turnip: skip unnecessary CP_REG_TEST for cond load/store

When no attachment allows conditional load/store, skip the unnecessary
CP_REG_TEST.

This is done to avoid a performance trap on a618 (gen1).  CP_REG_TEST or
any command that reads a register is slow on a618.

glmark2 score goes from 830 to 1001.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/8162
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21206>

src/freedreno/vulkan/tu_cmd_buffer.c

@@ -736,7 +736,8 @@ static void
 tu6_emit_cond_for_load_stores(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
                               uint32_t pipe, uint32_t slot, bool skip_wfm)
 {
-   if (cmd->state.tiling->binning_possible) {
+   if (cmd->state.tiling->binning_possible &&
+       cmd->state.pass->has_cond_load_store) {
       tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
       tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(REG_A6XX_VSC_STATE_REG(pipe)) |
                      A6XX_CP_REG_TEST_0_BIT(slot) |
@@ -855,8 +856,10 @@ tu6_emit_tile_load(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
    tu6_emit_blit_scissor(cmd, cs, true);
 
+   const bool cond_exec_allowed = cmd->state.tiling->binning &&
+                                  cmd->state.pass->has_cond_load_store;
    for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
-      tu_load_gmem_attachment(cmd, cs, i, cmd->state.tiling->binning, false);
+      tu_load_gmem_attachment(cmd, cs, i, cond_exec_allowed, false);
 }
 
 static void
@@ -873,9 +876,11 @@ tu6_emit_tile_store(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 
    for (uint32_t a = 0; a < pass->attachment_count; ++a) {
       if (pass->attachments[a].gmem) {
+         const bool cond_exec_allowed = cmd->state.tiling->binning_possible &&
+                                        cmd->state.pass->has_cond_load_store;
          tu_store_gmem_attachment(cmd, cs, a, a,
                                   fb->layers, subpass->multiview_mask,
-                                  cmd->state.tiling->binning_possible);
+                                  cond_exec_allowed);
       }
    }
 

src/freedreno/vulkan/tu_pass.c

@@ -547,6 +547,9 @@ tu_render_pass_cond_config(struct tu_render_pass *pass)
          (att->load || att->load_stencil) && !att->clear_mask && !att->will_be_resolved;
       att->cond_store_allowed =
          (att->store || att->store_stencil) && !att->clear_mask;
+
+      pass->has_cond_load_store |=
+         att->cond_load_allowed | att->cond_store_allowed;
    }
 }
 

src/freedreno/vulkan/tu_pass.h

@@ -108,7 +108,10 @@ struct tu_render_pass
    uint32_t sysmem_bandwidth_per_pixel;
 
    struct tu_subpass_attachment *subpass_attachments;
+
    struct tu_render_pass_attachment *attachments;
+   bool has_cond_load_store;
+
    struct tu_subpass_barrier end_barrier;
    struct tu_subpass subpasses[0];
 };