From: Christian König Date: Thu, 20 Jan 2011 21:10:37 +0000 (+0100) Subject: Merge remote branch 'origin/master' into pipe-video X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=d2ff6b8715e817c1ef14d4bf12be58c19d894143;p=mesa.git Merge remote branch 'origin/master' into pipe-video Conflicts: src/gallium/drivers/r600/r600_asm.c src/gallium/drivers/r600/r600_shader.c --- d2ff6b8715e817c1ef14d4bf12be58c19d894143 diff --cc src/gallium/drivers/r600/eg_asm.c index 1881e633d54,67d742b3760..4f86e3b4c38 --- a/src/gallium/drivers/r600/eg_asm.c +++ b/src/gallium/drivers/r600/eg_asm.c @@@ -36,8 -35,9 +36,10 @@@ int eg_bc_cf_build(struct r600_bc *bc, switch (cf->inst) { case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3): + case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3): + case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3): case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3): + assert(!end_of_program); bc->bytecode[id++] = S_SQ_CF_ALU_WORD0_ADDR(cf->addr >> 1) | S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache[0].mode) | S_SQ_CF_ALU_WORD0_KCACHE_BANK0(cf->kcache[0].bank) | diff --cc src/gallium/drivers/r600/r600_asm.c index 31f386964a1,a473fb29c4d..61de24b31ae --- a/src/gallium/drivers/r600/r600_asm.c +++ b/src/gallium/drivers/r600/r600_asm.c @@@ -35,10 -35,7 +35,10 @@@ #define NUM_OF_CYCLES 3 #define NUM_OF_COMPONENTS 4 +#define PREV_ALU(alu) LIST_ENTRY(struct r600_bc_alu, alu->list.prev, list) +#define NEXT_ALU(alu) LIST_ENTRY(struct r600_bc_alu, alu->list.next, list) + - static inline unsigned int r600_bc_get_num_operands(struct r600_bc_alu *alu) + static inline unsigned int r600_bc_get_num_operands(struct r600_bc *bc, struct r600_bc_alu *alu) { if(alu->is_op3) return 3; @@@ -240,74 -254,121 +297,141 @@@ int r600_bc_add_output(struct r600_bc * return 0; } -/* alu instructions that can ony exits once per group */ -static int is_alu_once_inst(struct r600_bc *bc, struct r600_bc_alu *alu) +/* alu predicate instructions */ - static int is_alu_pred_inst(struct r600_bc_alu *alu) - { - return !alu->is_op3 && ( - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_UINT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_UINT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_INV || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_POP || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_CLR || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_RESTORE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_PUSH || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_PUSH || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_PUSH || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_PUSH || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_PUSH_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_PUSH_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_PUSH_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_PUSH_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETLT_PUSH_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETLE_PUSH_INT); ++static int is_alu_pred_inst(struct r600_bc *bc, struct r600_bc_alu *alu) + { + switch (bc->chiprev) { + case CHIPREV_R600: + case CHIPREV_R700: + return !alu->is_op3 && ( - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_UINT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_UINT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_INV || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_POP || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_CLR || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_RESTORE || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_PUSH || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_PUSH || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_PUSH || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_PUSH || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_PUSH_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_PUSH_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_PUSH_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_PUSH_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETLT_PUSH_INT || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETLE_PUSH_INT); + case CHIPREV_EVERGREEN: + default: + return !alu->is_op3 && ( - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT || - alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_UINT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_UINT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_INV || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_POP || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_CLR || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_RESTORE || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_PUSH || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_PUSH || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_PUSH || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_PUSH || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_PUSH_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_PUSH_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_PUSH_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_PUSH_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETLT_PUSH_INT || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETLE_PUSH_INT); + } } +/* alu kill instructions */ - static int is_alu_kill_inst(struct r600_bc_alu *alu) ++static int is_alu_kill_inst(struct r600_bc *bc, struct r600_bc_alu *alu) +{ - return !alu->is_op3 && ( - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT); ++ switch (bc->chiprev) { ++ case CHIPREV_R600: ++ case CHIPREV_R700: ++ return !alu->is_op3 && ( ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT || ++ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT); ++ case CHIPREV_EVERGREEN: ++ default: ++ return !alu->is_op3 && ( ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT || ++ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT); ++ } +} + +/* alu instructions that can ony exits once per group */ - static int is_alu_once_inst(struct r600_bc_alu *alu) ++static int is_alu_once_inst(struct r600_bc *bc, struct r600_bc_alu *alu) +{ - return is_alu_kill_inst(alu) || - is_alu_pred_inst(alu); ++ return is_alu_kill_inst(bc, alu) || ++ is_alu_pred_inst(bc, alu); +} + - static int is_alu_reduction_inst(struct r600_bc_alu *alu) + static int is_alu_reduction_inst(struct r600_bc *bc, struct r600_bc_alu *alu) { - return !alu->is_op3 && ( - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_CUBE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4 || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4_IEEE || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX4); + switch (bc->chiprev) { + case CHIPREV_R600: + case CHIPREV_R700: + return !alu->is_op3 && ( + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_CUBE || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4 || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4_IEEE || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX4); + case CHIPREV_EVERGREEN: + default: + return !alu->is_op3 && ( + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_CUBE || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4 || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4_IEEE || + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX4); + } } - static int is_alu_mova_inst(struct r600_bc_alu *alu) + static int is_alu_mova_inst(struct r600_bc *bc, struct r600_bc_alu *alu) { - return !alu->is_op3 && ( - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_FLOOR || - alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT); + switch (bc->chiprev) { + case CHIPREV_R600: + case CHIPREV_R700: + return !alu->is_op3 && ( + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_FLOOR || + alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT); + case CHIPREV_EVERGREEN: + default: + return !alu->is_op3 && ( + alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT); + } } /* alu instructions that can only execute on the vector unit */ @@@ -474,31 -569,26 +632,26 @@@ static int is_gpr(unsigned sel /* CB constants start at 512, and get translated to a kcache index when ALU * clauses are constructed. Note that we handle kcache constants the same way * as (the now gone) cfile constants, is that really required? */ - static int is_cb_const(int sel) + static int is_cfile(unsigned sel) { - if (sel > 511 && sel < 4607) - return 1; - return 0; + return (sel > 255 && sel < 512) || + (sel > 511 && sel < 4607) || // Kcache before translate + (sel > 127 && sel < 192); // Kcache after translate } - + static int is_const(int sel) { return is_cfile(sel) || - is_cb_const(sel) || - (sel >= V_SQ_ALU_SRC_0 && + (sel >= V_SQ_ALU_SRC_0 && sel <= V_SQ_ALU_SRC_LITERAL); } - + - static int check_vector(struct r600_bc_alu *alu, struct alu_bank_swizzle *bs, int bank_swizzle) + static int check_vector(struct r600_bc *bc, struct r600_bc_alu *alu, + struct alu_bank_swizzle *bs, int bank_swizzle) { int r, src, num_src, sel, elem, cycle; - + - num_src = r600_bc_get_num_operands(alu); + num_src = r600_bc_get_num_operands(bc, alu); for (src = 0; src < num_src; src++) { sel = alu->src[src].sel; elem = alu->src[src].chan; @@@ -522,12 -612,13 +675,13 @@@ } return 0; } - + - static int check_scalar(struct r600_bc_alu *alu, struct alu_bank_swizzle *bs, int bank_swizzle) + static int check_scalar(struct r600_bc *bc, struct r600_bc_alu *alu, + struct alu_bank_swizzle *bs, int bank_swizzle) { int r, src, num_src, const_count, sel, elem, cycle; - + - num_src = r600_bc_get_num_operands(alu); + num_src = r600_bc_get_num_operands(bc, alu); for (const_count = 0, src = 0; src < num_src; ++src) { sel = alu->src[src].sel; elem = alu->src[src].chan; @@@ -749,9 -846,9 +909,9 @@@ static int merge_inst_groups(struct r60 { struct r600_bc_alu *prev[5]; struct r600_bc_alu *result[5] = { NULL }; - + - uint32_t literal[4]; - unsigned nliteral = 0; + uint32_t literal[4], prev_literal[4]; + unsigned nliteral = 0, prev_nliteral = 0; int i, j, r, src, num_src; int num_once_inst = 0; @@@ -1009,21 -1115,15 +1178,16 @@@ int r600_bc_add_alu_type(struct r600_b if (!bc->cf_last->curr_bs_head) { bc->cf_last->curr_bs_head = nalu; } - /* at most 128 slots, one add alu can add 5 slots + 4 constants(2 slots) - * worst case */ - if (nalu->last && (bc->cf_last->ndw >> 1) >= 120) { - bc->force_add_cf = 1; - } - /* number of gpr == the last gpr used in any alu */ + /* replace special constants */ for (i = 0; i < 3; i++) { - if (nalu->src[i].sel >= bc->ngpr && nalu->src[i].sel < 128) { - bc->ngpr = nalu->src[i].sel + 1; - } if (nalu->src[i].sel == V_SQ_ALU_SRC_LITERAL) r600_bc_special_constants( - nalu->src[i].value[nalu->src[i].chan], + nalu->src[i].value[nalu->src[i].chan], &nalu->src[i].sel, &nalu->src[i].neg); + + if (nalu->src[i].sel >= bc->ngpr && nalu->src[i].sel < 128) { + bc->ngpr = nalu->src[i].sel + 1; + } } if (nalu->dst.sel >= bc->ngpr) { bc->ngpr = nalu->dst.sel + 1; @@@ -1374,765 -1448,6 +1561,766 @@@ static int r600_bc_cf_build(struct r600 return 0; } +struct gpr_usage_range { + int replacement; + int32_t start; + int32_t end; +}; + +struct gpr_usage { + unsigned channels:4; + int32_t first_write; + int32_t last_write[4]; + unsigned nranges; + struct gpr_usage_range *ranges; +}; + +static struct gpr_usage_range* add_gpr_usage_range(struct gpr_usage *usage) +{ + usage->nranges++; + usage->ranges = realloc(usage->ranges, usage->nranges * sizeof(struct gpr_usage_range)); + if (!usage->ranges) + return NULL; + return &usage->ranges[usage->nranges-1]; +} + +static void notice_gpr_read(struct gpr_usage *usage, int32_t id, unsigned chan) +{ + usage->channels |= 1 << chan; + usage->first_write = -1; + if (!usage->nranges) { + struct gpr_usage_range* range = add_gpr_usage_range(usage); + range->replacement = -1; + range->start = -1; + range->end = -1; + } + if (usage->ranges[usage->nranges-1].end < id) + usage->ranges[usage->nranges-1].end = id; +} + +static void notice_gpr_rel_read(struct gpr_usage usage[128], int32_t id, unsigned chan) +{ + unsigned i; + for (i = 0; i < 128; ++i) + notice_gpr_read(&usage[i], id, chan); +} + +static void notice_gpr_last_write(struct gpr_usage *usage, int32_t id, unsigned chan) +{ + usage->last_write[chan] = id; +} + +static void notice_gpr_write(struct gpr_usage *usage, int32_t id, unsigned chan, + int predicate, int prefered_replacement) +{ + int32_t start = usage->first_write != -1 ? usage->first_write : id; + usage->channels &= ~(1 << chan); + if (usage->channels) { + if (usage->first_write == -1) + usage->first_write = id; + } else if (!usage->nranges || (usage->ranges[usage->nranges-1].start != start && !predicate)) { + usage->first_write = start; + struct gpr_usage_range* range = add_gpr_usage_range(usage); + range->replacement = prefered_replacement; + range->start = start; + range->end = -1; + } else if (usage->ranges[usage->nranges-1].start == start && prefered_replacement != -1) { + usage->ranges[usage->nranges-1].replacement = prefered_replacement; + } + notice_gpr_last_write(usage, id, chan); +} + +static void notice_gpr_rel_last_write(struct gpr_usage usage[128], int32_t id, unsigned chan) +{ + unsigned i; + for (i = 0; i < 128; ++i) + notice_gpr_last_write(&usage[i], id, chan); +} + +static void notice_gpr_rel_write(struct gpr_usage usage[128], int32_t id, unsigned chan) +{ + unsigned i; + for (i = 0; i < 128; ++i) + notice_gpr_write(&usage[i], id, chan, 1, -1); +} + - static void notice_alu_src_gprs(struct r600_bc_alu *alu, struct gpr_usage usage[128], int32_t id) ++static void notice_alu_src_gprs(struct r600_bc *bc, struct r600_bc_alu *alu, ++ struct gpr_usage usage[128], int32_t id) +{ + unsigned src, num_src; + - num_src = r600_bc_get_num_operands(alu); ++ num_src = r600_bc_get_num_operands(bc, alu); + for (src = 0; src < num_src; ++src) { + // constants doesn't matter + if (!is_gpr(alu->src[src].sel)) + continue; + + if (alu->src[src].rel) + notice_gpr_rel_read(usage, id, alu->src[src].chan); + else + notice_gpr_read(&usage[alu->src[src].sel], id, alu->src[src].chan); + } +} + +static void notice_alu_dst_gprs(struct r600_bc_alu *alu_first, struct gpr_usage usage[128], + int32_t id, int predicate) +{ + struct r600_bc_alu *alu; + for (alu = alu_first; alu; alu = LIST_ENTRY(struct r600_bc_alu, alu->list.next, list)) { + if (alu->dst.write) { + if (alu->dst.rel) + notice_gpr_rel_write(usage, id, alu->dst.chan); + else if (alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV && is_gpr(alu->src[0].sel)) + notice_gpr_write(&usage[alu->dst.sel], id, alu->dst.chan, + predicate, alu->src[0].sel); + else + notice_gpr_write(&usage[alu->dst.sel], id, alu->dst.chan, predicate, -1); + } + + if (alu->last) + break; + } +} + +static void notice_tex_gprs(struct r600_bc_tex *tex, struct gpr_usage usage[128], + int32_t id, int predicate) +{ + if (tex->src_rel) { + if (tex->src_sel_x < 4) + notice_gpr_rel_read(usage, id, tex->src_sel_x); + if (tex->src_sel_y < 4) + notice_gpr_rel_read(usage, id, tex->src_sel_y); + if (tex->src_sel_z < 4) + notice_gpr_rel_read(usage, id, tex->src_sel_z); + if (tex->src_sel_w < 4) + notice_gpr_rel_read(usage, id, tex->src_sel_w); + } else { + if (tex->src_sel_x < 4) + notice_gpr_read(&usage[tex->src_gpr], id, tex->src_sel_x); + if (tex->src_sel_y < 4) + notice_gpr_read(&usage[tex->src_gpr], id, tex->src_sel_y); + if (tex->src_sel_z < 4) + notice_gpr_read(&usage[tex->src_gpr], id, tex->src_sel_z); + if (tex->src_sel_w < 4) + notice_gpr_read(&usage[tex->src_gpr], id, tex->src_sel_w); + } + if (tex->dst_rel) { + if (tex->dst_sel_x != 7) + notice_gpr_rel_write(usage, id, 0); + if (tex->dst_sel_y != 7) + notice_gpr_rel_write(usage, id, 1); + if (tex->dst_sel_z != 7) + notice_gpr_rel_write(usage, id, 2); + if (tex->dst_sel_w != 7) + notice_gpr_rel_write(usage, id, 3); + } else { + if (tex->dst_sel_x != 7) + notice_gpr_write(&usage[tex->dst_gpr], id, 0, predicate, -1); + if (tex->dst_sel_y != 7) + notice_gpr_write(&usage[tex->dst_gpr], id, 1, predicate, -1); + if (tex->dst_sel_z != 7) + notice_gpr_write(&usage[tex->dst_gpr], id, 2, predicate, -1); + if (tex->dst_sel_w != 7) + notice_gpr_write(&usage[tex->dst_gpr], id, 3, predicate, -1); + } +} + +static void notice_vtx_gprs(struct r600_bc_vtx *vtx, struct gpr_usage usage[128], + int32_t id, int predicate) +{ + notice_gpr_read(&usage[vtx->src_gpr], id, vtx->src_sel_x); + + if (vtx->dst_sel_x != 7) + notice_gpr_write(&usage[vtx->dst_gpr], id, 0, predicate, -1); + if (vtx->dst_sel_y != 7) + notice_gpr_write(&usage[vtx->dst_gpr], id, 1, predicate, -1); + if (vtx->dst_sel_z != 7) + notice_gpr_write(&usage[vtx->dst_gpr], id, 2, predicate, -1); + if (vtx->dst_sel_w != 7) + notice_gpr_write(&usage[vtx->dst_gpr], id, 3, predicate, -1); +} + +static void notice_export_gprs(struct r600_bc_cf *cf, struct gpr_usage usage[128], + struct r600_bc_cf *export_cf[128], int32_t export_remap[128]) +{ + //TODO handle other memory operations + struct gpr_usage *output = &usage[cf->output.gpr]; + int32_t id = (output->last_write[0] + 0x100) & ~0xFF; + + export_cf[cf->output.gpr] = cf; + export_remap[cf->output.gpr] = id; + if (cf->output.swizzle_x < 4) + notice_gpr_read(output, id, cf->output.swizzle_x); + if (cf->output.swizzle_y < 4) + notice_gpr_read(output, id, cf->output.swizzle_y); + if (cf->output.swizzle_z < 4) + notice_gpr_read(output, id, cf->output.swizzle_z); + if (cf->output.swizzle_w < 4) + notice_gpr_read(output, id, cf->output.swizzle_w); +} + +static struct gpr_usage_range *find_src_range(struct gpr_usage *usage, int32_t id) +{ + unsigned i; + for (i = 0; i < usage->nranges; ++i) { + struct gpr_usage_range* range = &usage->ranges[i]; + + if (range->start < id && id <= range->end) + return range; + } + return NULL; +} + +static struct gpr_usage_range *find_dst_range(struct gpr_usage *usage, int32_t id) +{ + unsigned i; + for (i = 0; i < usage->nranges; ++i) { + struct gpr_usage_range* range = &usage->ranges[i]; + int32_t end = range->end; + + if (range->start <= id && (id < end || end == -1)) + return range; + } + assert(0); /* should not happen */ + return NULL; +} + +static int is_barrier_needed(struct gpr_usage *usage, int32_t id, unsigned chan, int32_t last_barrier) +{ + if (usage->last_write[chan] != (id & ~0xFF)) + return usage->last_write[chan] >= last_barrier; + else + return 0; +} + +static int is_intersection(struct gpr_usage_range* a, struct gpr_usage_range* b) +{ + return a->start <= b->end && b->start < a->end; +} + +static int rate_replacement(struct gpr_usage *usage, struct gpr_usage_range* range) +{ + unsigned i; + int32_t best_start = 0x3FFFFFFF, best_end = 0x3FFFFFFF; + + for (i = 0; i < usage->nranges; ++i) { + if (usage->ranges[i].replacement != -1) + continue; /* ignore already remapped ranges */ + + if (is_intersection(&usage->ranges[i], range)) + return -1; /* forget it if usages overlap */ + + if (range->start >= usage->ranges[i].end) + best_start = MIN2(best_start, range->start - usage->ranges[i].end); + + if (range->end != -1 && range->end <= usage->ranges[i].start) + best_end = MIN2(best_end, usage->ranges[i].start - range->end); + } + return best_start + best_end; +} + +static void find_replacement(struct gpr_usage usage[128], unsigned current, + struct gpr_usage_range *range, int is_export) +{ + unsigned i; + int best_gpr = -1, best_rate = 0x7FFFFFFF; + + if (range->replacement != -1 && range->replacement <= current) { + struct gpr_usage_range *other = find_src_range(&usage[range->replacement], range->start); + if (other && other->replacement != -1) + range->replacement = other->replacement; + } + + if (range->replacement != -1 && range->replacement < current) { + int rate = rate_replacement(&usage[range->replacement], range); + + /* check if prefered replacement can be used */ + if (rate != -1) { + best_rate = rate; + best_gpr = range->replacement; + } + } + + if (best_gpr == -1 && (range->start & ~0xFF) == (range->end & ~0xFF)) { + /* register is just used inside one ALU clause */ + /* try to use clause temporaryis for it */ + for (i = 127; i > 123; --i) { + int rate = rate_replacement(&usage[i], range); + + if (rate == -1) /* can't be used because ranges overlap */ + continue; + + if (rate < best_rate) { + best_rate = rate; + best_gpr = i; + + /* can't get better than this */ + if (rate == 0 || is_export) + break; + } + } + } + + if (best_gpr == -1) { + for (i = 0; i < current; ++i) { + int rate = rate_replacement(&usage[i], range); + + if (rate == -1) /* can't be used because ranges overlap */ + continue; + + if (rate < best_rate) { + best_rate = rate; + best_gpr = i; + + /* can't get better than this */ + if (rate == 0) + break; + } + } + } + + range->replacement = best_gpr; + if (best_gpr != -1) { + struct gpr_usage_range *reservation = add_gpr_usage_range(&usage[best_gpr]); + reservation->replacement = -1; + reservation->start = range->start; + reservation->end = range->end; + } +} + +static void find_export_replacement(struct gpr_usage usage[128], + struct gpr_usage_range *range, struct r600_bc_cf *current, + struct r600_bc_cf *next, int32_t next_id) +{ + if (!next || next_id <= range->start || next_id > range->end) + return; + + if (current->output.type != next->output.type) + return; + + if ((current->output.array_base + 1) != next->output.array_base) + return; + + find_src_range(&usage[next->output.gpr], next_id)->replacement = range->replacement + 1; +} + - static void replace_alu_gprs(struct r600_bc_alu *alu, struct gpr_usage usage[128], ++static void replace_alu_gprs(struct r600_bc *bc, struct r600_bc_alu *alu, struct gpr_usage usage[128], + int32_t id, int32_t last_barrier, unsigned *barrier) +{ + struct gpr_usage *cur_usage; + struct gpr_usage_range *range; + unsigned src, num_src; + - num_src = r600_bc_get_num_operands(alu); ++ num_src = r600_bc_get_num_operands(bc, alu); + for (src = 0; src < num_src; ++src) { + // constants doesn't matter + if (!is_gpr(alu->src[src].sel)) + continue; + + cur_usage = &usage[alu->src[src].sel]; + range = find_src_range(cur_usage, id); + if (range->replacement != -1) + alu->src[src].sel = range->replacement; + + *barrier |= is_barrier_needed(cur_usage, id, alu->src[src].chan, last_barrier); + } + + if (alu->dst.write) { + cur_usage = &usage[alu->dst.sel]; + range = find_dst_range(cur_usage, id); + if (range->replacement == alu->dst.sel) { + if (!alu->is_op3) + alu->dst.write = 0; + else + /*TODO: really check that register 123 is useable */ + alu->dst.sel = 123; + } else if (range->replacement != -1) { + alu->dst.sel = range->replacement; + } + if (alu->dst.rel) + notice_gpr_rel_last_write(usage, id, alu->dst.chan); + else + notice_gpr_last_write(cur_usage, id, alu->dst.chan); + } +} + +static