Files
third_party_mesa3d/src/compiler/nir/nir_validate.c

1119 lines
33 KiB
C
Raw Normal View History

/*
* Copyright © 2014 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*
* Authors:
* Connor Abbott (cwabbott0@gmail.com)
*
*/
#include "nir.h"
#include <assert.h>
/*
* This file checks for invalid IR indicating a bug somewhere in the compiler.
*/
/* Since this file is just a pile of asserts, don't bother compiling it if
* we're not building a debug build.
*/
#ifdef DEBUG
/*
* Per-register validation state.
*/
typedef struct {
/*
* equivalent to the uses and defs in nir_register, but built up by the
* validator. At the end, we verify that the sets have the same entries.
*/
struct set *uses, *if_uses, *defs;
nir_function_impl *where_defined; /* NULL for global registers */
} reg_validate_state;
typedef struct {
/*
* equivalent to the uses in nir_ssa_def, but built up by the validator.
* At the end, we verify that the sets have the same entries.
*/
struct set *uses, *if_uses;
nir_function_impl *where_defined;
} ssa_def_validate_state;
typedef struct {
/* map of register -> validation state (struct above) */
struct hash_table *regs;
/* the current shader being validated */
nir_shader *shader;
/* the current instruction being validated */
nir_instr *instr;
/* the current basic block being validated */
nir_block *block;
/* the current if statement being validated */
nir_if *if_stmt;
/* the current loop being visited */
nir_loop *loop;
/* the parent of the current cf node being visited */
nir_cf_node *parent_node;
/* the current function implementation being validated */
nir_function_impl *impl;
/* map of SSA value -> function implementation where it is defined */
struct hash_table *ssa_defs;
/* bitset of ssa definitions we have found; used to check uniqueness */
BITSET_WORD *ssa_defs_found;
/* bitset of registers we have currently found; used to check uniqueness */
BITSET_WORD *regs_found;
/* map of local variable -> function implementation where it is defined */
struct hash_table *var_defs;
} validate_state;
static void validate_src(nir_src *src, validate_state *state);
static void
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
validate_reg_src(nir_src *src, validate_state *state)
{
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
assert(src->reg.reg != NULL);
struct hash_entry *entry;
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
entry = _mesa_hash_table_search(state->regs, src->reg.reg);
assert(entry);
reg_validate_state *reg_state = (reg_validate_state *) entry->data;
if (state->instr) {
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
_mesa_set_add(reg_state->uses, src);
} else {
assert(state->if_stmt);
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
_mesa_set_add(reg_state->if_uses, src);
}
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
if (!src->reg.reg->is_global) {
assert(reg_state->where_defined == state->impl &&
"using a register declared in a different function");
}
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
assert((src->reg.reg->num_array_elems == 0 ||
src->reg.base_offset < src->reg.reg->num_array_elems) &&
"definitely out-of-bounds array access");
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
if (src->reg.indirect) {
assert(src->reg.reg->num_array_elems != 0);
assert((src->reg.indirect->is_ssa ||
src->reg.indirect->reg.indirect == NULL) &&
"only one level of indirection allowed");
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
validate_src(src->reg.indirect, state);
}
}
static void
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
validate_ssa_src(nir_src *src, validate_state *state)
{
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
assert(src->ssa != NULL);
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
struct hash_entry *entry = _mesa_hash_table_search(state->ssa_defs, src->ssa);
assert(entry);
ssa_def_validate_state *def_state = (ssa_def_validate_state *)entry->data;
assert(def_state->where_defined == state->impl &&
"using an SSA value defined in a different function");
