4);
|
|
158
|
111
|
159 /* Quad float: 128-bit floating mode for long doubles. */
|
|
160 FLOAT_MODE (TF, 16, ieee_quad_format);
|
131
|
161
|
|
162 /* A 4-tuple of SVE vectors with the maximum -msve-vector-bits= setting.
|
|
163 Note that this is a limit only on the compile-time sizes of modes;
|
|
164 it is not a limit on the runtime sizes, since VL-agnostic code
|
|
165 must work with arbitary vector lengths. */
|
|
166 #define MAX_BITSIZE_MODE_ANY_MODE (2048 * 4)
|
|
167
|
|
168 /* Coefficient 1 is multiplied by the number of 128-bit chunks in an
|
|
169 SVE vector (referred to as "VQ") minus one. */
|
|
170 #define NUM_POLY_INT_COEFFS 2
|