Import Mesa 13.0.2

5 files changed, 436 insertions, 45 deletions

diff --git a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_drv.h b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_drv.h
index ffc973735..90f45397d 100644
--- a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_drv.h
+++ b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_drv.h
@@ -81,6 +81,8 @@ struct vc4_exec_info {
 	struct drm_gem_cma_object *tile_bo;
 	uint32_t tile_alloc_offset;
 
+	uint32_t tile_width, tile_height;
+
 	/**
 	 * Computed addresses pointing into exec_bo where we start the
 	 * bin thread (ct0) and render thread (ct1).
@@ -145,6 +147,9 @@ struct vc4_validated_shader_info
 	uint32_t uniforms_src_size;
 	uint32_t num_texture_samples;
 	struct vc4_texture_sample_info *texture_samples;
+
+	uint32_t num_uniform_addr_offsets;
+	uint32_t *uniform_addr_offsets;
 };
 
 /* vc4_validate.c */
diff --git a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_gem.c b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_gem.c
index 93f9ec7ed..12350b8d3 100644
--- a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_gem.c
+++ b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_gem.c
@@ -139,8 +139,17 @@ fail:
 int
 vc4_cl_validate(struct drm_device *dev, struct vc4_exec_info *exec)
 {
+	struct drm_vc4_submit_cl *args = exec->args;
 	int ret = 0;
 
+	if (args->color_write.bits & VC4_RENDER_CONFIG_MS_MODE_4X) {
+		exec->tile_width = 32;
+		exec->tile_height = 32;
+	} else {
+		exec->tile_width = 64;
+		exec->tile_height = 64;
+	}
+
 	if (exec->args->bin_cl_size != 0) {
 		ret = vc4_get_bcl(dev, exec);
 		if (ret)
diff --git a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_render_cl.c b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_render_cl.c
index d9c68423f..b926d35a6 100644
--- a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_render_cl.c
+++ b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_render_cl.c
@@ -95,7 +95,7 @@ static uint32_t vc4_full_res_offset(struct vc4_exec_info *exec,
 				    uint8_t x, uint8_t y)
 {
 	return bo->paddr + surf->offset + VC4_TILE_BUFFER_SIZE *
-		(DIV_ROUND_UP(exec->args->width, 32) * y + x);
+		(DIV_ROUND_UP(exec->args->width, exec->tile_width) * y + x);
 }
 
 /*
@@ -369,7 +369,8 @@ static int vc4_full_res_bounds_check(struct vc4_exec_info *exec,
 				     struct drm_vc4_submit_rcl_surface *surf)
 {
 	struct drm_vc4_submit_cl *args = exec->args;
-	u32 render_tiles_stride = DIV_ROUND_UP(exec->args->width, 32);
+	u32 render_tiles_stride = DIV_ROUND_UP(exec->args->width,
+					       exec->tile_width);
 
 	if (surf->offset > obj->base.size) {
 		DRM_ERROR("surface offset %d > BO size %zd\n",
diff --git a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_validate.c b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_validate.c
index c9e3934ab..4ef01108b 100644
--- a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_validate.c
+++ b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_validate.c
@@ -800,7 +800,7 @@ validate_gl_shader_rec(struct drm_device *dev,
 		uint32_t src_offset = *(uint32_t *)(pkt_u + o);
 		uint32_t *texture_handles_u;
 		void *uniform_data_u;
-		uint32_t tex;
+		uint32_t tex, uni;
 
 		*(uint32_t *)(pkt_v + o) = bo[i]->paddr + src_offset;
 