if (state->instr) {
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
_mesa_set_add(def_state->uses, src);
} else {
assert(state->if_stmt);
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
_mesa_set_add(def_state->if_uses, src);
}
/* TODO validate that the use is dominated by the definition */
}
static void
validate_src(nir_src *src, validate_state *state)
{
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
if (state->instr)
assert(src->parent_instr == state->instr);
else
assert(src->parent_if == state->if_stmt);
if (src->is_ssa)
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
validate_ssa_src(src, state);
else
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
validate_reg_src(src, state);
}
static void
validate_alu_src(nir_alu_instr *instr, unsigned index, validate_state *state)
{
nir_alu_src *src = &instr->src[index];
unsigned num_components;
unsigned src_bit_size;
if (src->src.is_ssa) {
src_bit_size = src->src.ssa->bit_size;
num_components = src->src.ssa->num_components;
} else {
src_bit_size = src->src.reg.reg->bit_size;
if (src->src.reg.reg->is_packed)
num_components = 4; /* can't check anything */
else
num_components = src->src.reg.reg->num_components;
}
for (unsigned i = 0; i < 4; i++) {
assert(src->swizzle[i] < 4);
if (nir_alu_instr_channel_used(instr, index, i))
assert(src->swizzle[i] < num_components);
}
nir_alu_type src_type = nir_op_infos[instr->op].input_types[index];
/* 8-bit float isn't a thing */
if (nir_alu_type_get_base_type(src_type) == nir_type_float)
assert(src_bit_size == 16 || src_bit_size == 32 || src_bit_size == 64);
if (nir_alu_type_get_type_size(src_type)) {
/* This source has an explicit bit size */
assert(nir_alu_type_get_type_size(src_type) == src_bit_size);
} else {
if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type)) {
unsigned dest_bit_size =
instr->dest.dest.is_ssa ? instr->dest.dest.ssa.bit_size
: instr->dest.dest.reg.reg->bit_size;
assert(dest_bit_size == src_bit_size);
}
}
validate_src(&src->src, state);
}
static void
validate_reg_dest(nir_reg_dest *dest, validate_state *state)
{
assert(dest->reg != NULL);
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
assert(dest->parent_instr == state->instr);
struct hash_entry *entry2;
entry2 = _mesa_hash_table_search(state->regs, dest->reg);
assert(entry2);
reg_validate_state *reg_state = (reg_validate_state *) entry2->data;
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
_mesa_set_add(reg_state->defs, dest);
if (!dest->reg->is_global) {
assert(reg_state->where_defined == state->impl &&
"writing to a register declared in a different function");
}
assert((dest->reg->num_array_elems == 0 ||
dest->base_offset < dest->reg->num_array_elems) &&
"definitely out-of-bounds array access");
if (dest->indirect) {
assert(dest->reg->num_array_elems != 0);
assert((dest->indirect->is_ssa || dest->indirect->reg.indirect == NULL) &&
"only one level of indirection allowed");
validate_src(dest->indirect, state);
}
}
static void
validate_ssa_def(nir_ssa_def *def, validate_state *state)
{
assert(def->index < state->impl->ssa_alloc);
assert(!BITSET_TEST(state->ssa_defs_found, def->index));
BITSET_SET(state->ssa_defs_found, def->index);
assert(def->parent_instr == state->instr);
assert(def->num_components <= 4);
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
list_validate(&def->uses);
list_validate(&def->if_uses);
ssa_def_validate_state *def_state = ralloc(state->ssa_defs,
ssa_def_validate_state);
def_state->where_defined = state->impl;
def_state->uses = _mesa_set_create(def_state, _mesa_hash_pointer,
_mesa_key_pointer_equal);
def_state->if_uses = _mesa_set_create(def_state, _mesa_hash_pointer,
_mesa_key_pointer_equal);
_mesa_hash_table_insert(state->ssa_defs, def, def_state);
}
static void
validate_dest(nir_dest *dest, validate_state *state)
{
if (dest->is_ssa)
validate_ssa_def(&dest->ssa, state);
else
validate_reg_dest(&dest->reg, state);
}
static void
validate_alu_dest(nir_alu_instr *instr, validate_state *state)
{
nir_alu_dest *dest = &instr->dest;
unsigned dest_size =
