summaryrefslogtreecommitdiff
path: root/tests/compute/generics-constrained.slang
diff options
context:
space:
mode:
authorYong He <yonghe@outlook.com>2024-06-03 14:12:48 -0700
committerGitHub <noreply@github.com>2024-06-03 14:12:48 -0700
commit89c1fd0dd1581221f583653a9dfa6d1cf990577c (patch)
treefe5b1832bd64225ddf03c0146aec6f6057f607b8 /tests/compute/generics-constrained.slang
parent68bf31e586b9d1805c19dd42492086c7c780c2a4 (diff)
Fix performance issue in source-map (#4261)
Diffstat (limited to 'tests/compute/generics-constrained.slang')
0 files changed, 0 insertions, 0 deletions
generated by cgit v1.2.3 (git 2.39.1) at 2026-09-06 17:51:08 +0000
02 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 
#include "slang-ir-defer-buffer-load.h"

#include "slang-ir-clone.h"
#include "slang-ir-dominators.h"
#include "slang-ir-insts.h"
#include "slang-ir-layout.h"
#include "slang-ir-redundancy-removal.h"
#include "slang-ir-util.h"
#include "slang-ir.h"

namespace Slang
{

// Generally, we want to specialize arguments that are large in size, or arguments that
// are arrays or composite type that contains arrays.
// This is because:
// 1. Struct types without arrays will eventually be SROA's into registers and then effectively
//    DCE'd, so they usually won't cause performance issues. In fact, front loading structs
//    and reusing the loaded value instead of repetitively loading from constant memory is
//    usually beneficial to performance. However large struct values can be SROA'd into a large
//    number of registers, causing slow downstream compilation. Therefore we should avoid/defer
//    loading them into registers if we can.
// 2. Arrays usually cannot be SROA'd into individual registers, which usually leads to
//    large register consumption if they ever get loaded, so we want to defer loading array
//    typed values as much as possible.

// If the argument data is bigger than this threshold, it is considered a large object
// and we will try to specialize it even if it doesn't contain arrays.
static const int kBufferLoadElementSizeSpecializationThreshold = 128;

// If the argument data is smaller than this threshold, it is considered a tiny object
// and we will not consider specializing it, even if it contains arrays.
static const int kBufferLoadElementSizeSpecializationMinThreshold = 16;

static bool isCompositeTypeContainingArrays(IRType* type)
{
    if (auto structType = as<IRStructType>(type))
    {
        for (auto field : structType->getFields())
        {
            if (const auto arrayType = as<IRArrayTypeBase>(field->getFieldType()))
            {
                return true;
            }
            if (auto subStructType = as<IRStructType>(field->getFieldType()))
            {
                if (isCompositeTypeContainingArrays(subStructType))
                    return true;
            }
        }
    }
    else if (as<IRArrayTypeBase>(type))
    {
        return true;
    }
    return false;
}

bool isTypePreferrableToDeferLoad(CodeGenContext* codeGenContext, IRType* type)
{
    // If parameter is a pointer/reference, we should consider specialize it.
    if (as<IROutParamTypeBase>(type) || as<IRRefParamType>(type) || as<IRBorrowInParamType>(type))
        return true;

    // We only want to defer loading values that are "large enough" that
    // we expect them to be expensive to pass by value.
    //
    IRSizeAndAlignment sizeAlignment = {};
    if (SLANG_FAILED(getNaturalSizeAndAlignment(
            codeGenContext->getTargetProgram()->getOptionSet(),
            type,
            &sizeAlignment)))
    {
        // If type contains fields that we don't know how to compute natural size
        // for, default to specialize if it contains arrays.
        return isCompositeTypeContainingArrays(type);
    }

    // If the argument is very small, don't bother specializing.
    if (sizeAlignment.size <= kBufferLoadElementSizeSpecializationMinThreshold)
        return false;

    // If the argument is somewhat small, don't specialize, unless it contains
    // arrays.
    if (sizeAlignment.size <= kBufferLoadElementSizeSpecializationThreshold)
    {
        // We generally do not specialize for small values, except it contains
        // arrays that usually present a challenge for the SROA pass to eliminate
        // unnecessary loads.
        if (!isCompositeTypeContainingArrays(type))
            return false;
    }
    return true;
}

// Returns true if memory loaded by `loadInst` is not modified before `userInst` after it is
// loaded.
// This method is currently implementing a very conservative analysis that only allows
// `loadInst` to be in the same block as `userInst`, with basic aliasing analysis for any
// stores in between. All other cases are conservatively treated as the memory location may be
// modified.
bool isMemoryLocationUnmodifiedBetweenLoadAndUser(
    TargetRequest* target,
    IRInst* loadInst,
    IRInst* userInst)
{
    auto func = getParentFunc(loadInst);
    if (!func)
        return false;

    // For now we only check if loadInst and userInst are in the same block.
    if (loadInst->getParent() != userInst->getParent())
        return false;

    for (IRInst* inst = loadInst->getNextInst(); inst; inst = inst->getNextInst())
    {
        // We found callInst before hitting any instruction that may modify the memory.
        if (inst == userInst)
            return true;

        if (!inst->mightHaveSideEffects())
            continue;

        // If we see any inst that has side effect, check if it is simple case that we can rule
        // out the possibility of modifying the memory location.
        switch (inst->getOp())
        {
        case kIROp_Store:
            {
                auto storedDest = inst->getOperand(0);