src/gallium/drivers/radeonsi/si_shader_llvm.c

   1 /*
   2  * Copyright 2016 Advanced Micro Devices, Inc.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * on the rights to use, copy, modify, merge, publish, distribute, sub
   9  * license, and/or sell copies of the Software, and to permit persons to whom
  10  * the Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 #include "ac_nir_to_llvm.h"
  26 #include "ac_rtld.h"
  27 #include "si_pipe.h"
  28 #include "si_shader_internal.h"
  29 #include "sid.h"
  30 #include "tgsi/tgsi_from_mesa.h"
  31 #include "util/u_memory.h"
  32
  33 struct si_llvm_diagnostics {
  34    struct pipe_debug_callback *debug;
  35    unsigned retval;
  36 };
  37
  38 static void si_diagnostic_handler(LLVMDiagnosticInfoRef di, void *context)
  39 {
  40    struct si_llvm_diagnostics *diag = (struct si_llvm_diagnostics *)context;
  41    LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
  42    const char *severity_str = NULL;
  43
  44    switch (severity) {
  45    case LLVMDSError:
  46       severity_str = "error";
  47       break;
  48    case LLVMDSWarning:
  49       severity_str = "warning";
  50       break;
  51    case LLVMDSRemark:
  52    case LLVMDSNote:
  53    default:
  54       return;
  55    }
  56
  57    char *description = LLVMGetDiagInfoDescription(di);
  58
  59    pipe_debug_message(diag->debug, SHADER_INFO, "LLVM diagnostic (%s): %s", severity_str,
  60                       description);
  61
  62    if (severity == LLVMDSError) {
  63       diag->retval = 1;
  64       fprintf(stderr, "LLVM triggered Diagnostic Handler: %s\n", description);
  65    }
  66
  67    LLVMDisposeMessage(description);
  68 }
  69
  70 bool si_compile_llvm(struct si_screen *sscreen, struct si_shader_binary *binary,
  71                      struct ac_shader_config *conf, struct ac_llvm_compiler *compiler,
  72                      struct ac_llvm_context *ac, struct pipe_debug_callback *debug,
  73                      gl_shader_stage stage, const char *name, bool less_optimized)
  74 {
  75    unsigned count = p_atomic_inc_return(&sscreen->num_compilations);
  76
  77    if (si_can_dump_shader(sscreen, stage)) {
  78       fprintf(stderr, "radeonsi: Compiling shader %d\n", count);
  79
  80       if (!(sscreen->debug_flags & (DBG(NO_IR) | DBG(PREOPT_IR)))) {
  81          fprintf(stderr, "%s LLVM IR:\n\n", name);
  82          ac_dump_module(ac->module);
  83          fprintf(stderr, "\n");
  84       }
  85    }
  86
  87    if (sscreen->record_llvm_ir) {
  88       char *ir = LLVMPrintModuleToString(ac->module);
  89       binary->llvm_ir_string = strdup(ir);
  90       LLVMDisposeMessage(ir);
  91    }
  92
  93    if (!si_replace_shader(count, binary)) {
  94       struct ac_compiler_passes *passes = compiler->passes;
  95
  96       if (ac->wave_size == 32)
  97          passes = compiler->passes_wave32;
  98       else if (less_optimized && compiler->low_opt_passes)
  99          passes = compiler->low_opt_passes;
 100
 101       struct si_llvm_diagnostics diag = {debug};
 102       LLVMContextSetDiagnosticHandler(ac->context, si_diagnostic_handler, &diag);
 103
 104       if (!ac_compile_module_to_elf(passes, ac->module, (char **)&binary->elf_buffer,
 105                                     &binary->elf_size))
 106          diag.retval = 1;
 107
 108       if (diag.retval != 0) {
 109          pipe_debug_message(debug, SHADER_INFO, "LLVM compilation failed");
 110          return false;
 111       }
 112    }
 113
 114    struct ac_rtld_binary rtld;
 115    if (!ac_rtld_open(&rtld, (struct ac_rtld_open_info){
 116                                .info = &sscreen->info,
 117                                .shader_type = stage,
 118                                .wave_size = ac->wave_size,
 119                                .num_parts = 1,
 120                                .elf_ptrs = &binary->elf_buffer,
 121                                .elf_sizes = &binary->elf_size}))
 122       return false;
 123
 124    bool ok = ac_rtld_read_config(&sscreen->info, &rtld, conf);
 125    ac_rtld_close(&rtld);
 126    return ok;
 127 }
 128
 129 void si_llvm_context_init(struct si_shader_context *ctx, struct si_screen *sscreen,
 130                           struct ac_llvm_compiler *compiler, unsigned wave_size)
 131 {
 132    memset(ctx, 0, sizeof(*ctx));
 133    ctx->screen = sscreen;
 134    ctx->compiler = compiler;
 135
 136    ac_llvm_context_init(&ctx->ac, compiler, sscreen->info.chip_class, sscreen->info.family,
 137                         AC_FLOAT_MODE_DEFAULT_OPENGL, wave_size, 64);
 138 }
 139
 140 void si_llvm_create_func(struct si_shader_context *ctx, const char *name, LLVMTypeRef *return_types,
 141                          unsigned num_return_elems, unsigned max_workgroup_size)
 142 {
 143    LLVMTypeRef ret_type;
 144    enum ac_llvm_calling_convention call_conv;
 145
 146    if (num_return_elems)
 147       ret_type = LLVMStructTypeInContext(ctx->ac.context, return_types, num_return_elems, true);
 148    else
 149       ret_type = ctx->ac.voidt;
 150
 151    gl_shader_stage real_stage = ctx->stage;
 152
 153    /* LS is merged into HS (TCS), and ES is merged into GS. */
 154    if (ctx->screen->info.chip_class >= GFX9) {
 155       if (ctx->shader->key.as_ls)
 156          real_stage = MESA_SHADER_TESS_CTRL;
 157       else if (ctx->shader->key.as_es || ctx->shader->key.as_ngg)
 158          real_stage = MESA_SHADER_GEOMETRY;
 159    }
 160
 161    switch (real_stage) {
 162    case MESA_SHADER_VERTEX:
 163    case MESA_SHADER_TESS_EVAL:
 164       call_conv = AC_LLVM_AMDGPU_VS;
 165       break;
 166    case MESA_SHADER_TESS_CTRL:
 167       call_conv = AC_LLVM_AMDGPU_HS;
 168       break;
 169    case MESA_SHADER_GEOMETRY:
 170       call_conv = AC_LLVM_AMDGPU_GS;
 171       break;
 172    case MESA_SHADER_FRAGMENT:
 173       call_conv = AC_LLVM_AMDGPU_PS;
 174       break;
 175    case MESA_SHADER_COMPUTE:
 176       call_conv = AC_LLVM_AMDGPU_CS;
 177       break;
 178    default:
 179       unreachable("Unhandle shader type");
 180    }
 181
 182    /* Setup the function */
 183    ctx->return_type = ret_type;
 184    ctx->main_fn = ac_build_main(&ctx->args, &ctx->ac, call_conv, name, ret_type, ctx->ac.module);
 185    ctx->return_value = LLVMGetUndef(ctx->return_type);
 186
 187    if (ctx->screen->info.address32_hi) {
 188       ac_llvm_add_target_dep_function_attr(ctx->main_fn, "amdgpu-32bit-address-high-bits",
 189                                            ctx->screen->info.address32_hi);
 190    }
 191
 192    ac_llvm_set_workgroup_size(ctx->main_fn, max_workgroup_size);
 193 }
 194
 195 void si_llvm_optimize_module(struct si_shader_context *ctx)
 196 {
 197    /* Dump LLVM IR before any optimization passes */
 198    if (ctx->screen->debug_flags & DBG(PREOPT_IR) && si_can_dump_shader(ctx->screen, ctx->stage))
 199       LLVMDumpModule(ctx->ac.module);
 200
 201    /* Run the pass */
 202    LLVMRunPassManager(ctx->compiler->passmgr, ctx->ac.module);
 203    LLVMDisposeBuilder(ctx->ac.builder);
 204 }
 205
 206 void si_llvm_dispose(struct si_shader_context *ctx)
 207 {
 208    LLVMDisposeModule(ctx->ac.module);
 209    LLVMContextDispose(ctx->ac.context);
 210    ac_llvm_context_dispose(&ctx->ac);
 211 }
 212
 213 /**
 214  * Load a dword from a constant buffer.