void replace_tex_gprs(struct r600_bc_tex *tex, struct gpr_usage usage[128], + int32_t id, int32_t last_barrier, unsigned *barrier) +{ + struct gpr_usage *cur_usage = &usage[tex->src_gpr]; + struct gpr_usage_range *range = find_src_range(cur_usage, id); + + if (tex->src_rel) { + *barrier = 1; + } else { + if (tex->src_sel_x < 4) + *barrier |= is_barrier_needed(cur_usage, id, tex->src_sel_x, last_barrier); + if (tex->src_sel_y < 4) + *barrier |= is_barrier_needed(cur_usage, id, tex->src_sel_y, last_barrier); + if (tex->src_sel_z < 4) + *barrier |= is_barrier_needed(cur_usage, id, tex->src_sel_z, last_barrier); + if (tex->src_sel_w < 4) + *barrier |= is_barrier_needed(cur_usage, id, tex->src_sel_w, last_barrier); + } + + if (range->replacement != -1) + tex->src_gpr = range->replacement; + + cur_usage = &usage[tex->dst_gpr]; + range = find_dst_range(cur_usage, id); + if (range->replacement != -1) + tex->dst_gpr = range->replacement; + + if (tex->dst_rel) { + if (tex->dst_sel_x != 7) + notice_gpr_rel_last_write(usage, id, tex->dst_sel_x); + if (tex->dst_sel_y != 7) + notice_gpr_rel_last_write(usage, id, tex->dst_sel_y); + if (tex->dst_sel_z != 7) + notice_gpr_rel_last_write(usage, id, tex->dst_sel_z); + if (tex->dst_sel_w != 7) + notice_gpr_rel_last_write(usage, id, tex->dst_sel_w); + } else { + if (tex->dst_sel_x != 7) + notice_gpr_last_write(cur_usage, id, tex->dst_sel_x); + if (tex->dst_sel_y != 7) + notice_gpr_last_write(cur_usage, id, tex->dst_sel_y); + if (tex->dst_sel_z != 7) + notice_gpr_last_write(cur_usage, id, tex->dst_sel_z); + if (tex->dst_sel_w != 7) + notice_gpr_last_write(cur_usage, id, tex->dst_sel_w); + } +} + +static void replace_vtx_gprs(struct r600_bc_vtx *vtx, struct gpr_usage usage[128], + int32_t id, int32_t last_barrier, unsigned *barrier) +{ + struct gpr_usage *cur_usage = &usage[vtx->src_gpr]; + struct gpr_usage_range *range = find_src_range(cur_usage, id); + + *barrier |= is_barrier_needed(cur_usage, id, vtx->src_sel_x, last_barrier); + + if (range->replacement != -1) + vtx->src_gpr = range->replacement; + + cur_usage = &usage[vtx->dst_gpr]; + range = find_dst_range(cur_usage, id); + if (range->replacement != -1) + vtx->dst_gpr = range->replacement; + + if (vtx->dst_sel_x != 7) + notice_gpr_last_write(cur_usage, id, vtx->dst_sel_x); + if (vtx->dst_sel_y != 7) + notice_gpr_last_write(cur_usage, id, vtx->dst_sel_y); + if (vtx->dst_sel_z != 7) + notice_gpr_last_write(cur_usage, id, vtx->dst_sel_z); + if (vtx->dst_sel_w != 7) + notice_gpr_last_write(cur_usage, id, vtx->dst_sel_w); +} + +static void replace_export_gprs(struct r600_bc_cf *cf, struct gpr_usage usage[128], + int32_t id, int32_t last_barrier) +{ + //TODO handle other memory operations + struct gpr_usage *cur_usage = &usage[cf->output.gpr]; + struct gpr_usage_range *range = find_src_range(cur_usage, id); + + cf->barrier = 0; + if (cf->output.swizzle_x < 4) + cf->barrier |= is_barrier_needed(cur_usage, -1, cf->output.swizzle_x, last_barrier); + if (cf->output.swizzle_y < 4) + cf->barrier |= is_barrier_needed(cur_usage, -1, cf->output.swizzle_y, last_barrier); + if (cf->output.swizzle_z < 4) + cf->barrier |= is_barrier_needed(cur_usage, -1, cf->output.swizzle_z, last_barrier); + if (cf->output.swizzle_w < 4) + cf->barrier |= is_barrier_needed(cur_usage, -1, cf->output.swizzle_w, last_barrier); + + if (range->replacement != -1) + cf->output.gpr = range->replacement; +} + - static void optimize_alu_inst(struct r600_bc_cf *cf, struct r600_bc_alu *alu) ++static void optimize_alu_inst(struct r600_bc *bc, struct r600_bc_cf *cf, struct r600_bc_alu *alu) +{ + struct r600_bc_alu *alu_next; + unsigned chan; + unsigned src, num_src; + + /* check if a MOV could be optimized away */ + if (alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV) { + + /* destination equals source? */ + if (alu->dst.sel != alu->src[0].sel || + alu->dst.chan != alu->src[0].chan) + return; + + /* any special handling for the source? */ + if (alu->src[0].rel || alu->src[0].neg || alu->src[0].abs) + return; + + /* any special handling for destination? */ + if (alu->dst.rel || alu->dst.clamp) + return; + + /* ok find next instruction group and check if ps/pv is used */ + for (alu_next = alu; !alu_next->last; alu_next = NEXT_ALU(alu_next)); + + if (alu_next->list.next != &cf->alu) { - chan = is_alu_reduction_inst(alu) ? 0 : alu->dst.chan; ++ chan = is_alu_reduction_inst(bc, alu) ? 