@@ -838,6 +838,17 @@ validate_gl_shader_rec(struct drm_device *dev,
 			}
 		}
 
+		/* Fill in the uniform slots that need this shader's
+		 * start-of-uniforms address (used for resetting the uniform
+		 * stream in the presence of control flow).
+		 */
+		for (uni = 0;
+		     uni < validated_shader->num_uniform_addr_offsets;
+		     uni++) {
+			uint32_t o = validated_shader->uniform_addr_offsets[uni];
+			((uint32_t *)exec->uniforms_v)[o] = exec->uniforms_p;
+		}
+
 		*(uint32_t *)(pkt_v + o + 4) = exec->uniforms_p;
 
 		exec->uniforms_u += validated_shader->uniforms_src_size;
diff --git a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_validate_shaders.c b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_validate_shaders.c
index 868a0ad1a..82717ca55 100644
--- a/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_validate_shaders.c
+++ b/lib/mesa/src/gallium/drivers/vc4/kernel/vc4_validate_shaders.c
@@ -40,7 +40,17 @@
 #include "vc4_qpu.h"
 #include "vc4_qpu_defines.h"
 
+#define LIVE_REG_COUNT (32 + 32 + 4)
+
 struct vc4_shader_validation_state {
+	/* Current IP being validated. */
+	uint32_t ip;
+
+	/* IP at the end of the BO, do not read shader[max_ip] */
+	uint32_t max_ip;
+
+	uint64_t *shader;
+
 	struct vc4_texture_sample_info tmu_setup[2];
 	int tmu_write_count[2];
 
@@ -50,8 +60,30 @@ struct vc4_shader_validation_state {
 	 *
 	 * This is used for the validation of direct address memory reads.
 	 */
-	uint32_t live_min_clamp_offsets[32 + 32 + 4];
-	bool live_max_clamp_regs[32 + 32 + 4];
+	uint32_t live_min_clamp_offsets[LIVE_REG_COUNT];
+	bool live_max_clamp_regs[LIVE_REG_COUNT];
+	uint32_t live_immediates[LIVE_REG_COUNT];
+
+	/* Bitfield of which IPs are used as branch targets.
+	 *
+	 * Used for validation that the uniform stream is updated at the right
+	 * points and clearing the texturing/clamping state.
+	 */
+	unsigned long *branch_targets;
+
+	/* Set when entering a basic block, and cleared when the uniform
+	 * address update is found.  This is used to make sure that we don't
+	 * read uniforms when the address is undefined.
+	 */
+	bool needs_uniform_address_update;
+
+	/* Set when we find a backwards branch.  If the branch is backwards,
+	 * the taraget is probably doing an address reset to read uniforms,
+	 * and so we need to be sure that a uniforms address is present in the
+	 * stream, even if the shader didn't need to read uniforms in later
+	 * basic blocks.
+	 */
+	bool needs_uniform_address_for_loop;
 };
 
 static uint32_t
@@ -130,11 +162,11 @@ record_texture_sample(struct vc4_validated_shader_info *validated_shader,
 }
 
 static bool
-check_tmu_write(uint64_t inst,
-		struct vc4_validated_shader_info *validated_shader,
+check_tmu_write(struct vc4_validated_shader_info *validated_shader,
 		struct vc4_shader_validation_state *validation_state,
 		bool is_mul)
 {
+	uint64_t inst = validation_state->shader[validation_state->ip];
 	uint32_t waddr = (is_mul ?
 			  QPU_GET_FIELD(inst, QPU_WADDR_MUL) :
 			  QPU_GET_FIELD(inst, QPU_WADDR_ADD));
@@ -163,7 +195,7 @@ check_tmu_write(uint64_t inst,
 			return false;
 		}
 
-		/* We assert that the the clamped address is the first
+		/* We assert that the clamped address is the first
 		 * argument, and the UBO base address is the second argument.
 		 * This is arbitrary, but simpler than supporting flipping the
 		 * two either way.
@@ -213,8 +245,14 @@ check_tmu_write(uint64_t inst,
 	/* Since direct uses a RADDR uniform reference, it will get counted in
 	 * check_instruction_reads()
 	 */
-	if (!is_direct)
+	if (!is_direct) {
+		if (validation_state->needs_uniform_address_update) {
+			DRM_ERROR("Texturing with undefined uniform address\n");
+			return false;
+		}
+
 		validated_shader->uniforms_size += 4;
+	}
 
 	if (submit) {
 		if (!record_texture_sample(validated_shader,
@@ -228,23 +266,138 @@ check_tmu_write(uint64_t inst,
 	return true;
 }
 