 215  */
 216 LLVMValueRef si_buffer_load_const(struct si_shader_context *ctx, LLVMValueRef resource,
 217                                   LLVMValueRef offset)
 218 {
 219    return ac_build_buffer_load(&ctx->ac, resource, 1, NULL, offset, NULL, 0, 0, true, true);
 220 }
 221
 222 void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret)
 223 {
 224    if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
 225       LLVMBuildRetVoid(ctx->ac.builder);
 226    else
 227       LLVMBuildRet(ctx->ac.builder, ret);
 228 }
 229
 230 LLVMValueRef si_insert_input_ret(struct si_shader_context *ctx, LLVMValueRef ret,
 231                                  struct ac_arg param, unsigned return_index)
 232 {
 233    return LLVMBuildInsertValue(ctx->ac.builder, ret, ac_get_arg(&ctx->ac, param), return_index, "");
 234 }
 235
 236 LLVMValueRef si_insert_input_ret_float(struct si_shader_context *ctx, LLVMValueRef ret,
 237                                        struct ac_arg param, unsigned return_index)
 238 {
 239    LLVMBuilderRef builder = ctx->ac.builder;
 240    LLVMValueRef p = ac_get_arg(&ctx->ac, param);
 241
 242    return LLVMBuildInsertValue(builder, ret, ac_to_float(&ctx->ac, p), return_index, "");
 243 }
 244
 245 LLVMValueRef si_insert_input_ptr(struct si_shader_context *ctx, LLVMValueRef ret,
 246                                  struct ac_arg param, unsigned return_index)
 247 {
 248    LLVMBuilderRef builder = ctx->ac.builder;
 249    LLVMValueRef ptr = ac_get_arg(&ctx->ac, param);
 250    ptr = LLVMBuildPtrToInt(builder, ptr, ctx->ac.i32, "");
 251    return LLVMBuildInsertValue(builder, ret, ptr, return_index, "");
 252 }
 253
 254 LLVMValueRef si_prolog_get_rw_buffers(struct si_shader_context *ctx)
 255 {
 256    LLVMValueRef ptr[2], list;
 257    bool merged_shader = si_is_merged_shader(ctx->shader);
 258
 259    ptr[0] = LLVMGetParam(ctx->main_fn, (merged_shader ? 8 : 0) + SI_SGPR_RW_BUFFERS);
 260    list =
 261       LLVMBuildIntToPtr(ctx->ac.builder, ptr[0], ac_array_in_const32_addr_space(ctx->ac.v4i32), "");
 262    return list;
 263 }
 264
 265 LLVMValueRef si_build_gather_64bit(struct si_shader_context *ctx, LLVMTypeRef type,
 266                                    LLVMValueRef val1, LLVMValueRef val2)
 267 {
 268    LLVMValueRef values[2] = {
 269       ac_to_integer(&ctx->ac, val1),
 270       ac_to_integer(&ctx->ac, val2),
 271    };
 272    LLVMValueRef result = ac_build_gather_values(&ctx->ac, values, 2);
 273    return LLVMBuildBitCast(ctx->ac.builder, result, type, "");
 274 }
 275
 276 void si_llvm_emit_barrier(struct si_shader_context *ctx)
 277 {
 278    /* GFX6 only (thanks to a hw bug workaround):
 279     * The real barrier instruction isn’t needed, because an entire patch
 280     * always fits into a single wave.
 281     */
 282    if (ctx->screen->info.chip_class == GFX6 && ctx->stage == MESA_SHADER_TESS_CTRL) {
 283       ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
 284       return;
 285    }
 286
 287    ac_build_s_barrier(&ctx->ac);
 288 }
 289
 290 /* Ensure that the esgs ring is declared.
 291  *
 292  * We declare it with 64KB alignment as a hint that the
 293  * pointer value will always be 0.