0 : alu->dst.chan; + for (alu_next = NEXT_ALU(alu_next); alu_next; alu_next = NEXT_ALU(alu_next)) { - num_src = r600_bc_get_num_operands(alu_next); ++ num_src = r600_bc_get_num_operands(bc, alu_next); + for (src = 0; src < num_src; ++src) { + if (alu_next->src[src].sel == V_SQ_ALU_SRC_PV && + alu_next->src[src].chan == chan) + return; + + if (alu_next->src[src].sel == V_SQ_ALU_SRC_PS) + return; + } + + if (alu_next->last) + break; + } + } + + r600_bc_remove_alu(cf, alu); + } +} + +static void optimize_export_inst(struct r600_bc *bc, struct r600_bc_cf *cf) +{ + struct r600_bc_cf *prev = LIST_ENTRY(struct r600_bc_cf, cf->list.prev, list); + if (&prev->list == &bc->cf || + prev->inst != cf->inst || + prev->output.type != cf->output.type || + prev->output.elem_size != cf->output.elem_size || + prev->output.swizzle_x != cf->output.swizzle_x || + prev->output.swizzle_y != cf->output.swizzle_y || + prev->output.swizzle_z != cf->output.swizzle_z || + prev->output.swizzle_w != cf->output.swizzle_w) + return; + + if ((prev->output.burst_count + cf->output.burst_count) > 16) + return; + + if ((prev->output.gpr + prev->output.burst_count) == cf->output.gpr && + (prev->output.array_base + prev->output.burst_count) == cf->output.array_base) { + + prev->output.burst_count += cf->output.burst_count; + r600_bc_remove_cf(bc, cf); + + } else if (prev->output.gpr == (cf->output.gpr + cf->output.burst_count) && + prev->output.array_base == (cf->output.array_base + cf->output.burst_count)) { + + cf->output.burst_count += prev->output.burst_count; + r600_bc_remove_cf(bc, prev); + } +} + +static void r600_bc_optimize(struct r600_bc *bc) +{ + struct r600_bc_cf *cf, *next_cf; + struct r600_bc_alu *first, *next_alu; + struct r600_bc_alu *alu; + struct r600_bc_vtx *vtx; + struct r600_bc_tex *tex; + struct gpr_usage usage[128]; + + /* assume that each gpr is exported only once */ + struct r600_bc_cf *export_cf[128] = { NULL }; + int32_t export_remap[128]; + + int32_t id, barrier[bc->nstack]; + unsigned i, j, stack, predicate, old_stack; + + memset(&usage, 0, sizeof(usage)); + for (i = 0; i < 128; ++i) { + usage[i].first_write = -1; + usage[i].last_write[0] = -1; + usage[i].last_write[1] = -1; + usage[i].last_write[2] = -1; + usage[i].last_write[3] = -1; + } + + /* first gather some informations about the gpr usage */ + id = 0; stack = 0; + LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) { + switch (get_cf_class(cf)) { + case CF_CLASS_ALU: + predicate = 0; + first = NULL; + LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) { + if (!first) + first = alu; - notice_alu_src_gprs(alu, usage, id); ++ notice_alu_src_gprs(bc, alu, usage, id); + if (alu->last) { + notice_alu_dst_gprs(first, usage, id, predicate || stack > 0); + first = NULL; + ++id; + } - if (is_alu_pred_inst(alu)) ++ if (is_alu_pred_inst(bc, alu)) + predicate++; + } + if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3) + stack += predicate; + else if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3) + stack -= 1; + else if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3) + stack -= 2; + break; + case CF_CLASS_TEXTURE: + LIST_FOR_EACH_ENTRY(tex, &cf->tex, list) { + notice_tex_gprs(tex, usage, id++, stack > 0); + } + break; + case CF_CLASS_VERTEX: + LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) { + notice_vtx_gprs(vtx, usage, id++, stack > 0); + } + break; + case CF_CLASS_EXPORT: + notice_export_gprs(cf, usage, export_cf, export_remap); + continue; // don't increment id + case CF_CLASS_OTHER: + switch (cf->inst) { + case V_SQ_CF_WORD1_SQ_CF_INST_JUMP: + case V_SQ_CF_WORD1_SQ_CF_INST_ELSE: + case V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS: + break; + + case V_SQ_CF_WORD1_SQ_CF_INST_POP: + stack -= cf->pop_count; + break; + + default: + // TODO implement loop handling + goto out; + } + } + id += 0x100; + id &= ~0xFF; + } + assert(stack == 0); + + /* try to optimize gpr usage */ + for (i = 0; i < 124; ++i) { + for (j = 0; j < usage[i].nranges; ++j) { + struct gpr_usage_range *range = &usage[i].ranges[j]; + int is_export = export_cf[i] && export_cf[i + 1] && + range->start < export_remap[i] && + export_remap[i] <= range->end; + + if (range->start == -1) + range->replacement = -1; + else if (range->end == -1) + range->replacement = i; + else + find_replacement(usage, i, range, is_export); + + if (range->replacement == -1) + bc->ngpr = i; + else if (range->replacement < i && range->replacement > bc->ngpr) + bc->ngpr = range->replacement; + + if (is_export && range->replacement != -1) { + find_export_replacement(usage, range, export_cf[i], + export_cf[i + 1], export_remap[i + 1]); + } + } + } + bc->ngpr++; + + /* apply