+static bool require_uniform_address_uniform(struct vc4_validated_shader_info *validated_shader)
+{
+	uint32_t o = validated_shader->num_uniform_addr_offsets;
+	uint32_t num_uniforms = validated_shader->uniforms_size / 4;
+
+	validated_shader->uniform_addr_offsets =
+		krealloc(validated_shader->uniform_addr_offsets,
+			 (o + 1) *
+			 sizeof(*validated_shader->uniform_addr_offsets),
+			 GFP_KERNEL);
+	if (!validated_shader->uniform_addr_offsets)
+		return false;
+
+	validated_shader->uniform_addr_offsets[o] = num_uniforms;
+	validated_shader->num_uniform_addr_offsets++;
+
+	return true;
+}
+
 static bool
-check_reg_write(uint64_t inst,
-		struct vc4_validated_shader_info *validated_shader,
+validate_uniform_address_write(struct vc4_validated_shader_info *validated_shader,
+			       struct vc4_shader_validation_state *validation_state,
+			       bool is_mul)
+{
+	uint64_t inst = validation_state->shader[validation_state->ip];
+	u32 add_b = QPU_GET_FIELD(inst, QPU_ADD_B);
+	u32 raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
+	u32 raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
+	u32 add_lri = raddr_add_a_to_live_reg_index(inst);
+	/* We want our reset to be pointing at whatever uniform follows the
+	 * uniforms base address.
+	 */
+	u32 expected_offset = validated_shader->uniforms_size + 4;
+
+	/* We only support absolute uniform address changes, and we
+	 * require that they be in the current basic block before any
+	 * of its uniform reads.
+	 *
+	 * One could potentially emit more efficient QPU code, by
+	 * noticing that (say) an if statement does uniform control
+	 * flow for all threads and that the if reads the same number
+	 * of uniforms on each side.  However, this scheme is easy to
+	 * validate so it's all we allow for now.
+	 */
+
+	if (QPU_GET_FIELD(inst, QPU_SIG) != QPU_SIG_NONE) {
+		DRM_ERROR("uniforms address change must be "
+			  "normal math\n");
+		return false;
+	}
+
+	if (is_mul || QPU_GET_FIELD(inst, QPU_OP_ADD) != QPU_A_ADD) {
+		DRM_ERROR("Uniform address reset must be an ADD.\n");
+		return false;
+	}
+
+	if (QPU_GET_FIELD(inst, QPU_COND_ADD) != QPU_COND_ALWAYS) {
+		DRM_ERROR("Uniform address reset must be unconditional.\n");
+		return false;
+	}
+
+	if (QPU_GET_FIELD(inst, QPU_PACK) != QPU_PACK_A_NOP &&
+	    !(inst & QPU_PM)) {
+		DRM_ERROR("No packing allowed on uniforms reset\n");
+		return false;
+	}
+
+	if (add_lri == -1) {
+		DRM_ERROR("First argument of uniform address write must be "
+			  "an immediate value.\n");
+		return false;
+	}
+
+	if (validation_state->live_immediates[add_lri] != expected_offset) {
+		DRM_ERROR("Resetting uniforms with offset %db instead of %db\n",
+			  validation_state->live_immediates[add_lri],
+			  expected_offset);
+		return false;
+	}
+
+	if (!(add_b == QPU_MUX_A && raddr_a == QPU_R_UNIF) &&
+	    !(add_b == QPU_MUX_B && raddr_b == QPU_R_UNIF)) {
+		DRM_ERROR("Second argument of uniform address write must be "
+			  "a uniform.\n");
+		return false;
+	}
+
+	validation_state->needs_uniform_address_update = false;
+	validation_state->needs_uniform_address_for_loop = false;
+	return require_uniform_address_uniform(validated_shader);
+}
+
+static bool
+check_reg_write(struct vc4_validated_shader_info *validated_shader,
 		struct vc4_shader_validation_state *validation_state,
 		bool is_mul)
 {
+	uint64_t inst = validation_state->shader[validation_state->ip];
 	uint32_t waddr = (is_mul ?
 			  QPU_GET_FIELD(inst, QPU_WADDR_MUL) :
 			  QPU_GET_FIELD(inst, QPU_WADDR_ADD));
+	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
+	bool ws = inst & QPU_WS;
+	bool is_b = is_mul ^ ws;
+	u32 lri = waddr_to_live_reg_index(waddr, is_b);
+
+	if (lri != -1) {
+		uint32_t cond_add = QPU_GET_FIELD(inst, QPU_COND_ADD);
+		uint32_t cond_mul = QPU_GET_FIELD(inst, QPU_COND_MUL);
+
+		if (sig == QPU_SIG_LOAD_IMM &&
+		    QPU_GET_FIELD(inst, QPU_PACK) == QPU_PACK_A_NOP &&
+		    ((is_mul && cond_mul == QPU_COND_ALWAYS) ||
+		     (!is_mul && cond_add == QPU_COND_ALWAYS))) {
+			validation_state->live_immediates[lri] =
+				QPU_GET_FIELD(inst, QPU_LOAD_IMM);
+		} else {
+			validation_state->live_immediates[lri] = ~0;
+		}
+	}
 
 	switch (waddr) {
 	case QPU_W_UNIFORMS_ADDRESS:
-		/* XXX: We'll probably need to support this for reladdr, but
-		 * it's definitely a security-related one.
-		 */
-		DRM_ERROR("uniforms address load unsupported\n");
-		return false;
+		if (is_b) {
+			DRM_ERROR("relative uniforms address change "
+				  "unsupported\n");
+			return false;
+		}
+
+		return validate_uniform_address_write(validated_shader,
+						      validation_state,
+						      is_mul);
 
 	case QPU_W_TLB_COLOR_MS:
 	case QPU_W_TLB_COLOR_ALL:
@@ -262,7 +415,7 @@ check_reg_write(uint64_t inst,
 	case QPU_W_TMU1_T:
 	case QPU_W_TMU1_R:
 	case QPU_W_TMU1_B:
-		return check_tmu_write(inst, validated_shader, validation_state,
+		return check_tmu_write(validated_shader, validation_state,
 				       is_mul);
 
 	case QPU_W_HOST_INT:
@@ -295,10 +448,10 @@ check_reg_write(uint64_t inst,
 }
 
 static void
-track_live_clamps(uint64_t inst,
-		  struct vc4_validated_shader_info *validated_shader,
+track_live_clamps(struct vc4_validated_shader_info *validated_shader,
 		  struct vc4_shader_validation_state *validation_state)
 {
+	uint64_t inst = validation_state->shader[validation_state->ip];
 	uint32_t op_add = QPU_GET_FIELD(inst, QPU_OP_ADD);
 	uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
 	uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
@@ -370,10 +523,10 @@ track_live_clamps(uint64_t inst,
 }
 
 static bool
-check_instruction_writes(uint64_t inst,
-			 struct vc4_validated_shader_info *validated_shader,
+check_instruction_writes(struct vc4_validated_shader_info *validated_shader,
 			 struct vc4_shader_validation_state *validation_state)
 {
+	uint64_t inst = validation_state->shader[validation_state->ip];
 	uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
 	uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
 	bool ok;
@@ -383,20 +536,44 @@ check_instruction_writes(uint64_t inst,
 		return false;
 	}
 
-	ok = (check_reg_write(inst, validated_shader, validation_state,
-			      false) &&
-	      check_reg_write(inst, validated_shader, validation_state,
-			      true));
+	ok = (check_reg_write(validated_shader, validation_state, false) &&
+	      check_reg_write(validated_shader, validation_state, true));
 
-	track_live_clamps(inst, validated_shader, validation_state);
+	track_live_clamps(validated_shader, validation_state);
 
 	return ok;
 }
 
 static bool
-check_instruction_reads(uint64_t inst,
-			struct vc4_validated_shader_info *validated_shader)
+check_branch(uint64_t inst,
+	     struct vc4_validated_shader_info *validated_shader,
+	     struct vc4_shader_validation_state *validation_state,
+	     int ip)
+{
+	int32_t branch_imm = QPU_GET_FIELD(inst, QPU_BRANCH_TARGET);
+	uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
+	uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
+
+	if ((int)branch_imm < 0)
+		validation_state->needs_uniform_address_for_loop = true;
+
+	/* We don't want to have to worry about validation of this, and
+	 * there's no need for it.
+	 */
+	if (waddr_add != QPU_W_NOP || waddr_mul != QPU_W_NOP) {
+		DRM_ERROR("branch instruction at %d wrote a register.\n",
+			  validation_state->ip);
+		return false;
+	}
+
+	return true;
+}
+
+static bool
+check_instruction_reads(struct vc4_validated_shader_info *validated_shader,
+			struct vc4_shader_validation_state *validation_state)
 {
+	uint64_t inst = validation_state->shader[validation_state->ip];
 	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
 	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
 	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
@@ -408,40 +585,204 @@ check_instruction_reads(uint64_t inst,
 		 * already be OOM.
 		 */
 		validated_shader->uniforms_size += 4;
+
+		if (validation_state->needs_uniform_address_update) {
+			DRM_ERROR("Uniform read with undefined uniform "
+				  "address\n");
+			return false;
+		}
 	}
 
 	return true;
 }
 
+/* Make sure that all branches are absolute and point within the shader, and
+ * note their targets for later.
+ */
+static bool
+vc4_validate_branches(struct vc4_shader_validation_state *validation_state)
+{
+	uint32_t max_branch_target = 0;
+	bool found_shader_end = false;
+	int ip;
+	int shader_end_ip = 0;
+	int last_branch = -2;
+
+	for (ip = 0; ip < validation_state->max_ip; ip++) {
+		uint64_t inst = validation_state->shader[ip];
+		int32_t branch_imm = QPU_GET_FIELD(inst, QPU_BRANCH_TARGET);
+		uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
+		uint32_t after_delay_ip = ip + 4;
+		uint32_t branch_target_ip;
+
+		if (sig == QPU_SIG_PROG_END) {
+			shader_end_ip = ip;
+			found_shader_end = true;
+			continue;
+		}
+
+		if (sig != QPU_SIG_BRANCH)
+			continue;
+
+		if (ip - last_branch < 4) {
+			DRM_ERROR("Branch at %d during delay slots\n", ip);
+			return false;
+		}
+		last_branch = ip;
+
+		if (inst & QPU_BRANCH_REG) {
+			DRM_ERROR("branching from register relative "
+				  "not supported\n");
+			return false;
+		}
+
+		if (!(inst & QPU_BRANCH_REL)) {
+			DRM_ERROR("relative branching required\n");
+			return false;
+		}
+
+		/* The actual branch target is the instruction after the delay
+		 * slots, plus whatever byte offset is in the low 32 bits of
+		 * the instruction.  Make sure we're not branching beyond the
+		 * end of the shader object.
+		 */
+		if (branch_imm % sizeof(inst) != 0) {
+			DRM_ERROR("branch target not aligned\n");
+			return false;
+		};
+
+		branch_target_ip = after_delay_ip + (branch_imm >> 3);
+		if (branch_target_ip >= validation_state->max_ip) {
+			DRM_ERROR("Branch at %d outside of shader (ip %d/%d)\n",
+				  ip, branch_target_ip,
+				  validation_state->max_ip);
+			return false;
+		}
+		set_bit(branch_target_ip, validation_state->branch_targets);
+
+		/* Make sure that the non-branching path is also not outside
+		 * the shader.
+		 */
+		if (after_delay_ip >= validation_state->max_ip) {
+			DRM_ERROR("Branch at %d continues past shader end "
+				  "(%d/%d)\n",
+				  ip, after_delay_ip, validation_state->max_ip);
+			return false;
+		}
+		set_bit(after_delay_ip, validation_state->branch_targets);
+		max_branch_target = max(max_branch_target, after_delay_ip);
+
+		/* There are two delay slots after program end is signaled
+		 * that are still executed, then we're finished.
+		 */
+		if (found_shader_end && ip == shader_end_ip + 2)
+			break;
+	}
+
+	if (max_branch_target > shader_end_ip) {
+		DRM_ERROR("Branch landed after QPU_SIG_PROG_END");
+		return false;
+	}
+
+	return true;
+}
+
+/* Resets any known state for the shader, used when we may be branched to from
+ * multiple locations in the program (or at shader start).
+ */
+static void
+reset_validation_state(struct vc4_shader_validation_state *validation_state)
+{
+	int i;
+
+	for (i = 0; i < 8; i++)
+		validation_state->tmu_setup[i / 4].p_offset[i % 4] = ~0;
+
+	for (i = 0; i < LIVE_REG_COUNT; i++) {
+		validation_state->live_min_clamp_offsets[i] = ~0;
+		validation_state->live_max_clamp_regs[i] = false;
+		validation_state->live_immediates[i] = ~0;
+	}
+}
+
+static bool
+texturing_in_progress(struct vc4_shader_validation_state *validation_state)
+{
+	return (validation_state->tmu_write_count[0] != 0 ||
+		validation_state->tmu_write_count[1] != 0);
+}
+
+static bool
+vc4_handle_branch_target(struct vc4_shader_validation_state *validation_state)
+{
+	uint32_t ip = validation_state->ip;
+
+	if (!test_bit(ip, validation_state->branch_targets))
+		return true;
+
+	if (texturing_in_progress(validation_state)) {
+		DRM_ERROR("Branch target landed during TMU setup\n");
+		return false;
+	}
+
+	/* Reset our live values tracking, since this instruction may have
+	 * multiple predecessors.
+	 *
+	 * One could potentially do analysis to determine that, for
+	 * example, all predecessors have a live max clamp in the same
+	 * register, but we don't bother with that.
+	 */
+	reset_validation_state(validation_state);
+
+	/* Since we've entered a basic block from potentially multiple
+	 * predecessors, we need the uniforms address to be updated before any
+	 * unforms are read.  We require that after any branch point, the next
+	 * uniform to be loaded is a uniform address offset.  That uniform's
+	 * offset will be marked by the uniform address register write
+	 * validation, or a one-off the end-of-program check.
+	 */
+	validation_state->needs_uniform_address_update = true;
+
+	return true;
+}
+
 struct vc4_validated_shader_info *
 vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 {
 	bool found_shader_end = false;
 	int shader_end_ip = 0;
-	uint32_t ip, max_ip;
-	uint64_t *shader;
-	struct vc4_validated_shader_info *validated_shader;
+	uint32_t ip;
+	struct vc4_validated_shader_info *validated_shader = NULL;
 	struct vc4_shader_validation_state validation_state;
-	int i;
 
 	memset(&validation_state, 0, sizeof(validation_state));
+	validation_state.shader = shader_obj->vaddr;
+	validation_state.max_ip = shader_obj->base.size / sizeof(uint64_t);
 
-	for (i = 0; i < 8; i++)
-		validation_state.tmu_setup[i / 4].p_offset[i % 4] = ~0;
-	for (i = 0; i < ARRAY_SIZE(validation_state.live_min_clamp_offsets); i++)
-		validation_state.live_min_clamp_offsets[i] = ~0;
+	reset_validation_state(&validation_state);
 