 294  */
 295 void si_llvm_declare_esgs_ring(struct si_shader_context *ctx)
 296 {
 297    if (ctx->esgs_ring)
 298       return;
 299
 300    assert(!LLVMGetNamedGlobal(ctx->ac.module, "esgs_ring"));
 301
 302    ctx->esgs_ring = LLVMAddGlobalInAddressSpace(ctx->ac.module, LLVMArrayType(ctx->ac.i32, 0),
 303                                                 "esgs_ring", AC_ADDR_SPACE_LDS);
 304    LLVMSetLinkage(ctx->esgs_ring, LLVMExternalLinkage);
 305    LLVMSetAlignment(ctx->esgs_ring, 64 * 1024);
 306 }
 307
 308 void si_init_exec_from_input(struct si_shader_context *ctx, struct ac_arg param, unsigned bitoffset)
 309 {
 310    LLVMValueRef args[] = {
 311       ac_get_arg(&ctx->ac, param),
 312       LLVMConstInt(ctx->ac.i32, bitoffset, 0),
 313    };
 314    ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.init.exec.from.input", ctx->ac.voidt, args, 2,
 315                       AC_FUNC_ATTR_CONVERGENT);
 316 }
 317
 318 /**
 319  * Get the value of a shader input parameter and extract a bitfield.
 320  */
 321 static LLVMValueRef unpack_llvm_param(struct si_shader_context *ctx, LLVMValueRef value,
 322                                       unsigned rshift, unsigned bitwidth)
 323 {
 324    if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMFloatTypeKind)
 325       value = ac_to_integer(&ctx->ac, value);
 326
 327    if (rshift)
 328       value = LLVMBuildLShr(ctx->ac.builder, value, LLVMConstInt(ctx->ac.i32, rshift, 0), "");
 329
 330    if (rshift + bitwidth < 32) {
 331       unsigned mask = (1 << bitwidth) - 1;
 332       value = LLVMBuildAnd(ctx->ac.builder, value, LLVMConstInt(ctx->ac.i32, mask, 0), "");
 333    }
 334
 335    return value;
 336 }
 337
 338 LLVMValueRef si_unpack_param(struct si_shader_context *ctx, struct ac_arg param, unsigned rshift,
 339                              unsigned bitwidth)
 340 {
 341    LLVMValueRef value = ac_get_arg(&ctx->ac, param);
 342
 343    return unpack_llvm_param(ctx, value, rshift, bitwidth);
 344 }
 345
 346 LLVMValueRef si_get_primitive_id(struct si_shader_context *ctx, unsigned swizzle)
 347 {
 348    if (swizzle > 0)
 349       return ctx->ac.i32_0;
 350
 351    switch (ctx->stage) {
 352    case MESA_SHADER_VERTEX:
 353       return ac_get_arg(&ctx->ac, ctx->vs_prim_id);
 354    case MESA_SHADER_TESS_CTRL:
 355       return ac_get_arg(&ctx->ac, ctx->args.tcs_patch_id);
 356    case MESA_SHADER_TESS_EVAL:
 357       return ac_get_arg(&ctx->ac, ctx->args.tes_patch_id);
 358    case MESA_SHADER_GEOMETRY:
 359       return ac_get_arg(&ctx->ac, ctx->args.gs_prim_id);
 360    default:
 361       assert(0);
 362       return ctx->ac.i32_0;
 363    }
 364 }
 365
 366 LLVMValueRef si_llvm_get_block_size(struct ac_shader_abi *abi)
 367 {
 368    struct si_shader_context *ctx = si_shader_context_from_abi(abi);
 369
 370    LLVMValueRef values[3];
 371    LLVMValueRef result;
 372    unsigned i;
 373    unsigned *properties = ctx->shader->selector->info.properties;
 374
 375    if (properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] != 0) {
 376       unsigned sizes[3] = {properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH],
 377                            properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT],
 378                            properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH]};
 379
 380       for (i = 0; i < 3; ++i)
 381          values[i] = LLVMConstInt(ctx->ac.i32, sizes[i], 0);
 382
 383       result = ac_build_gather_values(&ctx->ac, values, 3);
 384    } else {
 385       result = ac_get_arg(&ctx->ac, ctx->block_size);
 386    }
 387
 388    return result;
 389 }
 390
 391 void si_llvm_declare_compute_memory(struct si_shader_context *ctx)
 392 {
 393    struct si_shader_selector *sel = ctx->shader->selector;
 394    unsigned lds_size = sel->info.properties[TGSI_PROPERTY_CS_LOCAL_SIZE];
 395
 396    LLVMTypeRef i8p = LLVMPointerType(ctx->ac.i8, AC_ADDR_SPACE_LDS);
 397    LLVMValueRef var;
 398
 399    assert(!ctx->ac.lds);
 400
 401    var = LLVMAddGlobalInAddressSpace(ctx->ac.module, LLVMArrayType(ctx->ac.i8, lds_size),
 402                                      "compute_lds", AC_ADDR_SPACE_LDS);
 403    LLVMSetAlignment(var, 64 * 1024);
 404
 405    ctx->ac.lds = LLVMBuildBitCast(ctx->ac.builder, var, i8p, "");
 406 }
 407
 408 bool si_nir_build_llvm(struct si_shader_context *ctx, struct nir_shader *nir)
 409 {
 410    if (nir->info.stage == MESA_SHADER_VERTEX) {
 411       si_llvm_load_vs_inputs(ctx, nir);
 412    } else if (nir->info.stage == MESA_SHADER_FRAGMENT) {
 413       unsigned colors_read = ctx->shader->selector->info.colors_read;
 414       LLVMValueRef main_fn = ctx->main_fn;
 415
 416       LLVMValueRef undef = LLVMGetUndef(ctx->ac.f32);
 417
 418       unsigned offset = SI_PARAM_POS_FIXED_PT + 1;
 419
 420       if (colors_read & 0x0f) {
 421          unsigned mask = colors_read & 0x0f;
 422          LLVMValueRef values[4];
 423          values[0] = mask & 0x1 ? LLVMGetParam(main_fn, offset++) : undef;
 424          values[1] = mask & 0x2 ? LLVMGetParam(main_fn, offset++) : undef;
 425          values[2] = mask & 0x4 ? LLVMGetParam(main_fn, offset++) : undef;
 426          values[3] = mask & 0x8 ? LLVMGetParam(main_fn, offset++) : undef;
 427          ctx->abi.color0 = ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, 4));
 428       }
 429       if (colors_read & 0xf0) {
 430          unsigned mask = (colors_read & 0xf0) >> 4;
 431          LLVMValueRef values[4];
 432          values[0] = mask & 0x1 ? LLVMGetParam(main_fn, offset++) : undef;
 433          values[1] = mask & 0x2 ? LLVMGetParam(main_fn, offset++) : undef;
 434          values[2] = mask & 0x4 ? LLVMGetParam(main_fn, offset++) : undef;
 435          values[3] = mask & 0x8 ? LLVMGetParam(main_fn, offset++) : undef;
 436          ctx->abi.color1 = ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, 4));
 437       }
 438
 439       ctx->abi.interp_at_sample_force_center =
 440          ctx->shader->key.mono.u.ps.interpolate_at_sample_force_center;
 441
 442       ctx->abi.kill_ps_if_inf_interp =
 443          (ctx->screen->debug_flags & DBG(KILL_PS_INF_INTERP)) &&
 444          (ctx->shader->selector->info.uses_persp_center ||
 445           ctx->shader->selector->info.uses_persp_centroid ||
 446           ctx->shader->selector->info.uses_persp_sample);
 447
 448    } else if (nir->info.stage == MESA_SHADER_COMPUTE) {
 449       if (nir->info.cs.user_data_components_amd) {
 450          ctx->abi.user_data = ac_get_arg(&ctx->ac, ctx->cs_user_data);
 451          ctx->abi.user_data = ac_build_expand_to_vec4(&ctx->ac, ctx->abi.user_data,
 452                                                       nir->info.cs.user_data_components_amd);
 453       }
 454    }
 455
 456    ctx->abi.inputs = &ctx->inputs[0];
 457    ctx->abi.clamp_shadow_reference = true;
 458    ctx->abi.robust_buffer_access = true;
 459    ctx->abi.convert_undef_to_zero = true;
 460    ctx->abi.clamp_div_by_zero = ctx->screen->options.clamp_div_by_zero;
 461
 462    if (ctx->shader->selector->info.properties[TGSI_PROPERTY_CS_LOCAL_SIZE]) {
 463       assert(gl_shader_stage_is_compute(nir->info.stage));
 464       si_llvm_declare_compute_memory(ctx);
 465    }
 466
 467    const struct si_shader_info *info = &ctx->shader->selector->info;
 468    for (unsigned i = 0; i < info->num_outputs; i++) {
 469       for (unsigned j = 0; j < 4; j++)
 470          ctx->abi.outputs[i * 4 + j] = ac_build_alloca_undef(&ctx->ac, ctx->ac.f32, "");
 471    }
 472
 473    ac_nir_translate(&ctx->ac, &ctx->abi, &ctx->args, nir);
 474
 475    return true;
 476 }
 477
 478 /**
 479  * Given a list of shader part functions, build a wrapper function that
 480  * runs them in sequence to form a monolithic shader.
