• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* -*- c++ -*- */
2 /*
3  * Copyright © 2010-2016 Intel Corporation
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22  * IN THE SOFTWARE.
23  */
24 
25 #pragma once
26 
27 #include <assert.h>
28 #include "elk_reg.h"
29 #include "compiler/glsl/list.h"
30 
31 #define MAX_SAMPLER_MESSAGE_SIZE 11
32 
33 /* The sampler can return a vec5 when sampling with sparse residency. In
34  * SIMD32, each component takes up 4 GRFs, so we need to allow up to size-20
35  * VGRFs to hold the result.
36  */
37 #define MAX_VGRF_SIZE(devinfo) ((devinfo)->ver >= 20 ? 40 : 20)
38 
39 #ifdef __cplusplus
40 struct elk_backend_reg : private elk_reg
41 {
elk_backend_regelk_backend_reg42    elk_backend_reg() {}
elk_backend_regelk_backend_reg43    elk_backend_reg(const struct elk_reg &reg) : elk_reg(reg), offset(0) {}
44 
as_elk_regelk_backend_reg45    const elk_reg &as_elk_reg() const
46    {
47       assert(file == ARF || file == FIXED_GRF || file == MRF || file == IMM);
48       assert(offset == 0);
49       return static_cast<const elk_reg &>(*this);
50    }
51 
as_elk_regelk_backend_reg52    elk_reg &as_elk_reg()
53    {
54       assert(file == ARF || file == FIXED_GRF || file == MRF || file == IMM);
55       assert(offset == 0);
56       return static_cast<elk_reg &>(*this);
57    }
58 
59    bool equals(const elk_backend_reg &r) const;
60    bool negative_equals(const elk_backend_reg &r) const;
61 
62    bool is_zero() const;
63    bool is_one() const;
64    bool is_negative_one() const;
65    bool is_null() const;
66    bool is_accumulator() const;
67 
68    /** Offset from the start of the (virtual) register in bytes. */
69    uint16_t offset;
70 
71    using elk_reg::type;
72    using elk_reg::file;
73    using elk_reg::negate;
74    using elk_reg::abs;
75    using elk_reg::address_mode;
76    using elk_reg::subnr;
77    using elk_reg::nr;
78 
79    using elk_reg::swizzle;
80    using elk_reg::writemask;
81    using elk_reg::indirect_offset;
82    using elk_reg::vstride;
83    using elk_reg::width;
84    using elk_reg::hstride;
85 
86    using elk_reg::df;
87    using elk_reg::f;
88    using elk_reg::d;
89    using elk_reg::ud;
90    using elk_reg::d64;
91    using elk_reg::u64;
92 };
93 
94 struct elk_bblock_t;
95 
96 struct elk_backend_instruction : public exec_node {
97    bool elk_is_3src(const struct elk_compiler *compiler) const;
98    bool is_math() const;
99    bool is_control_flow_begin() const;
100    bool is_control_flow_end() const;
101    bool is_control_flow() const;
102    bool is_commutative() const;
103    bool can_do_source_mods() const;
104    bool can_do_saturate() const;
105    bool can_do_cmod() const;
106    bool reads_accumulator_implicitly() const;
107    bool writes_accumulator_implicitly(const struct intel_device_info *devinfo) const;
108 
109    /**
110     * Instructions that use indirect addressing have additional register
111     * regioning restrictions.
112     */
113    bool uses_indirect_addressing() const;
114 
115    void remove(elk_bblock_t *block, bool defer_later_block_ip_updates = false);
116    void insert_after(elk_bblock_t *block, elk_backend_instruction *inst);
117    void insert_before(elk_bblock_t *block, elk_backend_instruction *inst);
118 
119    /**
120     * True if the instruction has side effects other than writing to
121     * its destination registers.  You are expected not to reorder or
122     * optimize these out unless you know what you are doing.
123     */
124    bool has_side_effects() const;
125 
126    /**
127     * True if the instruction might be affected by side effects of other
128     * instructions.
129     */
130    bool is_volatile() const;
131 #else
132 struct elk_backend_instruction {
133    struct exec_node link;
134 #endif
135    /** @{
136     * Annotation for the generated IR.  One of the two can be set.
137     */
138    const void *ir;
139    const char *annotation;
140    /** @} */
141 
142    /**
143     * Execution size of the instruction.  This is used by the generator to
144     * generate the correct binary for the given instruction.  Current valid
145     * values are 1, 4, 8, 16, 32.
146     */
147    uint8_t exec_size;
148 
149    /**
150     * Channel group from the hardware execution and predication mask that
151     * should be applied to the instruction.  The subset of channel enable
152     * signals (calculated from the EU control flow and predication state)
153     * given by [group, group + exec_size) will be used to mask GRF writes and
154     * any other side effects of the instruction.
155     */
156    uint8_t group;
157 
158    uint32_t offset; /**< spill/unspill offset or texture offset bitfield */
159    uint8_t mlen; /**< SEND message length */
160    int8_t base_mrf; /**< First MRF in the SEND message, if mlen is nonzero. */
161    uint8_t target; /**< MRT target. */
162    uint8_t sfid; /**< SFID for SEND instructions */
163    uint32_t desc; /**< SEND[S] message descriptor immediate */
164    unsigned size_written; /**< Data written to the destination register in bytes. */
165 
166    enum elk_opcode opcode; /* ELK_OPCODE_* or ELK_FS_OPCODE_* */
167    enum elk_conditional_mod conditional_mod; /**< ELK_CONDITIONAL_* */
168    enum elk_predicate predicate;
169    bool predicate_inverse:1;
170    bool writes_accumulator:1; /**< instruction implicitly writes accumulator */
171    bool force_writemask_all:1;
172    bool no_dd_clear:1;
173    bool no_dd_check:1;
174    bool saturate:1;
175    bool shadow_compare:1;
176    bool check_tdr:1; /**< Only valid for SEND; turns it into a SENDC */
177    bool send_has_side_effects:1; /**< Only valid for ELK_SHADER_OPCODE_SEND */
178    bool send_is_volatile:1; /**< Only valid for ELK_SHADER_OPCODE_SEND */
179    bool predicate_trivial:1; /**< The predication mask applied to this
180                               *   instruction is guaranteed to be uniform and
181                               *   a superset of the execution mask of the
182                               *   present block, no currently enabled channels
183                               *   will be disabled by the predicate.
184                               */
185    bool eot:1;
186 
187    /* Chooses which flag subregister (f0.0 to f3.1) is used for conditional
188     * mod and predication.
189     */
190    unsigned flag_subreg:3;
191 
192    /** The number of hardware registers used for a message header. */
193    uint8_t header_size;
194 };
195