-	shader = shader_obj->vaddr;
-	max_ip = shader_obj->base.size / sizeof(uint64_t);
+	validation_state.branch_targets =
+		kcalloc(BITS_TO_LONGS(validation_state.max_ip),
+			sizeof(unsigned long), GFP_KERNEL);
+	if (!validation_state.branch_targets)
+		goto fail;
 
 	validated_shader = kcalloc(1, sizeof(*validated_shader), GFP_KERNEL);
 	if (!validated_shader)
-		return NULL;
+		goto fail;
+
+	if (!vc4_validate_branches(&validation_state))
+		goto fail;
 
-	for (ip = 0; ip < max_ip; ip++) {
-		uint64_t inst = shader[ip];
+	for (ip = 0; ip < validation_state.max_ip; ip++) {
+		uint64_t inst = validation_state.shader[ip];
 		uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
 
+		validation_state.ip = ip;
+
+		if (!vc4_handle_branch_target(&validation_state))
+			goto fail;
+
 		switch (sig) {
 		case QPU_SIG_NONE:
 		case QPU_SIG_WAIT_FOR_SCOREBOARD:
@@ -451,13 +792,14 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 		case QPU_SIG_LOAD_TMU1:
 		case QPU_SIG_PROG_END:
 		case QPU_SIG_SMALL_IMM:
-			if (!check_instruction_writes(inst, validated_shader,
+			if (!check_instruction_writes(validated_shader,
 						      &validation_state)) {
 				DRM_ERROR("Bad write at ip %d\n", ip);
 				goto fail;
 			}
 
-			if (!check_instruction_reads(inst, validated_shader))
+			if (!check_instruction_reads(validated_shader,
+						     &validation_state))
 				goto fail;
 
 			if (sig == QPU_SIG_PROG_END) {
@@ -468,13 +810,18 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 			break;
 
 		case QPU_SIG_LOAD_IMM:
-			if (!check_instruction_writes(inst, validated_shader,
+			if (!check_instruction_writes(validated_shader,
 						      &validation_state)) {
 				DRM_ERROR("Bad LOAD_IMM write at ip %d\n", ip);
 				goto fail;
 			}
 			break;
 
+		case QPU_SIG_BRANCH:
+			if (!check_branch(inst, validated_shader,
+					  &validation_state, ip))
+				goto fail;
+			break;
 		default:
 			DRM_ERROR("Unsupported QPU signal %d at "
 				  "instruction %d\n", sig, ip);
@@ -488,13 +835,28 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 			break;
 	}
 
-	if (ip == max_ip) {
+	if (ip == validation_state.max_ip) {
 		DRM_ERROR("shader failed to terminate before "
 			  "shader BO end at %zd\n",
 			  shader_obj->base.size);
 		goto fail;
 	}
 
+	/* If we did a backwards branch and we haven't emitted a uniforms
+	 * reset since then, we still need the uniforms stream to have the
+	 * uniforms address available so that the backwards branch can do its
+	 * uniforms reset.
+	 *
+	 * We could potentially prove that the backwards branch doesn't
+	 * contain any uses of uniforms until program exit, but that doesn't
+	 * seem to be worth the trouble.
+	 */
+	if (validation_state.needs_uniform_address_for_loop) {
+		if (!require_uniform_address_uniform(validated_shader))
+			goto fail;
+		validated_shader->uniforms_size += 4;
+	}
+
 	/* Again, no chance of integer overflow here because the worst case
 	 * scenario is 8 bytes of uniforms plus handles per 8-byte
 	 * instruction.
@@ -503,9 +865,12 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 		(validated_shader->uniforms_size +
 		 4 * validated_shader->num_texture_samples);
 
+	kfree(validation_state.branch_targets);
+
 	return validated_shader;
 
 fail:
+	kfree(validation_state.branch_targets);
 	if (validated_shader) {
 		kfree(validated_shader->texture_samples);
 		kfree(validated_shader);