 481  */
 482 void si_build_wrapper_function(struct si_shader_context *ctx, LLVMValueRef *parts,
 483                                unsigned num_parts, unsigned main_part,
 484                                unsigned next_shader_first_part)
 485 {
 486    LLVMBuilderRef builder = ctx->ac.builder;
 487    /* PS epilog has one arg per color component; gfx9 merged shader
 488     * prologs need to forward 40 SGPRs.
 489     */
 490    LLVMValueRef initial[AC_MAX_ARGS], out[AC_MAX_ARGS];
 491    LLVMTypeRef function_type;
 492    unsigned num_first_params;
 493    unsigned num_out, initial_num_out;
 494    ASSERTED unsigned num_out_sgpr;         /* used in debug checks */
 495    ASSERTED unsigned initial_num_out_sgpr; /* used in debug checks */
 496    unsigned num_sgprs, num_vgprs;
 497    unsigned gprs;
 498
 499    memset(&ctx->args, 0, sizeof(ctx->args));
 500
 501    for (unsigned i = 0; i < num_parts; ++i) {
 502       ac_add_function_attr(ctx->ac.context, parts[i], -1, AC_FUNC_ATTR_ALWAYSINLINE);
 503       LLVMSetLinkage(parts[i], LLVMPrivateLinkage);
 504    }
 505
 506    /* The parameters of the wrapper function correspond to those of the
 507     * first part in terms of SGPRs and VGPRs, but we use the types of the
 508     * main part to get the right types. This is relevant for the
 509     * dereferenceable attribute on descriptor table pointers.
 510     */
 511    num_sgprs = 0;
 512    num_vgprs = 0;
 513
 514    function_type = LLVMGetElementType(LLVMTypeOf(parts[0]));
 515    num_first_params = LLVMCountParamTypes(function_type);
 516
 517    for (unsigned i = 0; i < num_first_params; ++i) {
 518       LLVMValueRef param = LLVMGetParam(parts[0], i);
 519
 520       if (ac_is_sgpr_param(param)) {
 521          assert(num_vgprs == 0);
 522          num_sgprs += ac_get_type_size(LLVMTypeOf(param)) / 4;
 523       } else {
 524          num_vgprs += ac_get_type_size(LLVMTypeOf(param)) / 4;
 525       }
 526    }
 527
 528    gprs = 0;
 529    while (gprs < num_sgprs + num_vgprs) {
 530       LLVMValueRef param = LLVMGetParam(parts[main_part], ctx->args.arg_count);
 531       LLVMTypeRef type = LLVMTypeOf(param);
 532       unsigned size = ac_get_type_size(type) / 4;
 533
 534       /* This is going to get casted anyways, so we don't have to
 535        * have the exact same type. But we do have to preserve the
 536        * pointer-ness so that LLVM knows about it.