dest->dest.is_ssa ? dest->dest.ssa.num_components
: dest->dest.reg.reg->num_components;
bool is_packed = !dest->dest.is_ssa && dest->dest.reg.reg->is_packed;
/*
* validate that the instruction doesn't write to components not in the
* register/SSA value
*/
assert(is_packed || !(dest->write_mask & ~((1 << dest_size) - 1)));
/* validate that saturate is only ever used on instructions with
* destinations of type float
*/
nir_alu_instr *alu = nir_instr_as_alu(state->instr);
assert(nir_op_infos[alu->op].output_type == nir_type_float ||
!dest->saturate);
unsigned bit_size = dest->dest.is_ssa ? dest->dest.ssa.bit_size
: dest->dest.reg.reg->bit_size;
nir_alu_type type = nir_op_infos[instr->op].output_type;
/* 8-bit float isn't a thing */
if (nir_alu_type_get_base_type(type) == nir_type_float)
assert(bit_size == 16 || bit_size == 32 || bit_size == 64);
assert(nir_alu_type_get_type_size(type) == 0 ||
nir_alu_type_get_type_size(type) == bit_size);
validate_dest(&dest->dest, state);
}
static void
validate_alu_instr(nir_alu_instr *instr, validate_state *state)
{
assert(instr->op < nir_num_opcodes);
for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
validate_alu_src(instr, i, state);
}
validate_alu_dest(instr, state);
}
static void
validate_deref_chain(nir_deref *deref, validate_state *state)
{
assert(deref->child == NULL || ralloc_parent(deref->child) == deref);
nir_deref *parent = NULL;
while (deref != NULL) {
switch (deref->deref_type) {
case nir_deref_type_array:
assert(deref->type == glsl_get_array_element(parent->type));
if (nir_deref_as_array(deref)->deref_array_type ==
nir_deref_array_type_indirect)
validate_src(&nir_deref_as_array(deref)->indirect, state);
break;
case nir_deref_type_struct:
assert(deref->type ==
glsl_get_struct_field(parent->type,
nir_deref_as_struct(deref)->index));
break;
case nir_deref_type_var:
break;
default:
assert(!"Invalid deref type");
break;
}
parent = deref;
deref = deref->child;
}
}
static void
validate_var_use(nir_variable *var, validate_state *state)
{
if (var->data.mode == nir_var_local) {
struct hash_entry *entry = _mesa_hash_table_search(state->var_defs, var);
assert(entry);
assert((nir_function_impl *) entry->data == state->impl);
}
}
static void
validate_deref_var(void *parent_mem_ctx, nir_deref_var *deref, validate_state *state)
{
assert(deref != NULL);
assert(ralloc_parent(deref) == parent_mem_ctx);
assert(deref->deref.type == deref->var->type);
validate_var_use(deref->var, state);
validate_deref_chain(&deref->deref, state);
}
static void
validate_intrinsic_instr(nir_intrinsic_instr *instr, validate_state *state)
{
unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
for (unsigned i = 0; i < num_srcs; i++) {
unsigned components_read =
nir_intrinsic_infos[instr->intrinsic].src_components[i];
if (components_read == 0)
components_read = instr->num_components;
assert(components_read > 0);
if (instr->src[i].is_ssa) {
assert(components_read <= instr->src[i].ssa->num_components);
} else if (!instr->src[i].reg.reg->is_packed) {
assert(components_read <= instr->src[i].reg.reg->num_components);
}
validate_src(&instr->src[i], state);
}
unsigned num_vars = nir_intrinsic_infos[instr->intrinsic].num_variables;
for (unsigned i = 0; i < num_vars; i++) {
validate_deref_var(instr, instr->variables[i], state);
}
if (nir_intrinsic_infos[instr->intrinsic].has_dest) {
unsigned components_written =
nir_intrinsic_infos[instr->intrinsic].dest_components;
if (components_written == 0)
components_written = instr->num_components;
assert(components_written > 0);
if (instr->dest.is_ssa) {
assert(components_written <= instr->dest.ssa.num_components);
} else if (!instr->dest.reg.reg->is_packed) {
assert(components_written <= instr->dest.reg.reg->num_components);
}
validate_dest(&instr->dest, state);
}
switch (instr->intrinsic) {
case nir_intrinsic_load_var: {
const struct glsl_type *type =
nir_deref_tail(&instr->variables[0]->deref)->type;
assert(glsl_type_is_vector_or_scalar(type) ||
(instr->variables[0]->var->data.mode == nir_var_uniform &&
glsl_get_base_type(type) == GLSL_TYPE_SUBROUTINE));
assert(instr->num_components == glsl_get_vector_elements(type));
break;
}
case nir_intrinsic_store_var: {
const struct glsl_type *type =
nir_deref_tail(&instr->variables[0]->deref)->type;
assert(glsl_type_is_vector_or_scalar(type) ||
(instr->variables[0]->var->data.mode == nir_var_uniform &&
glsl_get_base_type(type) == GLSL_TYPE_SUBROUTINE));
assert(instr->num_components == glsl_get_vector_elements(type));
assert(instr->variables[0]->var->data.mode != nir_var_shader_in &&
instr->variables[0]->var->data.mode != nir_var_uniform &&
instr->variables[0]->var->data.mode != nir_var_shader_storage);
assert((nir_intrinsic_write_mask(instr) & ~((1 << instr->num_components) - 1)) == 0);
break;
}
case nir_intrinsic_copy_var:
assert(nir_deref_tail(&instr->variables[0]->deref)->type ==
nir_deref_tail(&instr->variables[1]->deref)->type);
assert(instr->variables[0]->var->data.mode != nir_var_shader_in &&
instr->variables[0]->var->data.mode != nir_var_uniform &&
instr->variables[0]->var->data.mode != nir_var_shader_storage);
break;
default:
break;
}
}
static void
validate_tex_instr(nir_tex_instr *instr, validate_state *state)
{
bool src_type_seen[nir_num_tex_src_types];
for (unsigned i = 0; i < nir_num_tex_src_types; i++)
src_type_seen[i] = false;
for (unsigned i = 0; i < instr->num_srcs; i++) {
assert(!src_type_seen[instr->src[i].src_type]);
src_type_seen[instr->src[i].src_type] = true;
validate_src(&instr->src[i].src, state);
}
if (instr->texture != NULL)
validate_deref_var(instr, instr->texture, state);
if (instr->sampler != NULL)
validate_deref_var(instr, instr->sampler, state);
validate_dest(&instr->dest, state);
}
static void
validate_call_instr(nir_call_instr *instr, validate_state *state)
{
if (instr->return_deref == NULL) {
assert(glsl_type_is_void(instr->callee->return_type));
} else {
assert(instr->return_deref->deref.type == instr->callee->return_type);
validate_deref_var(instr, instr->return_deref, state);
}
assert(instr->num_params == instr->callee->num_params);
for (unsigned i = 0; i < instr->num_params; i++) {
assert(instr->callee->params[i].type == instr->params[i]->deref.type);
validate_deref_var(instr, instr->params[i], state);
}
}
static void
validate_load_const_instr(nir_load_const_instr *instr, validate_state *state)
{
validate_ssa_def(&instr->def, state);
}
static void
validate_ssa_undef_instr(nir_ssa_undef_instr *instr, validate_state *state)
{
validate_ssa_def(&instr->def, state);
}
static void
validate_phi_instr(nir_phi_instr *instr, validate_state *state)
{
/*
* don't validate the sources until we get to them from their predecessor
* basic blocks, to avoid validating an SSA use before its definition.
*/
validate_dest(&instr->dest, state);
exec_list_validate(&instr->srcs);
assert(exec_list_length(&instr->srcs) ==
state->block->predecessors->entries);
}
static void
validate_instr(nir_instr *instr, validate_state *state)
{
assert(instr->block == state->block);
state->instr = instr;
switch (instr->type) {
case nir_instr_type_alu:
validate_alu_instr(nir_instr_as_alu(instr), state);
break;
case nir_instr_type_call:
validate_call_instr(nir_instr_as_call(instr), state);
break;
case nir_instr_type_intrinsic:
validate_intrinsic_instr(nir_instr_as_intrinsic(instr), state);
break;
case nir_instr_type_tex:
validate_tex_instr(nir_instr_as_tex(instr), state);
break;
case nir_instr_type_load_const:
validate_load_const_instr(nir_instr_as_load_const(instr), state);
break;
case nir_instr_type_phi:
validate_phi_instr(nir_instr_as_phi(instr), state);
break;
case nir_instr_type_ssa_undef:
validate_ssa_undef_instr(nir_instr_as_ssa_undef(instr), state);
break;
case nir_instr_type_jump:
break;
default:
assert(!"Invalid ALU instruction type");
break;
}
state->instr = NULL;
}
static void
validate_phi_src(nir_phi_instr *instr, nir_block *pred, validate_state *state)
{
state->instr = &instr->instr;
assert(instr->dest.is_ssa);
exec_list_validate(&instr->srcs);
nir_foreach_phi_src(src, instr) {
if (src->pred == pred) {
assert(src->src.is_ssa);
assert(src->src.ssa->num_components ==
