3 files changed, 62 insertions, 1 deletions

diff --git a/src/broadcom/compiler/vir_to_qpu.c b/src/broadcom/compiler/vir_to_qpu.c
index 4baadce294c..0ab0f26c953 100644
--- a/src/broadcom/compiler/vir_to_qpu.c
+++ b/src/broadcom/compiler/vir_to_qpu.c
@@ -353,6 +353,33 @@ v3d_generate_code_block(struct v3d_compile *c,
         }
 }
 
+static bool
+reads_uniform(const struct v3d_device_info *devinfo, uint64_t instruction)
+{
+        struct v3d_qpu_instr qpu;
+        MAYBE_UNUSED bool ok = v3d_qpu_instr_unpack(devinfo, instruction, &qpu);
+        assert(ok);
+
+        if (qpu.sig.ldunif ||
+            qpu.sig.ldunifarf ||
+            qpu.sig.wrtmuc) {
+                return true;
+        }
+
+        if (qpu.type == V3D_QPU_INSTR_TYPE_ALU) {
+                if (qpu.alu.add.magic_write &&
+                    v3d_qpu_magic_waddr_loads_unif(qpu.alu.add.waddr)) {
+                        return true;
+                }
+
+                if (qpu.alu.mul.magic_write &&
+                    v3d_qpu_magic_waddr_loads_unif(qpu.alu.mul.waddr)) {
+                        return true;
+                }
+        }
+
+        return false;
+}
 
 static void
 v3d_dump_qpu(struct v3d_compile *c)
@@ -361,11 +388,30 @@ v3d_dump_qpu(struct v3d_compile *c)
                 vir_get_stage_name(c),
                 c->program_id, c->variant_id);
 
+        int next_uniform = 0;
         for (int i = 0; i < c->qpu_inst_count; i++) {
                 const char *str = v3d_qpu_disasm(c->devinfo, c->qpu_insts[i]);
-                fprintf(stderr, "0x%016"PRIx64" %s\n", c->qpu_insts[i], str);
+                fprintf(stderr, "0x%016"PRIx64" %s", c->qpu_insts[i], str);
+
+                /* We can only do this on 4.x, because we're not tracking TMU
+                 * implicit uniforms here on 3.x.
+                 */
+                if (c->devinfo->ver >= 40 &&
+                    reads_uniform(c->devinfo, c->qpu_insts[i])) {
+                        fprintf(stderr, " (");
+                        vir_dump_uniform(c->uniform_contents[next_uniform],
+                                         c->uniform_data[next_uniform]);
+                        fprintf(stderr, ")");
+                        next_uniform++;
+                }
+                fprintf(stderr, "\n");
                 ralloc_free((void *)str);
         }
+
+        /* Make sure our dumping lined up. */
+        if (c->devinfo->ver >= 40)
+                assert(next_uniform == c->num_uniforms);
+
         fprintf(stderr, "\n");
 }
 
diff --git a/src/broadcom/qpu/qpu_instr.c b/src/broadcom/qpu/qpu_instr.c
index 147017a6594..08f29a40937 100644
--- a/src/broadcom/qpu/qpu_instr.c
+++ b/src/broadcom/qpu/qpu_instr.c
@@ -555,6 +555,20 @@ v3d_qpu_magic_waddr_is_tsy(enum v3d_qpu_waddr waddr)
                 waddr == V3D_QPU_WADDR_SYNCU);
 }
 
+bool
+v3d_qpu_magic_waddr_loads_unif(enum v3d_qpu_waddr waddr)
+{
+        switch (waddr) {
+        case V3D_QPU_WADDR_VPMU:
+        case V3D_QPU_WADDR_TLBU:
+        case V3D_QPU_WADDR_TMUAU:
+        case V3D_QPU_WADDR_SYNCU:
+                return true;
+        default:
+                return false;
+        }
+}
+
 static bool
 v3d_qpu_add_op_reads_vpm(enum  v3d_qpu_add_op op)
 {
diff --git a/src/broadcom/qpu/qpu_instr.h b/src/broadcom/qpu/qpu_instr.h
index c2b4ebd1995..49d66ee888f 100644
--- a/src/broadcom/qpu/qpu_instr.h
+++ b/src/broadcom/qpu/qpu_instr.h
@@ -443,6 +443,7 @@ bool v3d_qpu_magic_waddr_is_tmu(enum v3d_qpu_waddr waddr) ATTRIBUTE_CONST;
 bool v3d_qpu_magic_waddr_is_tlb(enum v3d_qpu_waddr waddr) ATTRIBUTE_CONST;
 bool v3d_qpu_magic_waddr_is_vpm(enum v3d_qpu_waddr waddr) ATTRIBUTE_CONST;
 bool v3d_qpu_magic_waddr_is_tsy(enum v3d_qpu_waddr waddr) ATTRIBUTE_CONST;
+bool v3d_qpu_magic_waddr_loads_unif(enum v3d_qpu_waddr waddr) ATTRIBUTE_CONST;
 bool v3d_qpu_uses_tlb(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;
 bool v3d_qpu_uses_sfu(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;
 bool v3d_qpu_writes_tmu(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;