 537        */
 538       enum ac_arg_type arg_type = AC_ARG_INT;
 539       if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
 540          type = LLVMGetElementType(type);
 541
 542          if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
 543             if (LLVMGetVectorSize(type) == 4)
 544                arg_type = AC_ARG_CONST_DESC_PTR;
 545             else if (LLVMGetVectorSize(type) == 8)
 546                arg_type = AC_ARG_CONST_IMAGE_PTR;
 547             else
 548                assert(0);
 549          } else if (type == ctx->ac.f32) {
 550             arg_type = AC_ARG_CONST_FLOAT_PTR;
 551          } else {
 552             assert(0);
 553          }
 554       }
 555
 556       ac_add_arg(&ctx->args, gprs < num_sgprs ? AC_ARG_SGPR : AC_ARG_VGPR, size, arg_type, NULL);
 557
 558       assert(ac_is_sgpr_param(param) == (gprs < num_sgprs));
 559       assert(gprs + size <= num_sgprs + num_vgprs &&
 560              (gprs >= num_sgprs || gprs + size <= num_sgprs));
 561
 562       gprs += size;
 563    }
 564
 565    /* Prepare the return type. */
 566    unsigned num_returns = 0;
 567    LLVMTypeRef returns[AC_MAX_ARGS], last_func_type, return_type;
 568
 569    last_func_type = LLVMGetElementType(LLVMTypeOf(parts[num_parts - 1]));
 570    return_type = LLVMGetReturnType(last_func_type);
 571
 572    switch (LLVMGetTypeKind(return_type)) {
 573    case LLVMStructTypeKind:
 574       num_returns = LLVMCountStructElementTypes(return_type);
 575       assert(num_returns <= ARRAY_SIZE(returns));
 576       LLVMGetStructElementTypes(return_type, returns);
 577       break;
 578    case LLVMVoidTypeKind:
 579       break;
 580    default:
 581       unreachable("unexpected type");
 582    }
 583
 584    si_llvm_create_func(ctx, "wrapper", returns, num_returns,
 585                        si_get_max_workgroup_size(ctx->shader));
 586
 587    if (si_is_merged_shader(ctx->shader))
 588       ac_init_exec_full_mask(&ctx->ac);
 589
 590    /* Record the arguments of the function as if they were an output of
 591     * a previous part.
 592     */
 593    num_out = 0;
 594    num_out_sgpr = 0;
 595
 596    for (unsigned i = 0; i < ctx->args.arg_count; ++i) {
 597       LLVMValueRef param = LLVMGetParam(ctx->main_fn, i);
 598       LLVMTypeRef param_type = LLVMTypeOf(param);
 599       LLVMTypeRef out_type = ctx->args.args[i].file == AC_ARG_SGPR ? ctx->ac.i32 : ctx->ac.f32;
 600       unsigned size = ac_get_type_size(param_type) / 4;
 601
 602       if (size == 1) {
 603          if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
 604             param = LLVMBuildPtrToInt(builder, param, ctx->ac.i32, "");
 605             param_type = ctx->ac.i32;
 606          }
 607
 608          if (param_type != out_type)
 609             param = LLVMBuildBitCast(builder, param, out_type, "");
 610          out[num_out++] = param;
 611       } else {
 612          LLVMTypeRef vector_type = LLVMVectorType(out_type, size);
 613
 614          if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
 615             param = LLVMBuildPtrToInt(builder, param, ctx->ac.i64, "");
 616             param_type = ctx->ac.i64;
 617          }
 618
 619          if (param_type != vector_type)
 620             param = LLVMBuildBitCast(builder, param, vector_type, "");
 621
 622          for (unsigned j = 0; j < size; ++j)
 623             out[num_out++] =
 624                LLVMBuildExtractElement(builder, param, LLVMConstInt(ctx->ac.i32, j, 0), "");
 625       }
 626
 627       if (ctx->args.args[i].file == AC_ARG_SGPR)
 628          num_out_sgpr = num_out;
 629    }
 630
 631    memcpy(initial, out, sizeof(out));
 632    initial_num_out = num_out;
 633    initial_num_out_sgpr = num_out_sgpr;
 634
 635    /* Now chain the parts. */
 636    LLVMValueRef ret = NULL;
 637    for (unsigned part = 0; part < num_parts; ++part) {
 638       LLVMValueRef in[AC_MAX_ARGS];
 639       LLVMTypeRef ret_type;
 640       unsigned out_idx = 0;
 641       unsigned num_params = LLVMCountParams(parts[part]);
 642
 643       /* Merged shaders are executed conditionally depending
 644        * on the number of enabled threads passed in the input SGPRs. */
 645       if (si_is_multi_part_shader(ctx->shader) && part == 0) {
 646          LLVMValueRef ena, count = initial[3];
 647
 648          count = LLVMBuildAnd(builder, count, LLVMConstInt(ctx->ac.i32, 0x7f, 0), "");
 649          ena = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), count, "");
 650          ac_build_ifcc(&ctx->ac, ena, 6506);
 651       }
 652
 653       /* Derive arguments for the next part from outputs of the
 654        * previous one.