instr->dest.ssa.num_components);
validate_src(&src->src, state);
state->instr = NULL;
return;
}
}
abort();
}
static void
validate_phi_srcs(nir_block *block, nir_block *succ, validate_state *state)
{
nir_foreach_instr(instr, succ) {
if (instr->type != nir_instr_type_phi)
break;
validate_phi_src(nir_instr_as_phi(instr), block, state);
}
}
static void validate_cf_node(nir_cf_node *node, validate_state *state);
static void
validate_block(nir_block *block, validate_state *state)
{
assert(block->cf_node.parent == state->parent_node);
state->block = block;
exec_list_validate(&block->instr_list);
nir_foreach_instr(instr, block) {
if (instr->type == nir_instr_type_phi) {
assert(instr == nir_block_first_instr(block) ||
nir_instr_prev(instr)->type == nir_instr_type_phi);
}
if (instr->type == nir_instr_type_jump) {
assert(instr == nir_block_last_instr(block));
}
validate_instr(instr, state);
}
assert(block->successors[0] != NULL);
assert(block->successors[0] != block->successors[1]);
for (unsigned i = 0; i < 2; i++) {
if (block->successors[i] != NULL) {
struct set_entry *entry =
_mesa_set_search(block->successors[i]->predecessors, block);
assert(entry);
validate_phi_srcs(block, block->successors[i], state);
}
}
struct set_entry *entry;
set_foreach(block->predecessors, entry) {
const nir_block *pred = entry->key;
assert(pred->successors[0] == block ||
pred->successors[1] == block);
}
if (!exec_list_is_empty(&block->instr_list) &&
nir_block_last_instr(block)->type == nir_instr_type_jump) {
assert(block->successors[1] == NULL);
nir_jump_instr *jump = nir_instr_as_jump(nir_block_last_instr(block));
switch (jump->type) {
case nir_jump_break: {
nir_block *after =
nir_cf_node_as_block(nir_cf_node_next(&state->loop->cf_node));
assert(block->successors[0] == after);
break;
}
case nir_jump_continue: {
nir_block *first =
nir_cf_node_as_block(nir_loop_first_cf_node(state->loop));
assert(block->successors[0] == first);
break;
}
case nir_jump_return:
assert(block->successors[0] == state->impl->end_block);
break;
default:
unreachable("bad jump type");
}
} else {
nir_cf_node *next = nir_cf_node_next(&block->cf_node);
if (next == NULL) {
switch (state->parent_node->type) {
case nir_cf_node_loop: {
nir_block *first =
nir_cf_node_as_block(nir_loop_first_cf_node(state->loop));
assert(block->successors[0] == first);
/* due to the hack for infinite loops, block->successors[1] may
* point to the block after the loop.
*/
break;
}
case nir_cf_node_if: {
nir_block *after =
nir_cf_node_as_block(nir_cf_node_next(state->parent_node));
assert(block->successors[0] == after);
assert(block->successors[1] == NULL);
break;
}
case nir_cf_node_function:
assert(block->successors[0] == state->impl->end_block);
assert(block->successors[1] == NULL);
break;
default:
unreachable("unknown control flow node type");
}
} else {
if (next->type == nir_cf_node_if) {
nir_if *if_stmt = nir_cf_node_as_if(next);
assert(&block->successors[0]->cf_node ==
nir_if_first_then_node(if_stmt));
assert(&block->successors[1]->cf_node ==
nir_if_first_else_node(if_stmt));
} else {
assert(next->type == nir_cf_node_loop);
nir_loop *loop = nir_cf_node_as_loop(next);
assert(&block->successors[0]->cf_node ==
nir_loop_first_cf_node(loop));
assert(block->successors[1] == NULL);
}
}
}
}
static void
validate_if(nir_if *if_stmt, validate_state *state)
{
state->if_stmt = if_stmt;
assert(!exec_node_is_head_sentinel(if_stmt->cf_node.node.prev));
nir_cf_node *prev_node = nir_cf_node_prev(&if_stmt->cf_node);
assert(prev_node->type == nir_cf_node_block);
assert(!exec_node_is_tail_sentinel(if_stmt->cf_node.node.next));
nir_cf_node *next_node = nir_cf_node_next(&if_stmt->cf_node);
assert(next_node->type == nir_cf_node_block);
validate_src(&if_stmt->condition, state);
assert(!exec_list_is_empty(&if_stmt->then_list));
assert(!exec_list_is_empty(&if_stmt->else_list));
nir_cf_node *old_parent = state->parent_node;
state->parent_node = &if_stmt->cf_node;
exec_list_validate(&if_stmt->then_list);
foreach_list_typed(nir_cf_node, cf_node, node, &if_stmt->then_list) {
