wine/vkd3d
0
0
Fork 0
mirror of https://gitlab.winehq.org/wine/vkd3d.git synced 2025-12-15 08:03:30 -08:00
Code Issues Packages Projects Releases Wiki Activity

vkd3d-shader/ir: Introduce a vsir DCE pass.

Browse Source
This commit is contained in:
Elizabeth Figura
2025-03-05 14:23:40 -06:00
committed by Henri Verbeet
parent c7529028eb
commit 322c91e3f8
Notes: Henri Verbeet 2025-10-06 19:48:04 +02:00 
Approved-by: Francisco Casas (@fcasas)
Approved-by: Henri Verbeet (@hverbeet)
Merge-Request: https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/1718
5 changed files with 449 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

431
libs/vkd3d-shader/ir.c
View File

@@ -28,6 +28,7 @@ struct vsir_transformation_context
const struct vkd3d_shader_compile_info *compile_info;
struct vkd3d_shader_message_context *message_context;
struct vkd3d_shader_location null_location;
bool progress;
};
static void vsir_transformation_context_init(struct vsir_transformation_context *ctx,
@@ -8850,7 +8851,7 @@ struct liveness_tracker
bool written;
bool fixed_mask;
uint8_t mask;
unsigned int first_write, last_access;
unsigned int first_write, last_access, last_read;
} *ssa_regs, *temp_regs;
};
@@ -8864,9 +8865,15 @@ static void liveness_track_src(struct liveness_tracker *tracker,
}
if (src->reg.type == VKD3DSPR_SSA)
{
tracker->ssa_regs[src->reg.idx[0].offset].last_read = index;
tracker->ssa_regs[src->reg.idx[0].offset].last_access = index;
}
else if (src->reg.type == VKD3DSPR_TEMP)
{
tracker->temp_regs[src->reg.idx[0].offset].last_read = index;
tracker->temp_regs[src->reg.idx[0].offset].last_access = index;
}
}
static void liveness_track_dst(struct liveness_tracker *tracker, struct vkd3d_shader_dst_param *dst,
@@ -9020,6 +9027,8 @@ static enum vkd3d_result track_liveness(struct vsir_program *program, struct liv
reg->first_write = loop_start;
if (reg->last_access < i)
reg->last_access = i;
if (reg->last_read < i)
reg->last_read = i;
}
for (unsigned int j = 0; j < program->temp_count; ++j)
@@ -9030,6 +9039,8 @@ static enum vkd3d_result track_liveness(struct vsir_program *program, struct liv
reg->first_write = loop_start;
if (reg->last_access < i)
reg->last_access = i;
if (reg->last_read < i)
reg->last_read = i;
}
}
}
@@ -12888,6 +12899,8 @@ static void vsir_transform_(
struct vsir_transformation_context *ctx, const char *step_name,
enum vkd3d_result (*step)(struct vsir_program *program, struct vsir_transformation_context *ctx))
{
ctx->progress = false;
if (ctx->result < 0)
return;
@@ -12905,6 +12918,422 @@ static void vsir_transform_(
}
}
static bool vsir_instruction_has_side_effects(const struct vkd3d_shader_instruction *ins)
{
switch (ins->opcode)
{
case VSIR_OP_ABS:
case VSIR_OP_ACOS:
case VSIR_OP_ADD:
case VSIR_OP_AND:
case VSIR_OP_ASIN:
case VSIR_OP_ATAN:
case VSIR_OP_BEM:
case VSIR_OP_BFI:
case VSIR_OP_BFREV:
case VSIR_OP_BUFINFO:
case VSIR_OP_CHECK_ACCESS_FULLY_MAPPED:
case VSIR_OP_CMP:
case VSIR_OP_CND:
case VSIR_OP_COS:
case VSIR_OP_COUNTBITS:
case VSIR_OP_CRS:
case VSIR_OP_DADD:
case VSIR_OP_DDIV:
case VSIR_OP_DEF:
case VSIR_OP_DEFB:
case VSIR_OP_DEFI:
case VSIR_OP_DEQO:
case VSIR_OP_DFMA:
case VSIR_OP_DGEO:
case VSIR_OP_DIV:
case VSIR_OP_DLT:
case VSIR_OP_DMAX:
case VSIR_OP_DMIN:
case VSIR_OP_DMOV:
case VSIR_OP_DMOVC:
case VSIR_OP_DMUL:
case VSIR_OP_DNE:
case VSIR_OP_DP2:
case VSIR_OP_DP2ADD:
case VSIR_OP_DP3:
case VSIR_OP_DP4:
case VSIR_OP_DRCP:
case VSIR_OP_DST:
case VSIR_OP_DSX:
case VSIR_OP_DSX_COARSE:
case VSIR_OP_DSX_FINE:
case VSIR_OP_DSY:
case VSIR_OP_DSY_COARSE:
case VSIR_OP_DSY_FINE:
case VSIR_OP_DTOF:
case VSIR_OP_DTOI:
case VSIR_OP_DTOU:
case VSIR_OP_EQO:
case VSIR_OP_EQU:
case VSIR_OP_EVAL_CENTROID:
case VSIR_OP_EVAL_SAMPLE_INDEX:
case VSIR_OP_EXP:
case VSIR_OP_EXPP:
case VSIR_OP_F16TOF32:
case VSIR_OP_F32TOF16:
case VSIR_OP_FIRSTBIT_HI:
case VSIR_OP_FIRSTBIT_LO:
case VSIR_OP_FIRSTBIT_SHI:
case VSIR_OP_FRC:
case VSIR_OP_FREM:
case VSIR_OP_FTOD:
case VSIR_OP_FTOI:
case VSIR_OP_FTOU:
case VSIR_OP_GATHER4:
case VSIR_OP_GATHER4_C:
case VSIR_OP_GATHER4_C_S:
case VSIR_OP_GATHER4_PO:
case VSIR_OP_GATHER4_PO_C:
case VSIR_OP_GATHER4_PO_C_S:
case VSIR_OP_GATHER4_PO_S:
case VSIR_OP_GATHER4_S:
case VSIR_OP_GEO:
case VSIR_OP_GEU:
case VSIR_OP_HCOS:
case VSIR_OP_HSIN:
case VSIR_OP_HTAN:
case VSIR_OP_IADD:
case VSIR_OP_IBFE:
case VSIR_OP_IDIV:
case VSIR_OP_IEQ:
case VSIR_OP_IGE:
case VSIR_OP_ILT:
case VSIR_OP_IMAD:
case VSIR_OP_IMAX:
case VSIR_OP_IMIN:
case VSIR_OP_IMUL:
case VSIR_OP_IMUL_LOW:
case VSIR_OP_INE:
case VSIR_OP_INEG:
case VSIR_OP_IREM:
case VSIR_OP_ISFINITE:
case VSIR_OP_ISHL:
case VSIR_OP_ISHR:
case VSIR_OP_ISINF:
case VSIR_OP_ISNAN:
case VSIR_OP_ITOD:
case VSIR_OP_ITOF:
case VSIR_OP_ITOI:
case VSIR_OP_LD:
case VSIR_OP_LD2DMS:
case VSIR_OP_LD2DMS_S:
case VSIR_OP_LD_RAW:
case VSIR_OP_LD_RAW_S:
case VSIR_OP_LD_S:
case VSIR_OP_LD_STRUCTURED:
case VSIR_OP_LD_STRUCTURED_S:
case VSIR_OP_LD_UAV_TYPED:
case VSIR_OP_LD_UAV_TYPED_S:
case VSIR_OP_LIT:
case VSIR_OP_LOD:
case VSIR_OP_LOG:
case VSIR_OP_LOGP:
case VSIR_OP_LRP:
case VSIR_OP_LTO:
case VSIR_OP_LTU:
case VSIR_OP_M3x2:
case VSIR_OP_M3x3:
case VSIR_OP_M3x4:
case VSIR_OP_M4x3:
case VSIR_OP_M4x4:
case VSIR_OP_MAD:
case VSIR_OP_MAX:
case VSIR_OP_MIN:
case VSIR_OP_MOV:
case VSIR_OP_MOVA:
case VSIR_OP_MOVC:
case VSIR_OP_MSAD:
case VSIR_OP_MUL:
case VSIR_OP_NEO:
case VSIR_OP_NEU:
case VSIR_OP_NOP:
case VSIR_OP_NOT:
case VSIR_OP_NRM:
case VSIR_OP_OR:
case VSIR_OP_ORD:
case VSIR_OP_PHI:
case VSIR_OP_POW:
case VSIR_OP_QUAD_READ_ACROSS_D:
case VSIR_OP_QUAD_READ_ACROSS_X:
case VSIR_OP_QUAD_READ_ACROSS_Y:
case VSIR_OP_QUAD_READ_LANE_AT:
case VSIR_OP_RCP:
case VSIR_OP_RESINFO:
case VSIR_OP_ROUND_NE:
case VSIR_OP_ROUND_NI:
case VSIR_OP_ROUND_PI:
case VSIR_OP_ROUND_Z:
case VSIR_OP_RSQ:
case VSIR_OP_SAMPLE:
case VSIR_OP_SAMPLE_B:
case VSIR_OP_SAMPLE_B_CL_S:
case VSIR_OP_SAMPLE_C:
case VSIR_OP_SAMPLE_C_CL_S:
case VSIR_OP_SAMPLE_C_LZ:
case VSIR_OP_SAMPLE_C_LZ_S:
case VSIR_OP_SAMPLE_CL_S:
case VSIR_OP_SAMPLE_GRAD:
case VSIR_OP_SAMPLE_GRAD_CL_S:
case VSIR_OP_SAMPLE_INFO:
case VSIR_OP_SAMPLE_LOD:
case VSIR_OP_SAMPLE_LOD_S:
case VSIR_OP_SAMPLE_POS:
case VSIR_OP_SETP:
case VSIR_OP_SGE:
case VSIR_OP_SGN:
case VSIR_OP_SIN:
case VSIR_OP_SINCOS:
case VSIR_OP_SLT:
case VSIR_OP_SQRT:
case VSIR_OP_SUB:
case VSIR_OP_SWAPC:
case VSIR_OP_TAN:
case VSIR_OP_TEX:
case VSIR_OP_TEXBEM:
case VSIR_OP_TEXBEML:
case VSIR_OP_TEXCOORD:
case VSIR_OP_TEXCRD:
case VSIR_OP_TEXDEPTH:
case VSIR_OP_TEXDP3:
case VSIR_OP_TEXDP3TEX:
case VSIR_OP_TEXLD:
case VSIR_OP_TEXLDD:
case VSIR_OP_TEXLDL:
case VSIR_OP_TEXM3x2DEPTH:
case VSIR_OP_TEXM3x2PAD:
case VSIR_OP_TEXM3x2TEX:
case VSIR_OP_TEXM3x3:
case VSIR_OP_TEXM3x3DIFF:
case VSIR_OP_TEXM3x3PAD:
case VSIR_OP_TEXM3x3SPEC:
case VSIR_OP_TEXM3x3TEX:
case VSIR_OP_TEXM3x3VSPEC:
case VSIR_OP_TEXREG2AR:
case VSIR_OP_TEXREG2GB:
case VSIR_OP_TEXREG2RGB:
case VSIR_OP_UBFE:
case VSIR_OP_UDIV:
case VSIR_OP_UDIV_SIMPLE:
case VSIR_OP_UGE:
case VSIR_OP_ULT:
case VSIR_OP_UMAX:
case VSIR_OP_UMIN:
case VSIR_OP_UMUL:
case VSIR_OP_UNO:
case VSIR_OP_UREM:
case VSIR_OP_USHR:
case VSIR_OP_UTOD:
case VSIR_OP_UTOF:
case VSIR_OP_UTOU:
case VSIR_OP_WAVE_ACTIVE_ALL_EQUAL:
case VSIR_OP_WAVE_ACTIVE_BALLOT:
case VSIR_OP_WAVE_ACTIVE_BIT_AND:
case VSIR_OP_WAVE_ACTIVE_BIT_OR:
case VSIR_OP_WAVE_ACTIVE_BIT_XOR:
case VSIR_OP_WAVE_ALL_BIT_COUNT:
case VSIR_OP_WAVE_ALL_TRUE:
case VSIR_OP_WAVE_ANY_TRUE:
case VSIR_OP_WAVE_IS_FIRST_LANE:
case VSIR_OP_WAVE_OP_ADD:
case VSIR_OP_WAVE_OP_IMAX:
case VSIR_OP_WAVE_OP_IMIN:
case VSIR_OP_WAVE_OP_MAX:
case VSIR_OP_WAVE_OP_MIN:
case VSIR_OP_WAVE_OP_MUL:
case VSIR_OP_WAVE_OP_UMAX:
case VSIR_OP_WAVE_OP_UMIN:
case VSIR_OP_WAVE_PREFIX_BIT_COUNT:
case VSIR_OP_WAVE_READ_LANE_AT:
case VSIR_OP_WAVE_READ_LANE_FIRST:
case VSIR_OP_XOR:
return false;
case VSIR_OP_ATOMIC_AND:
case VSIR_OP_ATOMIC_CMP_STORE:
case VSIR_OP_ATOMIC_IADD:
case VSIR_OP_ATOMIC_IMAX:
case VSIR_OP_ATOMIC_IMIN:
case VSIR_OP_ATOMIC_OR:
case VSIR_OP_ATOMIC_UMAX:
case VSIR_OP_ATOMIC_UMIN:
case VSIR_OP_ATOMIC_XOR:
case VSIR_OP_BRANCH:
case VSIR_OP_BREAK:
case VSIR_OP_BREAKC:
case VSIR_OP_BREAKP:
case VSIR_OP_CALL:
case VSIR_OP_CALLNZ:
case VSIR_OP_CASE:
case VSIR_OP_CONTINUE:
case VSIR_OP_CONTINUEP:
case VSIR_OP_CUT:
case VSIR_OP_CUT_STREAM:
case VSIR_OP_DCL:
case VSIR_OP_DCL_CONSTANT_BUFFER:
case VSIR_OP_DCL_FUNCTION_BODY:
case VSIR_OP_DCL_FUNCTION_TABLE:
case VSIR_OP_DCL_GLOBAL_FLAGS:
case VSIR_OP_DCL_GS_INSTANCES:
case VSIR_OP_DCL_HS_FORK_PHASE_INSTANCE_COUNT:
case VSIR_OP_DCL_HS_JOIN_PHASE_INSTANCE_COUNT:
case VSIR_OP_DCL_HS_MAX_TESSFACTOR:
case VSIR_OP_DCL_IMMEDIATE_CONSTANT_BUFFER:
case VSIR_OP_DCL_INDEX_RANGE:
case VSIR_OP_DCL_INDEXABLE_TEMP:
case VSIR_OP_DCL_INPUT:
case VSIR_OP_DCL_INPUT_CONTROL_POINT_COUNT:
case VSIR_OP_DCL_INPUT_PRIMITIVE:
case VSIR_OP_DCL_INPUT_PS:
case VSIR_OP_DCL_INPUT_PS_SGV:
case VSIR_OP_DCL_INPUT_PS_SIV:
case VSIR_OP_DCL_INPUT_SGV:
case VSIR_OP_DCL_INPUT_SIV:
case VSIR_OP_DCL_INTERFACE:
case VSIR_OP_DCL_OUTPUT:
case VSIR_OP_DCL_OUTPUT_CONTROL_POINT_COUNT:
case VSIR_OP_DCL_OUTPUT_SGV:
case VSIR_OP_DCL_OUTPUT_SIV:
case VSIR_OP_DCL_OUTPUT_TOPOLOGY:
case VSIR_OP_DCL_RESOURCE_RAW:
case VSIR_OP_DCL_RESOURCE_STRUCTURED:
case VSIR_OP_DCL_SAMPLER:
case VSIR_OP_DCL_STREAM:
case VSIR_OP_DCL_TEMPS:
case VSIR_OP_DCL_TESSELLATOR_DOMAIN:
case VSIR_OP_DCL_TESSELLATOR_OUTPUT_PRIMITIVE:
case VSIR_OP_DCL_TESSELLATOR_PARTITIONING:
case VSIR_OP_DCL_TGSM_RAW:
case VSIR_OP_DCL_TGSM_STRUCTURED:
case VSIR_OP_DCL_THREAD_GROUP:
case VSIR_OP_DCL_UAV_RAW:
case VSIR_OP_DCL_UAV_STRUCTURED:
case VSIR_OP_DCL_UAV_TYPED:
case VSIR_OP_DCL_VERTICES_OUT:
case VSIR_OP_DEFAULT:
case VSIR_OP_DISCARD:
case VSIR_OP_ELSE:
case VSIR_OP_EMIT:
case VSIR_OP_EMIT_STREAM:
case VSIR_OP_ENDIF:
case VSIR_OP_ENDLOOP:
case VSIR_OP_ENDREP:
case VSIR_OP_ENDSWITCH:
case VSIR_OP_FCALL:
case VSIR_OP_HS_CONTROL_POINT_PHASE:
case VSIR_OP_HS_DECLS:
case VSIR_OP_HS_FORK_PHASE:
case VSIR_OP_HS_JOIN_PHASE:
case VSIR_OP_IF:
case VSIR_OP_IFC:
case VSIR_OP_IMM_ATOMIC_ALLOC:
case VSIR_OP_IMM_ATOMIC_AND:
case VSIR_OP_IMM_ATOMIC_CMP_EXCH:
case VSIR_OP_IMM_ATOMIC_CONSUME:
case VSIR_OP_IMM_ATOMIC_EXCH:
case VSIR_OP_IMM_ATOMIC_IADD:
case VSIR_OP_IMM_ATOMIC_IMAX:
case VSIR_OP_IMM_ATOMIC_IMIN:
case VSIR_OP_IMM_ATOMIC_OR:
case VSIR_OP_IMM_ATOMIC_UMAX:
case VSIR_OP_IMM_ATOMIC_UMIN:
case VSIR_OP_IMM_ATOMIC_XOR:
case VSIR_OP_LABEL:
case VSIR_OP_LOOP:
case VSIR_OP_PHASE:
case VSIR_OP_REP:
case VSIR_OP_RET:
case VSIR_OP_RETP:
case VSIR_OP_STORE_RAW:
case VSIR_OP_STORE_STRUCTURED:
case VSIR_OP_STORE_UAV_TYPED:
case VSIR_OP_SWITCH:
case VSIR_OP_SWITCH_MONOLITHIC:
case VSIR_OP_SYNC:
case VSIR_OP_TEXKILL:
return true;
case VSIR_OP_INVALID:
case VSIR_OP_COUNT:
break;
}
vkd3d_unreachable();
}
static enum vkd3d_result vsir_program_dce(struct vsir_program *program,
struct vsir_transformation_context *ctx)
{
struct vsir_program_iterator it = vsir_program_iterator(&program->instructions);
struct vkd3d_shader_instruction *ins;
struct liveness_tracker tracker;
enum vkd3d_result ret;
unsigned int i;
if ((ret = track_liveness(program, &tracker)))
return ret;
for (ins = vsir_program_iterator_head(&it), i = 0; ins; ins = vsir_program_iterator_next(&it), ++i)
{
unsigned int used_dst_count = 0;
/* FIXME: IMM_ATOMIC_* can still be effectively DCE'd by turning them
* into non-immediate ATOMIC_* instructions. */
if (vsir_instruction_has_side_effects(ins))
continue;
for (unsigned int j = 0; j < ins->dst_count; ++j)
{
struct vkd3d_shader_dst_param *dst = &ins->dst[j];
if (dst->reg.type == VKD3DSPR_SSA && !tracker.ssa_regs[dst->reg.idx[0].offset].last_read)
{
vsir_dst_param_init_null(dst);
ctx->progress = true;
}
else if (dst->reg.type == VKD3DSPR_TEMP
&& tracker.temp_regs[dst->reg.idx[0].offset].last_read <= i
&& !(program->shader_version.major == 1 && dst->reg.idx[0].offset == 0))
{
vsir_dst_param_init_null(dst);
ctx->progress = true;
}
else if (dst->reg.type != VKD3DSPR_NULL)
{
++used_dst_count;
}
}
if (!used_dst_count)
vkd3d_shader_instruction_make_nop(ins);
}
liveness_tracker_cleanup(&tracker);
return VKD3D_OK;
}
enum vkd3d_result vsir_program_optimize(struct vsir_program *program, uint64_t config_flags,
const struct vkd3d_shader_compile_info *compile_info, struct vkd3d_shader_message_context *message_context)
{
struct vsir_transformation_context ctx;
vsir_transformation_context_init(&ctx, program, config_flags, compile_info, message_context);
do
vsir_transform(&ctx, vsir_program_dce);
while (ctx.progress);
if (TRACE_ON())
vsir_program_trace(program);
return ctx.result;
}
/* Transformations which should happen at parse time, i.e. before scan
* information is returned to the user.
*