diff options
| -rw-r--r-- | plugins/gpu_neon/psx_gpu/psx_gpu.c | 15 | ||||
| -rw-r--r-- | plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S | 13 |
2 files changed, 25 insertions, 3 deletions
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.c b/plugins/gpu_neon/psx_gpu/psx_gpu.c index 485ef27..2cba878 100644 --- a/plugins/gpu_neon/psx_gpu/psx_gpu.c +++ b/plugins/gpu_neon/psx_gpu/psx_gpu.c @@ -769,13 +769,18 @@ void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a, #ifndef NDEBUG #define setup_spans_debug_check(span_edge_data_element) \ - if (&span_edge_data_element - psx_gpu->span_edge_data < psx_gpu->num_spans) \ +{ \ + u32 _num_spans = &span_edge_data_element - psx_gpu->span_edge_data; \ + if (_num_spans > MAX_SPANS) \ + *(int *)0 = 1; \ + if (_num_spans < psx_gpu->num_spans) \ { \ if(span_edge_data_element.num_blocks > MAX_BLOCKS_PER_ROW) \ *(int *)0 = 1; \ if(span_edge_data_element.y > 2048) \ *(int *)0 = 1; \ } \ +} \ #else #define setup_spans_debug_check(span_edge_data_element) \ @@ -1423,12 +1428,16 @@ void setup_spans_up_down(psx_gpu_struct *psx_gpu, vertex_struct *v_a, y_x4.e[3] = y_a + 3; setup_spans_adjust_edges_alternate_no(index_left, index_right); + // FIXME: overflow corner case + if(psx_gpu->num_spans + height_minor_b == MAX_SPANS) + height_minor_b &= ~3; + psx_gpu->num_spans += height_minor_b; - do + while(height_minor_b > 0) { setup_spans_set_x4(none, down, no); height_minor_b -= 4; - } while(height_minor_b > 0); + } } left_split_triangles++; 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 a2bfa5b..4e1e403 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 |