the changes */ + for (i = 0; i < 128; ++i) { + usage[i].last_write[0] = -1; + usage[i].last_write[1] = -1; + usage[i].last_write[2] = -1; + usage[i].last_write[3] = -1; + } + barrier[0] = 0; + id = 0; stack = 0; + LIST_FOR_EACH_ENTRY_SAFE(cf, next_cf, &bc->cf, list) { + old_stack = stack; + switch (get_cf_class(cf)) { + case CF_CLASS_ALU: + predicate = 0; + first = NULL; + cf->barrier = 0; + LIST_FOR_EACH_ENTRY_SAFE(alu, next_alu, &cf->alu, list) { - replace_alu_gprs(alu, usage, id, barrier[stack], &cf->barrier); ++ replace_alu_gprs(bc, alu, usage, id, barrier[stack], &cf->barrier); + if (alu->last) + ++id; + - if (is_alu_pred_inst(alu)) ++ if (is_alu_pred_inst(bc, alu)) + predicate++; + + if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3) - optimize_alu_inst(cf, alu); ++ optimize_alu_inst(bc, cf, alu); + } + if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3) + stack += predicate; + else if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3) + stack -= 1; + else if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3) + stack -= 2; + if (LIST_IS_EMPTY(&cf->alu)) { + r600_bc_remove_cf(bc, cf); + cf = NULL; + } + break; + case CF_CLASS_TEXTURE: + cf->barrier = 0; + LIST_FOR_EACH_ENTRY(tex, &cf->tex, list) { + replace_tex_gprs(tex, usage, id++, barrier[stack], &cf->barrier); + } + break; + case CF_CLASS_VERTEX: + cf->barrier = 0; + LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) { + replace_vtx_gprs(vtx, usage, id++, barrier[stack], &cf->barrier); + } + break; + case CF_CLASS_EXPORT: + continue; // don't increment id + case CF_CLASS_OTHER: + if (cf->inst == V_SQ_CF_WORD1_SQ_CF_INST_POP) { + cf->barrier = 0; + stack -= cf->pop_count; + } + break; + } + + id &= ~0xFF; + if (cf && cf->barrier) + barrier[old_stack] = id; + + for (i = old_stack + 1; i <= stack; ++i) + barrier[i] = barrier[old_stack]; + + id += 0x100; + if (stack != 0) /* ensue exports are placed outside of conditional blocks */ + continue; + + for (i = 0; i < 128; ++i) { + if (!export_cf[i] || id < export_remap[i]) + continue; + + r600_bc_move_cf(bc, export_cf[i], next_cf); + replace_export_gprs(export_cf[i], usage, export_remap[i], barrier[stack]); + if (export_cf[i]->barrier) + barrier[stack] = id - 1; + next_cf = LIST_ENTRY(struct r600_bc_cf, export_cf[i]->list.next, list); + optimize_export_inst(bc, export_cf[i]); + export_cf[i] = NULL; + } + } + assert(stack == 0); + +out: + for (i = 0; i < 128; ++i) { + free(usage[i].ranges); + } +} + int r600_bc_build(struct r600_bc *bc) { struct r600_bc_cf *cf; @@@ -2151,27 -1465,19 +2339,17 @@@ bc->nstack = 1; } + r600_bc_optimize(bc); + /* first path compute addr of each CF block */ /* addr start after all the CF instructions */ - addr = bc->cf_last->id + 2; + addr = LIST_ENTRY(struct r600_bc_cf, bc->cf.prev, list)->id + 2; LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) { - switch (cf->inst) { - case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3): - case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3): - case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3): - case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3): + switch (get_cf_class(cf)) { + case CF_CLASS_ALU: - nliteral = 0; - LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) { - r = r600_bc_alu_nliterals(alu, literal, &nliteral); - if (r) - return r; - if (alu->last) { - cf->ndw += align(nliteral, 2); - nliteral = 0; - } - } break; - case V_SQ_CF_WORD1_SQ_CF_INST_TEX: - case V_SQ_CF_WORD1_SQ_CF_INST_VTX: - case V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC: + case CF_CLASS_TEXTURE: + case CF_CLASS_VERTEX: /* fetch node need to be 16 bytes aligned*/ addr += 3; addr &= 0xFFFFFFFCUL; @@@ -2211,14 -1517,18 +2389,15 @@@ r = r600_bc_cf_build(bc, cf); if (r) return r; - switch (cf->inst) { - case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3): - case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3): - case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3): - case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3): + switch (get_cf_class(cf)) { + case CF_CLASS_ALU: nliteral = 0; + memset(literal, 0, sizeof(literal)); LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) { - r = r600_bc_alu_nliterals(alu, literal, &nliteral); + r = r600_bc_alu_nliterals(bc, alu, literal, &nliteral); if (r) return r; - r600_bc_alu_adjust_literals(alu, literal, nliteral); + r600_bc_alu_adjust_literals(bc, alu, literal, nliteral); switch(bc->chiprev) { case CHIPREV_R600: r = r600_bc_alu_build(bc, alu, addr);