X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fvc4%2Fvc4_qir.c;h=0d104c0615057e8989649755cc8c646812f86c6d;hb=d233c8c914ce819147197b9327bc22d1ea58b2fb;hp=7894b081b1921d51015a19022d5ddcb522d0713a;hpb=01ca4f207efac555ff5f729dce1687a68ba65400;p=mesa.git diff --git a/src/gallium/drivers/vc4/vc4_qir.c b/src/gallium/drivers/vc4/vc4_qir.c index 7894b081b19..0d104c06150 100644 --- a/src/gallium/drivers/vc4/vc4_qir.c +++ b/src/gallium/drivers/vc4/vc4_qir.c @@ -31,7 +31,6 @@ struct qir_op_info { const char *name; uint8_t ndst, nsrc; bool has_side_effects; - bool multi_instruction; }; static const struct qir_op_info qir_op_info[] = { @@ -59,44 +58,36 @@ static const struct qir_op_info qir_op_info[] = { [QOP_ASR] = { "asr", 1, 2 }, [QOP_SHL] = { "shl", 1, 2 }, [QOP_MIN] = { "min", 1, 2 }, + [QOP_MIN_NOIMM] = { "min_noimm", 1, 2 }, [QOP_MAX] = { "max", 1, 2 }, [QOP_AND] = { "and", 1, 2 }, [QOP_OR] = { "or", 1, 2 }, [QOP_XOR] = { "xor", 1, 2 }, [QOP_NOT] = { "not", 1, 1 }, - [QOP_SEL_X_0_NS] = { "fsel_x_0_ns", 1, 1, false, true }, - [QOP_SEL_X_0_NC] = { "fsel_x_0_nc", 1, 1, false, true }, - [QOP_SEL_X_0_ZS] = { "fsel_x_0_zs", 1, 1, false, true }, - [QOP_SEL_X_0_ZC] = { "fsel_x_0_zc", 1, 1, false, true }, - [QOP_SEL_X_Y_NS] = { "fsel_x_y_ns", 1, 2, false, true }, - [QOP_SEL_X_Y_NC] = { "fsel_x_y_nc", 1, 2, false, true }, - [QOP_SEL_X_Y_ZS] = { "fsel_x_y_zs", 1, 2, false, true }, - [QOP_SEL_X_Y_ZC] = { "fsel_x_y_zc", 1, 2, false, true }, - - [QOP_RCP] = { "rcp", 1, 1, false, true }, - [QOP_RSQ] = { "rsq", 1, 1, false, true }, - [QOP_EXP2] = { "exp2", 1, 2, false, true }, - [QOP_LOG2] = { "log2", 1, 2, false, true }, - [QOP_TLB_DISCARD_SETUP] = { "discard", 0, 1, true }, - [QOP_TLB_STENCIL_SETUP] = { "tlb_stencil_setup", 0, 1, true }, - [QOP_TLB_Z_WRITE] = { "tlb_z", 0, 1, true }, - [QOP_TLB_COLOR_WRITE] = { "tlb_color", 0, 1, true }, + [QOP_RCP] = { "rcp", 1, 1 }, + [QOP_RSQ] = { "rsq", 1, 1 }, + [QOP_EXP2] = { "exp2", 1, 1 }, + [QOP_LOG2] = { "log2", 1, 1 }, [QOP_TLB_COLOR_READ] = { "tlb_color_read", 1, 0 }, + [QOP_MS_MASK] = { "ms_mask", 0, 1, true }, [QOP_VARY_ADD_C] = { "vary_add_c", 1, 1 }, - [QOP_FRAG_X] = { "frag_x", 1, 0 }, - [QOP_FRAG_Y] = { "frag_y", 1, 0 }, [QOP_FRAG_Z] = { "frag_z", 1, 0 }, [QOP_FRAG_W] = { "frag_w", 1, 0 }, - [QOP_FRAG_REV_FLAG] = { "frag_rev_flag", 1, 0 }, - [QOP_TEX_S] = { "tex_s", 0, 2 }, - [QOP_TEX_T] = { "tex_t", 0, 2 }, - [QOP_TEX_R] = { "tex_r", 0, 2 }, - [QOP_TEX_B] = { "tex_b", 0, 2 }, - [QOP_TEX_DIRECT] = { "tex_direct", 0, 2 }, [QOP_TEX_RESULT] = { "tex_result", 1, 0, true }, + + [QOP_THRSW] = { "thrsw", 0, 0, true }, + + [QOP_LOAD_IMM] = { "load_imm", 0, 1 }, + [QOP_LOAD_IMM_U2] = { "load_imm_u2", 0, 1 }, + [QOP_LOAD_IMM_I2] = { "load_imm_i2", 0, 1 }, + + [QOP_ROT_MUL] = { "rot_mul", 0, 2 }, + + [QOP_BRANCH] = { "branch", 0, 0, true }, + [QOP_UNIFORMS_RESET] = { "uniforms_reset", 0, 2, true }, }; static const char * @@ -109,12 +100,35 @@ qir_get_op_name(enum qop qop) } int -qir_get_op_nsrc(enum qop qop) +qir_get_non_sideband_nsrc(struct qinst *inst) { - if (qop < ARRAY_SIZE(qir_op_info) && qir_op_info[qop].name) - return qir_op_info[qop].nsrc; - else - abort(); + assert(qir_op_info[inst->op].name); + return qir_op_info[inst->op].nsrc; +} + +int +qir_get_nsrc(struct qinst *inst) +{ + assert(qir_op_info[inst->op].name); + + int nsrc = qir_get_non_sideband_nsrc(inst); + + /* Normal (non-direct) texture coordinate writes also implicitly load + * a uniform for the texture parameters. + */ + if (qir_is_tex(inst) && inst->dst.file != QFILE_TEX_S_DIRECT) + nsrc++; + + return nsrc; +} + +/* The sideband uniform for textures gets stored after the normal ALU + * arguments. + */ +int +qir_get_tex_uniform_src(struct qinst *inst) +{ + return qir_get_nsrc(inst) - 1; } /** @@ -124,6 +138,21 @@ qir_get_op_nsrc(enum qop qop) bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst) { + switch (inst->dst.file) { + case QFILE_TLB_Z_WRITE: + case QFILE_TLB_COLOR_WRITE: + case QFILE_TLB_COLOR_WRITE_MS: + case QFILE_TLB_STENCIL_SETUP: + case QFILE_TEX_S_DIRECT: + case QFILE_TEX_S: + case QFILE_TEX_T: + case QFILE_TEX_R: + case QFILE_TEX_B: + return true; + default: + break; + } + return qir_op_info[inst->op].has_side_effects; } @@ -135,7 +164,7 @@ qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst) * point/line coordinates reads, because they're generated by * fixed-function hardware. */ - for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) { + for (int i = 0; i < qir_get_nsrc(inst); i++) { if (inst->src[i].file == QFILE_VARY && c->input_slots[inst->src[i].index].slot == 0xff) { return true; @@ -152,9 +181,14 @@ qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst) } bool -qir_is_multi_instruction(struct qinst *inst) +qir_has_uniform_read(struct qinst *inst) { - return qir_op_info[inst->op].multi_instruction; + for (int i = 0; i < qir_get_nsrc(inst); i++) { + if (inst->src[i].file == QFILE_UNIF) + return true; + } + + return false; } bool @@ -169,6 +203,7 @@ qir_is_mul(struct qinst *inst) case QOP_V8MAX: case QOP_V8ADDS: case QOP_V8SUBS: + case QOP_ROT_MUL: return true; default: return false; @@ -200,6 +235,7 @@ qir_is_raw_mov(struct qinst *inst) return ((inst->op == QOP_MOV || inst->op == QOP_FMOV || inst->op == QOP_MMOV) && + inst->cond == QPU_COND_ALWAYS && !inst->dst.pack && !inst->src[0].pack); } @@ -207,27 +243,43 @@ qir_is_raw_mov(struct qinst *inst) bool qir_is_tex(struct qinst *inst) { - return inst->op >= QOP_TEX_S && inst->op <= QOP_TEX_DIRECT; + switch (inst->dst.file) { + case QFILE_TEX_S_DIRECT: + case QFILE_TEX_S: + case QFILE_TEX_T: + case QFILE_TEX_R: + case QFILE_TEX_B: + return true; + default: + return false; + } } bool -qir_depends_on_flags(struct qinst *inst) +qir_has_implicit_tex_uniform(struct qinst *inst) { - switch (inst->op) { - case QOP_SEL_X_0_NS: - case QOP_SEL_X_0_NC: - case QOP_SEL_X_0_ZS: - case QOP_SEL_X_0_ZC: - case QOP_SEL_X_Y_NS: - case QOP_SEL_X_Y_NC: - case QOP_SEL_X_Y_ZS: - case QOP_SEL_X_Y_ZC: + switch (inst->dst.file) { + case QFILE_TEX_S: + case QFILE_TEX_T: + case QFILE_TEX_R: + case QFILE_TEX_B: return true; default: return false; } } +bool +qir_depends_on_flags(struct qinst *inst) +{ + if (inst->op == QOP_BRANCH) { + return inst->cond != QPU_COND_BRANCH_ALWAYS; + } else { + return (inst->cond != QPU_COND_ALWAYS && + inst->cond != QPU_COND_NEVER); + } +} + bool qir_writes_r4(struct qinst *inst) { @@ -244,58 +296,197 @@ qir_writes_r4(struct qinst *inst) } } +uint8_t +qir_channels_written(struct qinst *inst) +{ + if (qir_is_mul(inst)) { + switch (inst->dst.pack) { + case QPU_PACK_MUL_NOP: + case QPU_PACK_MUL_8888: + return 0xf; + case QPU_PACK_MUL_8A: + return 0x1; + case QPU_PACK_MUL_8B: + return 0x2; + case QPU_PACK_MUL_8C: + return 0x4; + case QPU_PACK_MUL_8D: + return 0x8; + } + } else { + switch (inst->dst.pack) { + case QPU_PACK_A_NOP: + case QPU_PACK_A_8888: + case QPU_PACK_A_8888_SAT: + case QPU_PACK_A_32_SAT: + return 0xf; + case QPU_PACK_A_8A: + case QPU_PACK_A_8A_SAT: + return 0x1; + case QPU_PACK_A_8B: + case QPU_PACK_A_8B_SAT: + return 0x2; + case QPU_PACK_A_8C: + case QPU_PACK_A_8C_SAT: + return 0x4; + case QPU_PACK_A_8D: + case QPU_PACK_A_8D_SAT: + return 0x8; + case QPU_PACK_A_16A: + case QPU_PACK_A_16A_SAT: + return 0x3; + case QPU_PACK_A_16B: + case QPU_PACK_A_16B_SAT: + return 0xc; + } + } + unreachable("Bad pack field"); +} + +char * +qir_describe_uniform(enum quniform_contents contents, uint32_t data, + const uint32_t *uniforms) +{ + static const char *quniform_names[] = { + [QUNIFORM_VIEWPORT_X_SCALE] = "vp_x_scale", + [QUNIFORM_VIEWPORT_Y_SCALE] = "vp_y_scale", + [QUNIFORM_VIEWPORT_Z_OFFSET] = "vp_z_offset", + [QUNIFORM_VIEWPORT_Z_SCALE] = "vp_z_scale", + [QUNIFORM_TEXTURE_CONFIG_P0] = "tex_p0", + [QUNIFORM_TEXTURE_CONFIG_P1] = "tex_p1", + [QUNIFORM_TEXTURE_CONFIG_P2] = "tex_p2", + [QUNIFORM_TEXTURE_FIRST_LEVEL] = "tex_first_level", + }; + + switch (contents) { + case QUNIFORM_CONSTANT: + return ralloc_asprintf(NULL, "0x%08x / %f", data, uif(data)); + case QUNIFORM_UNIFORM: + if (uniforms) { + uint32_t unif = uniforms[data]; + return ralloc_asprintf(NULL, "unif[%d] = 0x%08x / %f", + data, unif, uif(unif)); + } else { + return ralloc_asprintf(NULL, "unif[%d]", data); + } + + case QUNIFORM_TEXTURE_CONFIG_P0: + case QUNIFORM_TEXTURE_CONFIG_P1: + case QUNIFORM_TEXTURE_CONFIG_P2: + case QUNIFORM_TEXTURE_FIRST_LEVEL: + return ralloc_asprintf(NULL, "%s[%d]", + quniform_names[contents], data); + + default: + if (contents < ARRAY_SIZE(quniform_names) && + quniform_names[contents]) { + return ralloc_asprintf(NULL, "%s", + quniform_names[contents]); + } else { + return ralloc_asprintf(NULL, "??? %d", contents); + } + } +} + static void qir_print_reg(struct vc4_compile *c, struct qreg reg, bool write) { static const char *files[] = { [QFILE_TEMP] = "t", [QFILE_VARY] = "v", - [QFILE_UNIF] = "u", + [QFILE_TLB_COLOR_WRITE] = "tlb_c", + [QFILE_TLB_COLOR_WRITE_MS] = "tlb_c_ms", + [QFILE_TLB_Z_WRITE] = "tlb_z", + [QFILE_TLB_STENCIL_SETUP] = "tlb_stencil", + [QFILE_FRAG_X] = "frag_x", + [QFILE_FRAG_Y] = "frag_y", + [QFILE_FRAG_REV_FLAG] = "frag_rev_flag", + [QFILE_QPU_ELEMENT] = "elem", + [QFILE_TEX_S_DIRECT] = "tex_s_direct", + [QFILE_TEX_S] = "tex_s", + [QFILE_TEX_T] = "tex_t", + [QFILE_TEX_R] = "tex_r", + [QFILE_TEX_B] = "tex_b", }; - if (reg.file == QFILE_NULL) { + switch (reg.file) { + + case QFILE_NULL: fprintf(stderr, "null"); - } else if (reg.file == QFILE_SMALL_IMM) { + break; + + case QFILE_LOAD_IMM: + fprintf(stderr, "0x%08x (%f)", reg.index, uif(reg.index)); + break; + + case QFILE_SMALL_IMM: if ((int)reg.index >= -16 && (int)reg.index <= 15) fprintf(stderr, "%d", reg.index); else fprintf(stderr, "%f", uif(reg.index)); - } else if (reg.file == QFILE_VPM) { + break; + + case QFILE_VPM: if (write) { fprintf(stderr, "vpm"); } else { fprintf(stderr, "vpm%d.%d", reg.index / 4, reg.index % 4); } - } else { - fprintf(stderr, "%s%d", files[reg.file], reg.index); + break; + + case QFILE_TLB_COLOR_WRITE: + case QFILE_TLB_COLOR_WRITE_MS: + case QFILE_TLB_Z_WRITE: + case QFILE_TLB_STENCIL_SETUP: + case QFILE_TEX_S_DIRECT: + case QFILE_TEX_S: + case QFILE_TEX_T: + case QFILE_TEX_R: + case QFILE_TEX_B: + fprintf(stderr, "%s", files[reg.file]); + break; + + case QFILE_UNIF: { + char *desc = qir_describe_uniform(c->uniform_contents[reg.index], + c->uniform_data[reg.index], + NULL); + fprintf(stderr, "u%d (%s)", reg.index, desc); + ralloc_free(desc); + break; } - if (reg.file == QFILE_UNIF && - c->uniform_contents[reg.index] == QUNIFORM_CONSTANT) { - fprintf(stderr, " (0x%08x / %f)", - c->uniform_data[reg.index], - uif(c->uniform_data[reg.index])); + default: + fprintf(stderr, "%s%d", files[reg.file], reg.index); + break; } } void qir_dump_inst(struct vc4_compile *c, struct qinst *inst) { - fprintf(stderr, "%s%s ", - qir_get_op_name(inst->op), - inst->sf ? ".sf" : ""); + fprintf(stderr, "%s", qir_get_op_name(inst->op)); + if (inst->op == QOP_BRANCH) + vc4_qpu_disasm_cond_branch(stderr, inst->cond); + else + vc4_qpu_disasm_cond(stderr, inst->cond); + if (inst->sf) + fprintf(stderr, ".sf"); + fprintf(stderr, " "); - qir_print_reg(c, inst->dst, true); - if (inst->dst.pack) { + if (inst->op != QOP_BRANCH) { + qir_print_reg(c, inst->dst, true); if (inst->dst.pack) { - if (qir_is_mul(inst)) - vc4_qpu_disasm_pack_mul(stderr, inst->dst.pack); - else - vc4_qpu_disasm_pack_a(stderr, inst->dst.pack); + if (inst->dst.pack) { + if (qir_is_mul(inst)) + vc4_qpu_disasm_pack_mul(stderr, inst->dst.pack); + else + vc4_qpu_disasm_pack_a(stderr, inst->dst.pack); + } } } - for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) { + + for (int i = 0; i < qir_get_nsrc(inst); i++) { fprintf(stderr, ", "); qir_print_reg(c, inst->src[i], false); vc4_qpu_disasm_unpack(stderr, inst->src[i].pack); @@ -305,9 +496,70 @@ qir_dump_inst(struct vc4_compile *c, struct qinst *inst) void qir_dump(struct vc4_compile *c) { - list_for_each_entry(struct qinst, inst, &c->instructions, link) { - qir_dump_inst(c, inst); - fprintf(stderr, "\n"); + int ip = 0; + int pressure = 0; + + qir_for_each_block(block, c) { + fprintf(stderr, "BLOCK %d:\n", block->index); + qir_for_each_inst(inst, block) { + if (c->temp_start) { + bool first = true; + + fprintf(stderr, "%3d ", pressure); + + for (int i = 0; i < c->num_temps; i++) { + if (c->temp_start[i] != ip) + continue; + + if (first) { + first = false; + } else { + fprintf(stderr, ", "); + } + fprintf(stderr, "S%4d", i); + pressure++; + } + + if (first) + fprintf(stderr, " "); + else + fprintf(stderr, " "); + } + + if (c->temp_end) { + bool first = true; + + for (int i = 0; i < c->num_temps; i++) { + if (c->temp_end[i] != ip) + continue; + + if (first) { + first = false; + } else { + fprintf(stderr, ", "); + } + fprintf(stderr, "E%4d", i); + pressure--; + } + + if (first) + fprintf(stderr, " "); + else + fprintf(stderr, " "); + } + + qir_dump_inst(c, inst); + fprintf(stderr, "\n"); + ip++; + } + if (block->successors[1]) { + fprintf(stderr, "-> BLOCK %d, %d\n", + block->successors[0]->index, + block->successors[1]->index); + } else if (block->successors[0]) { + fprintf(stderr, "-> BLOCK %d\n", + block->successors[0]->index); + } } } @@ -339,46 +591,98 @@ qir_inst(enum qop op, struct qreg dst, struct qreg src0, struct qreg src1) inst->op = op; inst->dst = dst; - inst->src = calloc(2, sizeof(inst->src[0])); inst->src[0] = src0; inst->src[1] = src1; + inst->cond = QPU_COND_ALWAYS; return inst; } -struct qinst * -qir_inst4(enum qop op, struct qreg dst, - struct qreg a, - struct qreg b, - struct qreg c, - struct qreg d) +static void +qir_emit(struct vc4_compile *c, struct qinst *inst) { - struct qinst *inst = CALLOC_STRUCT(qinst); + list_addtail(&inst->link, &c->cur_block->instructions); +} - inst->op = op; - inst->dst = dst; - inst->src = calloc(4, sizeof(*inst->src)); - inst->src[0] = a; - inst->src[1] = b; - inst->src[2] = c; - inst->src[3] = d; +/* Updates inst to write to a new temporary, emits it, and notes the def. */ +struct qreg +qir_emit_def(struct vc4_compile *c, struct qinst *inst) +{ + assert(inst->dst.file == QFILE_NULL); - return inst; + inst->dst = qir_get_temp(c); + + if (inst->dst.file == QFILE_TEMP) + c->defs[inst->dst.index] = inst; + + qir_emit(c, inst); + + return inst->dst; } -void -qir_emit(struct vc4_compile *c, struct qinst *inst) +struct qinst * +qir_emit_nondef(struct vc4_compile *c, struct qinst *inst) { if (inst->dst.file == QFILE_TEMP) - c->defs[inst->dst.index] = inst; + c->defs[inst->dst.index] = NULL; - qir_emit_nodef(c, inst); + qir_emit(c, inst); + + return inst; } bool qir_reg_equals(struct qreg a, struct qreg b) { - return a.file == b.file && a.index == b.index; + return a.file == b.file && a.index == b.index && a.pack == b.pack; +} + +struct qblock * +qir_new_block(struct vc4_compile *c) +{ + struct qblock *block = rzalloc(c, struct qblock); + + list_inithead(&block->instructions); + list_inithead(&block->qpu_inst_list); + + block->predecessors = _mesa_set_create(block, + _mesa_hash_pointer, + _mesa_key_pointer_equal); + + block->index = c->next_block_index++; + + return block; +} + +void +qir_set_emit_block(struct vc4_compile *c, struct qblock *block) +{ + c->cur_block = block; + list_addtail(&block->link, &c->blocks); +} + +struct qblock * +qir_entry_block(struct vc4_compile *c) +{ + return list_first_entry(&c->blocks, struct qblock, link); +} + +struct qblock * +qir_exit_block(struct vc4_compile *c) +{ + return list_last_entry(&c->blocks, struct qblock, link); +} + +void +qir_link_blocks(struct qblock *predecessor, struct qblock *successor) +{ + _mesa_set_add(successor->predecessors, predecessor); + if (predecessor->successors[0]) { + assert(!predecessor->successors[1]); + predecessor->successors[1] = successor; + } else { + predecessor->successors[0] = successor; + } } struct vc4_compile * @@ -386,11 +690,14 @@ qir_compile_init(void) { struct vc4_compile *c = rzalloc(NULL, struct vc4_compile); - list_inithead(&c->instructions); + list_inithead(&c->blocks); + qir_set_emit_block(c, qir_new_block(c)); + c->last_top_block = c->cur_block; c->output_position_index = -1; c->output_color_index = -1; c->output_point_size_index = -1; + c->output_sample_mask_index = -1; c->def_ht = _mesa_hash_table_create(c, _mesa_hash_pointer, _mesa_key_pointer_equal); @@ -405,30 +712,38 @@ qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst) c->defs[qinst->dst.index] = NULL; list_del(&qinst->link); - free(qinst->src); free(qinst); } struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg) { + int pack = reg.pack; + while (reg.file == QFILE_TEMP && c->defs[reg.index] && - c->defs[reg.index]->op == QOP_MOV && - !c->defs[reg.index]->dst.pack) { + (c->defs[reg.index]->op == QOP_MOV || + c->defs[reg.index]->op == QOP_FMOV || + c->defs[reg.index]->op == QOP_MMOV)&& + !c->defs[reg.index]->dst.pack && + !c->defs[reg.index]->src[0].pack) { reg = c->defs[reg.index]->src[0]; } + reg.pack = pack; return reg; } void qir_compile_destroy(struct vc4_compile *c) { - while (!list_empty(&c->instructions)) { - struct qinst *qinst = - (struct qinst *)c->instructions.next; - qir_remove_instruction(c, qinst); + qir_for_each_block(block, c) { + while (!list_is_empty(&block->instructions)) { + struct qinst *qinst = + list_first_entry(&block->instructions, + struct qinst, link); + qir_remove_instruction(c, qinst); + } } ralloc_free(c); @@ -454,12 +769,11 @@ qir_uniform(struct vc4_compile *c, for (int i = 0; i < c->num_uniforms; i++) { if (c->uniform_contents[i] == contents && c->uniform_data[i] == data) { - return (struct qreg) { QFILE_UNIF, i }; + return qir_reg(QFILE_UNIF, i); } } uint32_t uniform = c->num_uniforms++; - struct qreg u = { QFILE_UNIF, uniform }; if (uniform >= c->uniform_array_size) { c->uniform_array_size = MAX2(MAX2(16, uniform + 1), @@ -476,22 +790,25 @@ qir_uniform(struct vc4_compile *c, c->uniform_contents[uniform] = contents; c->uniform_data[uniform] = data; - return u; + return qir_reg(QFILE_UNIF, uniform); } void qir_SF(struct vc4_compile *c, struct qreg src) { struct qinst *last_inst = NULL; - if (!list_empty(&c->instructions)) - last_inst = (struct qinst *)c->instructions.prev; - - if (!last_inst || - last_inst->dst.file != src.file || - last_inst->dst.index != src.index || - qir_is_multi_instruction(last_inst)) { - src = qir_MOV(c, src); - last_inst = (struct qinst *)c->instructions.prev; + + if (!list_is_empty(&c->cur_block->instructions)) + last_inst = (struct qinst *)c->cur_block->instructions.prev; + + /* We don't have any way to guess which kind of MOV is implied. */ + assert(!src.pack); + + if (src.file != QFILE_TEMP || + !c->defs[src.index] || + last_inst != c->defs[src.index]) { + last_inst = qir_MOV_dest(c, qir_reg(QFILE_NULL, 0), src); + last_inst = (struct qinst *)c->cur_block->instructions.prev; } last_inst->sf = true; } @@ -506,6 +823,7 @@ qir_SF(struct vc4_compile *c, struct qreg src) "QIR opt pass %2d: %s progress\n", \ pass, #func); \ } \ + qir_validate(c); \ } \ } while (0) @@ -519,12 +837,13 @@ qir_optimize(struct vc4_compile *c) bool progress = false; OPTPASS(qir_opt_algebraic); - OPTPASS(qir_opt_cse); OPTPASS(qir_opt_constant_folding); OPTPASS(qir_opt_copy_propagation); + OPTPASS(qir_opt_peephole_sf); OPTPASS(qir_opt_dead_code); OPTPASS(qir_opt_small_immediates); - OPTPASS(qir_opt_vpm_writes); + OPTPASS(qir_opt_vpm); + OPTPASS(qir_opt_coalesce_ff_writes); if (!progress) break;