validate_cf_node(cf_node, state);
}
exec_list_validate(&if_stmt->else_list);
foreach_list_typed(nir_cf_node, cf_node, node, &if_stmt->else_list) {
validate_cf_node(cf_node, state);
}
state->parent_node = old_parent;
state->if_stmt = NULL;
}
static void
validate_loop(nir_loop *loop, validate_state *state)
{
assert(!exec_node_is_head_sentinel(loop->cf_node.node.prev));
nir_cf_node *prev_node = nir_cf_node_prev(&loop->cf_node);
assert(prev_node->type == nir_cf_node_block);
assert(!exec_node_is_tail_sentinel(loop->cf_node.node.next));
nir_cf_node *next_node = nir_cf_node_next(&loop->cf_node);
assert(next_node->type == nir_cf_node_block);
assert(!exec_list_is_empty(&loop->body));
nir_cf_node *old_parent = state->parent_node;
state->parent_node = &loop->cf_node;
nir_loop *old_loop = state->loop;
state->loop = loop;
exec_list_validate(&loop->body);
foreach_list_typed(nir_cf_node, cf_node, node, &loop->body) {
validate_cf_node(cf_node, state);
}
state->parent_node = old_parent;
state->loop = old_loop;
}
static void
validate_cf_node(nir_cf_node *node, validate_state *state)
{
assert(node->parent == state->parent_node);
switch (node->type) {
case nir_cf_node_block:
validate_block(nir_cf_node_as_block(node), state);
break;
case nir_cf_node_if:
validate_if(nir_cf_node_as_if(node), state);
break;
case nir_cf_node_loop:
validate_loop(nir_cf_node_as_loop(node), state);
break;
default:
unreachable("Invalid CF node type");
}
}
static void
prevalidate_reg_decl(nir_register *reg, bool is_global, validate_state *state)
{
assert(reg->is_global == is_global);
if (is_global)
assert(reg->index < state->shader->reg_alloc);
else
assert(reg->index < state->impl->reg_alloc);
assert(!BITSET_TEST(state->regs_found, reg->index));
BITSET_SET(state->regs_found, reg->index);
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
list_validate(&reg->uses);
list_validate(&reg->defs);
list_validate(&reg->if_uses);
reg_validate_state *reg_state = ralloc(state->regs, reg_validate_state);
reg_state->uses = _mesa_set_create(reg_state, _mesa_hash_pointer,
_mesa_key_pointer_equal);
reg_state->if_uses = _mesa_set_create(reg_state, _mesa_hash_pointer,
_mesa_key_pointer_equal);
reg_state->defs = _mesa_set_create(reg_state, _mesa_hash_pointer,
_mesa_key_pointer_equal);
reg_state->where_defined = is_global ? NULL : state->impl;
_mesa_hash_table_insert(state->regs, reg, reg_state);
}
static void
postvalidate_reg_decl(nir_register *reg, validate_state *state)
{
struct hash_entry *entry = _mesa_hash_table_search(state->regs, reg);
reg_validate_state *reg_state = (reg_validate_state *) entry->data;
nir_foreach_use(src, reg) {
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
struct set_entry *entry = _mesa_set_search(reg_state->uses, src);
assert(entry);
_mesa_set_remove(reg_state->uses, entry);
}
if (reg_state->uses->entries != 0) {
printf("extra entries in register uses:\n");
struct set_entry *entry;
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
set_foreach(reg_state->uses, entry)
printf("%p\n", entry->key);
abort();
}
nir_foreach_if_use(src, reg) {
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
struct set_entry *entry = _mesa_set_search(reg_state->if_uses, src);
assert(entry);
_mesa_set_remove(reg_state->if_uses, entry);
}
if (reg_state->if_uses->entries != 0) {
printf("extra entries in register if_uses:\n");
struct set_entry *entry;
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
set_foreach(reg_state->if_uses, entry)
printf("%p\n", entry->key);
abort();
}
nir_foreach_def(src, reg) {
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
struct set_entry *entry = _mesa_set_search(reg_state->defs, src);
assert(entry);
_mesa_set_remove(reg_state->defs, entry);
}
if (reg_state->defs->entries != 0) {
printf("extra entries in register defs:\n");
struct set_entry *entry;
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
set_foreach(reg_state->defs, entry)
printf("%p\n", entry->key);
abort();
}
}
static void
validate_var_decl(nir_variable *var, bool is_global, validate_state *state)
{
assert(is_global == nir_variable_is_global(var));
/* Must have exactly one mode set */
assert(util_bitcount(var->data.mode) == 1);
/*
* TODO validate some things ir_validate.cpp does (requires more GLSL type
* support)
*/
if (!is_global) {
_mesa_hash_table_insert(state->var_defs, var, state->impl);
}
}
static bool
postvalidate_ssa_def(nir_ssa_def *def, void *void_state)
{
validate_state *state = void_state;
struct hash_entry *entry = _mesa_hash_table_search(state->ssa_defs, def);
ssa_def_validate_state *def_state = (ssa_def_validate_state *)entry->data;
nir_foreach_use(src, def) {
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
struct set_entry *entry = _mesa_set_search(def_state->uses, src);
assert(entry);
_mesa_set_remove(def_state->uses, entry);
}
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
if (def_state->uses->entries != 0) {
printf("extra entries in register uses:\n");
struct set_entry *entry;
set_foreach(def_state->uses, entry)
printf("%p\n", entry->key);
abort();
}
nir_foreach_if_use(src, def) {
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
struct set_entry *entry = _mesa_set_search(def_state->if_uses, src);
assert(entry);
_mesa_set_remove(def_state->if_uses, entry);
}
nir/nir: Use a linked list instead of a hash set for use/def sets This commit switches us from the current setup of using hash sets for use/def sets to using linked lists. Doing so should save us quite a bit of memory because we aren't carrying around 3 hash sets per register and 2 per SSA value. It should also save us CPU time because adding/removing things from use/def sets is 4 pointer manipulations instead of a hash lookup. Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists: GLSL IR Only: 586.4 +/- 1.653833 NIR with hash sets: 675.4 +/- 2.502108 NIR + use/def lists: 641.2 +/- 1.557043 I also ran a memory usage experiment with Ken's patch to delete GLSL IR and keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over and above what we gained by deleting the GLSL IR on the same dota trace. On the code complexity side of things, some things are now much easier and others are a bit harder. One of the operations we perform constantly in optimization passes is to replace one source with another. Due to the fact that an instruction can use the same SSA value multiple times, we had to iterate through the sources of the instruction and determine if the use we were replacing was the only one before removing it from the set of uses. With this patch, uses are per-source not per-instruction so we can just remove it safely. On the other hand, trying to iterate over all of the instructions that use a given value is more difficult. Fortunately, the two places we do that are the ffma peephole where it doesn't matter and GCM where we already gracefully handle duplicates visits to an instruction. Another aspect here is that using linked lists in this way can be tricky to get right. With sets, things were quite forgiving and the worst that happened if you didn't properly remove a use was that it would get caught in the validator. With linked lists, it can lead to linked list corruption which can be harder to track. However, we do just as much validation of the linked lists as we did of the sets so the validator should still catch these problems. While working on this series, the vast majority of the bugs I had to fix were caught by assertions. I don't think the lists are going to be that much worse than the sets. Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
2015-04-24 10:16:27 -07:00
if (def_state->if_uses->entries != 0) {
printf("extra entries in register uses:\n");
struct set_entry *entry;
set_foreach(def_state->if_uses, entry)
printf("%p\n", entry->key);
abort();
}
return true;
}
static void
validate_function_impl(nir_function_impl *impl, validate_state *state)
{
assert(impl->function->impl == impl);
assert(impl->cf_node.parent == NULL);
assert(impl->num_params == impl->function->num_params);
for (unsigned i = 0; i < impl->num_params; i++) {
assert(impl->params[i]->type == impl->function->params[i].type);
assert(impl->params[i]->data.mode == nir_var_param);
assert(impl->params[i]->data.location == i);
validate_var_decl(impl->params[i], false, state);
}
if (glsl_type_is_void(impl->function->return_type)) {
assert(impl->return_var == NULL);
} else {
assert(impl->return_var->type == impl->function->return_type);
assert(impl->return_var->data.mode == nir_var_param);
assert(impl->return_var->data.location == -1);
validate_var_decl(impl->return_var, false, state);
}
assert(exec_list_is_empty(&impl->end_block->instr_list));
assert(impl->end_block->successors[0] == NULL);
assert(impl->end_block->successors[1] == NULL);
state->impl = impl;
state->parent_node = &impl->cf_node;
exec_list_validate(&impl->locals);
nir_foreach_variable(var, &impl->locals) {
validate_var_decl(var, false, state);
}
state->regs_found = realloc(state->regs_found,
BITSET_WORDS(impl->reg_alloc) *
sizeof(BITSET_WORD));
memset(state->regs_found, 0, BITSET_WORDS(impl->reg_alloc) *
sizeof(BITSET_WORD));
exec_list_validate(&impl->registers);
foreach_list_typed(nir_register, reg, node, &impl->registers) {
prevalidate_reg_decl(reg, false, state);
}
state->ssa_defs_found = realloc(state->ssa_defs_found,
BITSET_WORDS(impl->ssa_alloc) *
sizeof(BITSET_WORD));
memset(state->ssa_defs_found, 0, BITSET_WORDS(impl->ssa_alloc) *
sizeof(BITSET_WORD));
exec_list_validate(&impl->body);
foreach_list_typed(nir_cf_node, node, node, &impl->body) {
validate_cf_node(node, state);
}
foreach_list_typed(nir_register, reg, node, &impl->registers) {
postvalidate_reg_decl(reg, state);
}
nir_foreach_block(block, impl) {
nir_foreach_instr(instr, block)
nir_foreach_ssa_def(instr, postvalidate_ssa_def, state);
}
}
static void
validate_function(nir_function *func, validate_state *state)
{
if (func->impl != NULL) {
assert(func->impl->function == func);
validate_function_impl(func->impl, state);
}
}
static void
init_validate_state(validate_state *state)
{
state->regs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
_mesa_key_pointer_equal);
state->ssa_defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
_mesa_key_pointer_equal);
state->ssa_defs_found = NULL;
state->regs_found = NULL;
state->var_defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
_mesa_key_pointer_equal);
state->loop = NULL;
}
static void
destroy_validate_state(validate_state *state)
{
_mesa_hash_table_destroy(state->regs, NULL);
_mesa_hash_table_destroy(state->ssa_defs, NULL);
free(state->ssa_defs_found);
free(state->regs_found);
_mesa_hash_table_destroy(state->var_defs, NULL);
}
void
nir_validate_shader(nir_shader *shader)
{
validate_state state;
init_validate_state(&state);
state.shader = shader;
exec_list_validate(&shader->uniforms);
nir_foreach_variable(var, &shader->uniforms) {
validate_var_decl(var, true, &state);
}
exec_list_validate(&shader->inputs);
nir_foreach_variable(var, &shader->inputs) {
validate_var_decl(var, true, &state);
}
exec_list_validate(&shader->outputs);
nir_foreach_variable(var, &shader->outputs) {
validate_var_decl(var, true, &state);
}
exec_list_validate(&shader->shared);
nir_foreach_variable(var, &shader->shared) {
validate_var_decl(var, true, &state);
}
exec_list_validate(&shader->globals);
nir_foreach_variable(var, &shader->globals) {
validate_var_decl(var, true, &state);
}
exec_list_validate(&shader->system_values);
nir_foreach_variable(var, &shader->system_values) {
validate_var_decl(var, true, &state);
}
state.regs_found = realloc(state.regs_found,
BITSET_WORDS(shader->reg_alloc) *
sizeof(BITSET_WORD));
memset(state.regs_found, 0, BITSET_WORDS(shader->reg_alloc) *
sizeof(BITSET_WORD));
exec_list_validate(&shader->registers);
foreach_list_typed(nir_register, reg, node, &shader->registers) {
prevalidate_reg_decl(reg, true, &state);
}
exec_list_validate(&shader->functions);
foreach_list_typed(nir_function, func, node, &shader->functions) {
validate_function(func, &state);
}
foreach_list_typed(nir_register, reg, node, &shader->registers) {
postvalidate_reg_decl(reg, &state);
}
destroy_validate_state(&state);
}
#endif /* NDEBUG */