 655        */
 656       for (unsigned param_idx = 0; param_idx < num_params; ++param_idx) {
 657          LLVMValueRef param;
 658          LLVMTypeRef param_type;
 659          bool is_sgpr;
 660          unsigned param_size;
 661          LLVMValueRef arg = NULL;
 662
 663          param = LLVMGetParam(parts[part], param_idx);
 664          param_type = LLVMTypeOf(param);
 665          param_size = ac_get_type_size(param_type) / 4;
 666          is_sgpr = ac_is_sgpr_param(param);
 667
 668          if (is_sgpr) {
 669             ac_add_function_attr(ctx->ac.context, parts[part], param_idx + 1, AC_FUNC_ATTR_INREG);
 670          } else if (out_idx < num_out_sgpr) {
 671             /* Skip returned SGPRs the current part doesn't
 672              * declare on the input. */
 673             out_idx = num_out_sgpr;
 674          }
 675
 676          assert(out_idx + param_size <= (is_sgpr ? num_out_sgpr : num_out));
 677
 678          if (param_size == 1)
 679             arg = out[out_idx];
 680          else
 681             arg = ac_build_gather_values(&ctx->ac, &out[out_idx], param_size);
 682
 683          if (LLVMTypeOf(arg) != param_type) {
 684             if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
 685                if (LLVMGetPointerAddressSpace(param_type) == AC_ADDR_SPACE_CONST_32BIT) {
 686                   arg = LLVMBuildBitCast(builder, arg, ctx->ac.i32, "");
 687                   arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
 688                } else {
 689                   arg = LLVMBuildBitCast(builder, arg, ctx->ac.i64, "");
 690                   arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
 691                }
 692             } else {
 693                arg = LLVMBuildBitCast(builder, arg, param_type, "");
 694             }
 695          }
 696
 697          in[param_idx] = arg;
 698          out_idx += param_size;
 699       }
 700
 701       ret = ac_build_call(&ctx->ac, parts[part], in, num_params);
 702
 703       if (si_is_multi_part_shader(ctx->shader) && part + 1 == next_shader_first_part) {
 704          ac_build_endif(&ctx->ac, 6506);
 705
 706          /* The second half of the merged shader should use
 707           * the inputs from the toplevel (wrapper) function,
 708           * not the return value from the last call.
 709           *
 710           * That's because the last call was executed condi-
 711           * tionally, so we can't consume it in the main
 712           * block.
 713           */
 714          memcpy(out, initial, sizeof(initial));
 715          num_out = initial_num_out;
 716          num_out_sgpr = initial_num_out_sgpr;
 717          continue;
 718       }
 719
 720       /* Extract the returned GPRs. */
 721       ret_type = LLVMTypeOf(ret);
 722       num_out = 0;
 723       num_out_sgpr = 0;
 724
 725       if (LLVMGetTypeKind(ret_type) != LLVMVoidTypeKind) {
 726          assert(LLVMGetTypeKind(ret_type) == LLVMStructTypeKind);
 727
 728          unsigned ret_size = LLVMCountStructElementTypes(ret_type);
 729
 730          for (unsigned i = 0; i < ret_size; ++i) {
 731             LLVMValueRef val = LLVMBuildExtractValue(builder, ret, i, "");
 732
 733             assert(num_out < ARRAY_SIZE(out));
 734             out[num_out++] = val;
 735
 736             if (LLVMTypeOf(val) == ctx->ac.i32) {
 737                assert(num_out_sgpr + 1 == num_out);
 738                num_out_sgpr = num_out;
 739             }
 740          }
 741       }
 742    }
 743
 744    /* Return the value from the last part. */
 745    if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
 746       LLVMBuildRetVoid(builder);
 747    else
 748       LLVMBuildRet(builder, ret);
 749 }