From: Rob Clark Date: Wed, 29 May 2019 20:25:11 +0000 (-0700) Subject: freedreno/ir3: set (ss) on last_input if ldlv X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=5d43b806ba3a2026d11e20ec655a9ead116c88ef;p=mesa.git freedreno/ir3: set (ss) on last_input if ldlv It seems like (ei) handling doesn't sync on (ss), so we could end up in a situation where we release varying storage before an ldlv for flat shaded varyings completes. Keep track if we've done an (ss) since the last ldlv, and if not add (ss) flag to last_input which gets (ei). Noticed with dEQP-GLES3.functional.fragment_out.random.24 and dEQP-GLES3.functional.fragment_out.random.27, which previously passed by luck because ir3_sched ordered instructions in a way that resulted in a lucky (ss). Signed-off-by: Rob Clark Acked-by: Eric Anholt --- diff --git a/src/freedreno/ir3/ir3_legalize.c b/src/freedreno/ir3/ir3_legalize.c index a5e5fd1fa26..c39b3a11011 100644 --- a/src/freedreno/ir3/ir3_legalize.c +++ b/src/freedreno/ir3/ir3_legalize.c @@ -87,6 +87,7 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block) struct list_head instr_list; struct ir3_legalize_state prev_state = bd->state; struct ir3_legalize_state *state = &bd->state; + bool last_input_needs_ss = false; /* our input state is the OR of all predecessor blocks' state: */ for (unsigned i = 0; i < block->predecessors_count; i++) { @@ -125,8 +126,10 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block) ctx->max_bary = MAX2(ctx->max_bary, inloc->iim_val); } - if (last_n && is_barrier(last_n)) + if (last_n && is_barrier(last_n)) { n->flags |= IR3_INSTR_SS | IR3_INSTR_SY; + last_input_needs_ss = false; + } /* NOTE: consider dst register too.. it could happen that * texture sample instruction (for example) writes some @@ -145,6 +148,7 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block) */ if (regmask_get(&state->needs_ss, reg)) { n->flags |= IR3_INSTR_SS; + last_input_needs_ss = false; regmask_init(&state->needs_ss_war); regmask_init(&state->needs_ss); } @@ -167,6 +171,7 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block) reg = n->regs[0]; if (regmask_get(&state->needs_ss_war, reg)) { n->flags |= IR3_INSTR_SS; + last_input_needs_ss = false; regmask_init(&state->needs_ss_war); regmask_init(&state->needs_ss); } @@ -237,6 +242,7 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block) } else if (n->opc == OPC_RESINFO) { regmask_set(&state->needs_ss, n->regs[0]); ir3_NOP(block)->flags |= IR3_INSTR_SS; + last_input_needs_ss = false; } else if (is_load(n)) { /* seems like ldlv needs (ss) bit instead?? which is odd but * makes a bunch of flat-varying tests start working on a4xx. @@ -271,8 +277,10 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block) } } - if (is_input(n)) + if (is_input(n)) { last_input = n; + last_input_needs_ss |= (n->opc == OPC_LDLV); + } last_n = n; } @@ -289,7 +297,6 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block) /* (ss)bary.f (ei)r63.x, 0, r0.x */ baryf = ir3_instr_create(block, OPC_BARY_F); - baryf->flags |= IR3_INSTR_SS; ir3_reg_create(baryf, regid(63, 0), 0); ir3_reg_create(baryf, 0, IR3_REG_IMMED)->iim_val = 0; ir3_reg_create(baryf, regid(0, 0), 0); @@ -301,6 +308,8 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block) last_input = baryf; } last_input->regs[0]->flags |= IR3_REG_EI; + if (last_input_needs_ss) + last_input->flags |= IR3_INSTR_SS; } if (last_rel)