panfrost: fix undefined value access on mir_set_intr_mask()

[mesa.git] / src / panfrost / midgard / midgard_compile.c

diff --git a/src/panfrost/midgard/midgard_compile.c b/src/panfrost/midgard/midgard_compile.c
index 4b15ef93209ca6578bf3161aa3a242afb92f3da3..06ed3f62cf44a74677620f54405513c55be823c2 100644 (file)
--- a/src/panfrost/midgard/midgard_compile.c
+++ b/src/panfrost/midgard/midgard_compile.c
@@ -497,6 +497,7 @@ optimise_nir(nir_shader *nir, unsigned quirks, bool is_blend)
                 .lower_txp = ~0,
                 .lower_tex_without_implicit_lod =
                         (quirks & MIDGARD_EXPLICIT_LOD),
+                .lower_tg4_broadcom_swizzle = true,
 
                 /* TODO: we have native gradient.. */
                 .lower_txd = true,
@@ -1290,8 +1291,8 @@ mir_set_intr_mask(nir_instr *instr, midgard_instruction *ins, bool is_read)
 
         /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
         unsigned bytemask = pan_to_bytemask(dsize, nir_mask);
-        mir_set_bytemask(ins, bytemask);
         ins->dest_type = nir_type_uint | dsize;
+        mir_set_bytemask(ins, bytemask);
 }
 
 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
@@ -1357,6 +1358,15 @@ emit_global(
         mir_set_offset(ctx, &ins, offset, is_shared);
         mir_set_intr_mask(instr, &ins, is_read);
 
+        /* Set a valid swizzle for masked out components */
+        assert(ins.mask);
+        unsigned first_component = __builtin_ffs(ins.mask) - 1;
+
+        for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[0]); ++i) {
+                if (!(ins.mask & (1 << i)))
+                        ins.swizzle[0][i] = first_component;
+        }
+
         emit_mir_instruction(ctx, ins);
 }
 
@@ -1570,13 +1580,13 @@ emit_vertex_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
 }
 
 static void
-emit_msaa_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
+emit_special(compiler_context *ctx, nir_intrinsic_instr *instr, unsigned idx)
 {
         unsigned reg = nir_dest_index(&instr->dest);
 
         midgard_instruction ld = m_ld_color_buffer_32u(reg, 0);
         ld.op = midgard_op_ld_color_buffer_32u_old;
-        ld.load_store.address = 97;
+        ld.load_store.address = idx;
         ld.load_store.arg_2 = 0x1E;
 
         for (int i = 0; i < 4; ++i)
@@ -1970,8 +1980,12 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
                 emit_vertex_builtin(ctx, instr);
                 break;
 
+        case nir_intrinsic_load_sample_mask_in:
+                emit_special(ctx, instr, 96);
+                break;
+
         case nir_intrinsic_load_sample_id:
-                emit_msaa_builtin(ctx, instr);
+                emit_special(ctx, instr, 97);
                 break;
 
         case nir_intrinsic_memory_barrier_buffer:
@@ -2052,6 +2066,19 @@ pan_attach_constant_bias(
         return true;
 }
 
+static enum mali_texture_mode
+mdg_texture_mode(nir_tex_instr *instr)
+{
+        if (instr->op == nir_texop_tg4 && instr->is_shadow)
+                return TEXTURE_GATHER_SHADOW;
+        else if (instr->op == nir_texop_tg4)
+                return TEXTURE_GATHER_X + instr->component;
+        else if (instr->is_shadow)
+                return TEXTURE_SHADOW;
+        else
+                return TEXTURE_NORMAL;
+}
+
 static void
 emit_texop_native(compiler_context *ctx, nir_tex_instr *instr,
                   unsigned midgard_texop)
@@ -2086,11 +2113,11 @@ emit_texop_native(compiler_context *ctx, nir_tex_instr *instr,
                         .format = midgard_tex_format(instr->sampler_dim),
                         .texture_handle = texture_index,
                         .sampler_handle = sampler_index,
-                        .shadow = instr->is_shadow,
+                        .mode = mdg_texture_mode(instr)
                 }
         };
 
-        if (instr->is_shadow && !instr->is_new_style_shadow)
+        if (instr->is_shadow && !instr->is_new_style_shadow && instr->op != nir_texop_tg4)
            for (int i = 0; i < 4; ++i)
               ins.swizzle[0][i] = COMPONENT_X;
 
@@ -2260,6 +2287,7 @@ emit_tex(compiler_context *ctx, nir_tex_instr *instr)
                 emit_texop_native(ctx, instr, TEXTURE_OP_NORMAL);
                 break;
         case nir_texop_txl:
+        case nir_texop_tg4:
                 emit_texop_native(ctx, instr, TEXTURE_OP_LOD);
                 break;
         case nir_texop_txf: