agx: Add schedule-specialized get_sr variants

Some special registers imply scheduling constraints. We want to have a single
scheduling class per instruction in the IR, so fork off various get_sr variants
depending on what kind of SR we're reading, and validate that we use the right
kind.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>

diff --git a/src/asahi/compiler/agx_compile.c b/src/asahi/compiler/agx_compile.c
index ee98ca5..83e7ba2 100644 (file)
--- a/src/asahi/compiler/agx_compile.c
+++ b/src/asahi/compiler/agx_compile.c
@@ -1038,15 +1038,15 @@ agx_emit_intrinsic(agx_builder *b, nir_intrinsic_instr *instr)
       return agx_get_sr_to(b, dst, AGX_SR_INPUT_SAMPLE_MASK);
 
    case nir_intrinsic_load_sample_mask:
-      return agx_get_sr_to(b, dst, AGX_SR_COVERAGE_MASK);
+      return agx_get_sr_coverage_to(b, dst, AGX_SR_COVERAGE_MASK);
 
    case nir_intrinsic_load_helper_invocation:
       /* Compare special register to zero. We could lower this in NIR (letting
        * us fold in an inot) but meh?
        */
-      return agx_icmpsel_to(b, dst, agx_get_sr(b, 32, AGX_SR_IS_ACTIVE_THREAD),
-                            agx_zero(), agx_immediate(1), agx_zero(),
-                            AGX_ICOND_UEQ);
+      return agx_icmpsel_to(
+         b, dst, agx_get_sr_coverage(b, 32, AGX_SR_IS_ACTIVE_THREAD),
+         agx_zero(), agx_immediate(1), agx_zero(), AGX_ICOND_UEQ);
 
    case nir_intrinsic_load_vertex_id:
       assert(b->shader->stage == MESA_SHADER_VERTEX);

diff --git a/src/asahi/compiler/agx_opcodes.py b/src/asahi/compiler/agx_opcodes.py
index 2585809..6c81d2c 100644 (file)
--- a/src/asahi/compiler/agx_opcodes.py
+++ b/src/asahi/compiler/agx_opcodes.py
@@ -310,7 +310,9 @@ op("local_atomic",
 op("wait", (0x38, 0xFF, 2, _), dests = 0,
       can_eliminate = False, imms = [SCOREBOARD], schedule_class = "invalid")
 
-op("get_sr", (0x72, 0x7F | L, 4, _), dests = 1, imms = [SR])
+for (suffix, schedule_class) in [("", "none"), ("_coverage", "coverage")]:
+    op(f"get_sr{suffix}", (0x72, 0x7F | L, 4, _), dests = 1, imms = [SR],
+       schedule_class = schedule_class, can_reorder = schedule_class == "none")
 
 op("sample_mask", (0x7fc1, 0xffff, 6, _), dests = 0, srcs = 2,
     can_eliminate = False, schedule_class = "coverage")

diff --git a/src/asahi/compiler/agx_validate.c b/src/asahi/compiler/agx_validate.c
index 458fdfb..a19a794 100644 (file)
--- a/src/asahi/compiler/agx_validate.c
+++ b/src/asahi/compiler/agx_validate.c
@@ -208,6 +208,34 @@ agx_validate_predecessors(agx_block *block)
    return true;
 }
 
+static bool
+agx_validate_sr(const agx_instr *I)
+{
+   bool none = (I->op == AGX_OPCODE_GET_SR);
+   bool coverage = (I->op == AGX_OPCODE_GET_SR_COVERAGE);
+   bool barrier = false; /* unused so far, will be GET_SR_BARRIER */
+
+   /* Filter get_sr instructions */
+   if (!(none || coverage || barrier))
+      return true;
+
+   switch (I->sr) {
+   case AGX_SR_ACTIVE_THREAD_INDEX_IN_QUAD:
+   case AGX_SR_ACTIVE_THREAD_INDEX_IN_SUBGROUP:
+   case AGX_SR_COVERAGE_MASK:
+   case AGX_SR_IS_ACTIVE_THREAD:
+      return coverage;
+
+   case AGX_SR_OPFIFO_CMD:
+   case AGX_SR_OPFIFO_DATA_L:
+   case AGX_SR_OPFIFO_DATA_H:
+      return barrier;
+
+   default:
+      return none;
+   }
+}
+
 void
 agx_validate(agx_context *ctx, const char *after)
 {
@@ -261,6 +289,12 @@ agx_validate(agx_context *ctx, const char *after)
          agx_print_instr(I, stderr);
          fail = true;
       }
+
+      if (!agx_validate_sr(I)) {
+         fprintf(stderr, "Invalid SR after %s\n", after);
+         agx_print_instr(I, stdout);
+         fail = true;
+      }
    }
 
    if (!agx_validate_width(ctx)) {