X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=blobdiff_plain;f=plugins%2Fgpu_neon%2Fpsx_gpu%2Fpsx_gpu_arm_neon.S;h=4e1e4032dcf3fbb759f4b0f3cc54fd59325de42c;hp=085e11b07f258a4b5999bc0481794d045fb0f9b1;hb=b7569147823a8fc5a9de98e5d491da906e119296;hpb=f0931e56b2428fe5e0f6b4d7d6d0f41462cfc551

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
index 085e11b0..4e1e4032 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
@@ -1175,6 +1175,10 @@ function(setup_spans_up_down)
 
   ldrh temp, [ psx_gpu, #psx_gpu_num_spans_offset ]
   add temp, temp, height_minor_b
+
+  cmp temp, #MAX_SPANS
+  beq 5f
+
   strh temp, [ psx_gpu, #psx_gpu_num_spans_offset ]
 
  2:                                                     
@@ -1190,6 +1194,15 @@ function(setup_spans_up_down)
   setup_spans_prologue_b()
   bal 4b
 
+ 5:
+  // FIXME: overflow corner case
+  sub temp, temp, height_minor_b
+  bics height_minor_b, #3
+  add temp, temp, height_minor_b
+  strh temp, [ psx_gpu, #psx_gpu_num_spans_offset ]
+  bne 2b
+  bal 1b
+
 .pool
 
 #undef span_uvrg_offset
@@ -3392,10 +3405,12 @@ function(shade_blocks_textured_unmodulated_direct)
    [ draw_mask_bits_ptr, :16 ], c_64
   vbif.u16 fb_pixels, pixels, draw_mask_combined
 
-  vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64
-
   sub fb_ptr_cmp, fb_ptr_next, fb_ptr
+  pld [ fb_ptr_next, #64 ]
+
   add fb_ptr_cmp, fb_ptr_cmp, #14
+  vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64
+
   cmp fb_ptr_cmp, #28
   bls 4f
 
@@ -3754,11 +3769,15 @@ function(blend_blocks_textured_add_##mask_evaluate)                            \
   vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g;                            \
   vand.u16 pixels_mg, pixels, d128_0x83E0;                                     \
                                                                                \
-  vbit.u16 blend_pixels, fb_pixels, draw_mask;                                 \
-  vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64;                       \
+  sub fb_ptr_cmp, fb_ptr_next, fb_ptr;                                         \
+  pld [ fb_ptr_next, #64 ];                                                    \
                                                                                \
   sub fb_ptr_cmp, fb_ptr_next, fb_ptr;                                         \
+  vbit.u16 blend_pixels, fb_pixels, draw_mask;                                 \
+                                                                               \
   add fb_ptr_cmp, fb_ptr_cmp, #14;                                             \
+  vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64;                       \
+                                                                               \
   cmp fb_ptr_cmp, #28;                                                         \
   bls 2f;                                                                      \
                                                                                \
@@ -4917,12 +4936,12 @@ setup_sprite_update_texture_8bpp_cache:
    draw_mask_fb_ptr_left_b);                                                   \
                                                                                \
   add texture_block_ptr, texture_page_ptr, texture_block_ptr;                  \
-  add fb_ptr, fb_ptr, #16*2;                                                   \
+  pld [ fb_ptr, #2048 ];                                                       \
                                                                                \
   vld1.u32 { texels }, [ texture_block_ptr, :64 ];                             \
-  vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels;                       \
+  add fb_ptr, fb_ptr, #16*2;                                                   \
                                                                                \
-  pld [ fb_ptr ];                                                              \
+  vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels;                       \
   vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels;                    \
                                                                                \
   vzip.8 texels_low, texels_high;                                              \
@@ -4961,9 +4980,10 @@ setup_sprite_update_texture_8bpp_cache:
   do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a,              \
    draw_mask_fb_ptr_##edge##_b);                                               \
                                                                                \
+  pld [ fb_ptr, #2048 ];                                                       \
   add fb_ptr, fb_ptr, #2048 * 2;                                               \
-  subs sub_tile_height, sub_tile_height, #1;                                   \
                                                                                \
+  subs sub_tile_height, sub_tile_height, #1;                                   \
   bne 4b;                                                                      \
                                                                                \
   ldr column_data, [sp], #8; /* fb_ptr2 */                                     \
@@ -4987,13 +5007,13 @@ setup_sprite_update_texture_8bpp_cache:
   do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a,                   \
    draw_mask_fb_ptr_left_b);                                                   \
                                                                                \
+  pld [ fb_ptr, #2048 ];                                                       \
   and texture_block_ptr, texture_block_ptr, texture_mask;                      \
                                                                                \
   add fb_ptr, fb_ptr, #16*2;                                                   \
   add texture_block_ptr, texture_page_ptr, texture_block_ptr;                  \
                                                                                \
   vld1.u32 { texels }, [ texture_block_ptr, :64 ];                             \
-  pld [ fb_ptr ];                                                              \
                                                                                \
   do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a,                  \
    draw_mask_fb_ptr_right_b);                                                  \
@@ -5022,6 +5042,7 @@ setup_sprite_update_texture_8bpp_cache:
   add texture_block_ptr, texture_page_ptr, texture_block_ptr;                  \
   vld1.u32 { texels }, [ texture_block_ptr, :64 ];                             \
                                                                                \
+  pld [ fb_ptr, #2048 ];                                                       \
   do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a,               \
    draw_mask_fb_ptr_##edge##_b);                                               \
                                                                                \