panvk/utrace: Add support for storing registers

Add the ability to store specific register values during tracing by
using the indirect capture capability.

Reviewed-by: Christoph Pillmayer <christoph.pillmayer@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36519>

src/panfrost/vulkan/csf/panvk_vX_utrace.c

@@ -16,7 +16,10 @@ static void
 cmd_write_timestamp(const struct panvk_device *dev, struct cs_builder *b,
                     uint64_t addr, struct cs_async_op ts_async_op)
 {
-   const struct cs_index addr_reg = cs_scratch_reg64(b, 0);
+   /* Unless we dedicate a register, this will potentially overwrite something
+    * during begin/end. */
+   const struct cs_index addr_reg =
+      cs_scratch_reg64(b, CS_REG_SCRATCH_COUNT - 2);
 
    /* Overwrite the signal_slot. Note that this has no effect in case of
     * synchronous or indirect syncs. */
@@ -74,6 +77,44 @@ cmd_copy_data(struct cs_builder *b, uint64_t dst_addr, uint64_t src_addr,
    }
 }
 
+static void
+cmd_store_regs(struct cs_builder *b, uint64_t dst_addr, uint64_t src_addr,
+               uint32_t size, bool wait_for_timestamp)
+{
+   assert((dst_addr | size) % sizeof(uint32_t) == 0);
+   uint32_t num_regs = size / sizeof(uint32_t);
+   assert(num_regs <= CS_MAX_REG_TUPLE_SIZE);
+   assert(src_addr + num_regs <= PANVK_CS_REG_SCRATCH_END);
+
+   if (wait_for_timestamp)
+      cs_wait_slot(b, SB_ID(DEFERRED_SYNC));
+
+   bool valid_dst_regs = false;
+   struct cs_index dst_addr_reg;
+   /* Unless we dedicate a register, this will potentially overwrite
+    * something during indirect capture. For now, we only ensure we don't
+    * corrupt the registers we're capturing. */
+   for (uint32_t dst_scratch_base = PANVK_CS_REG_SCRATCH_END - 1;
+        dst_scratch_base >= PANVK_CS_REG_SCRATCH_START; dst_scratch_base -= 2) {
+      if (src_addr + num_regs <= dst_scratch_base ||
+          dst_scratch_base + 2 <= src_addr) {
+         dst_addr_reg =
+            cs_scratch_reg64(b, dst_scratch_base - PANVK_CS_REG_SCRATCH_START);
+         valid_dst_regs = true;
+         break;
+      }
+   }
+
+   if (!valid_dst_regs) {
+      assert(!"No unused scratch registers found");
+      return;
+   }
+
+   const struct cs_index src_addr_reg = cs_reg_tuple(b, src_addr, num_regs);
+   cs_move64_to(b, dst_addr_reg, dst_addr);
+   cs_store(b, src_addr_reg, dst_addr_reg, BITFIELD_MASK(num_regs), 0);
+}
+
 static struct cs_builder *
 get_builder(struct panvk_cmd_buffer *cmdbuf, struct u_trace *ut)
 {
@@ -112,11 +153,16 @@ panvk_utrace_capture_data(struct u_trace *ut, void *cs, void *dst_buffer,
    const uint64_t dst_addr = dst_bo->addr.dev + dst_offset_B;
    const uint64_t src_addr = src_offset_B;
 
-   /* src_offset_B is absolute */
-   assert(!src_buffer);
+   /* src_offset_B is absolute, src_buffer is used to indicate register capture */
+   assert(!src_buffer ||
+          (uintptr_t)src_buffer == PANVK_UTRACE_CAPTURE_REGISTERS);
 
-   cmd_copy_data(b, dst_addr, src_addr, size_B,
-                 cs_info->capture_data_wait_for_ts);
+   if ((uintptr_t)src_buffer == PANVK_UTRACE_CAPTURE_REGISTERS)
+      cmd_store_regs(b, dst_addr, src_addr, size_B,
+                     cs_info->capture_data_wait_for_ts);
+   else
+      cmd_copy_data(b, dst_addr, src_addr, size_B,
+                    cs_info->capture_data_wait_for_ts);
 }
 
 void

src/panfrost/vulkan/panvk_utrace.h

@@ -49,6 +49,10 @@ struct panvk_utrace_cs_info {
    bool capture_data_wait_for_ts;
 };
 
+/* Special value indicating that an indirect capture should be of registers
+ * rather than an address. */
+#define PANVK_UTRACE_CAPTURE_REGISTERS 0x1
+
 void panvk_per_arch(utrace_context_init)(struct panvk_device *dev);
 void panvk_per_arch(utrace_context_fini)(struct panvk_device *dev);