pan/bi: rework memory barriers

The MEMORY_BARRIER instruction has some issues, where we end up
dead-code eliminating it before it gets to do what it's supposed to do.

But even if we fix that, we have issues where we can end up inserting
flow control into it, which isn't going to work because we have nothing
to emit here either.

So let's rework this to a special-cased NOP instruction, which is marked
as a scheduling barrier. The beneft here is that NOPs are already properly
handled when it comes to flow control.

Note that this isn't perfect either; this only prevents memory operations
from crossing the scheduling barrier. We should really prevent any
operation with observable side effects from crossing the barrier. This
includes things like reading clocks etc.

But that's a larger change, and it's a step in the right direction to get
this to no longer be dead-code eliminated. So let's put this band-aid on
for now.

Fixes: f77a50e45e ("pan/bi: add a MEMORY_BARRIER pseudo-instruction")
Reviewed-by: Caterina Shablia <caterina.shablia@collabora.com>
Reviewed-by: Olivia Lee <olivia.lee@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Reviewed-by: Mary Guillemard <mary.guillemard@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35502>

src/panfrost/compiler/IR_pseudo.xml

@@ -191,6 +191,4 @@
     <src start="0"/>
   </ins>
 
-  <ins name="MEMORY_BARRIER" pseudo="true" unit="NONE"/>
-
 </bifrost>

src/panfrost/compiler/bi_pressure_schedule.c

@@ -154,8 +154,13 @@ create_dag(bi_context *ctx, bi_block *block, void *memctx)
          coverage = node;
       }
       if (I->op == BI_OPCODE_DISCARD_F32 ||
-          I->op == BI_OPCODE_MEMORY_BARRIER) {
-         /* Serialize against memory operations and barriers */
+          bi_is_scheduling_barrier(I)) {
+         /* Serialize against memory operations and barriers.
+          *
+          * TODO: This is *not* quite sufficient in the case of a scheduling
+          * barrier. We need to serialize against *all* operations with
+          * side-effects in that case.
+          */
          add_dep(node, memory_load);
          add_dep(node, memory_store);
          memory_load = node;

src/panfrost/compiler/bifrost_compile.c

@@ -2105,7 +2105,7 @@ bi_emit_intrinsic(bi_builder *b, nir_intrinsic_instr *instr)
           * scheduler from reordering memory operations around the barrier.
           * Avail and vis are trivially established.
           */
-         bi_memory_barrier(b);
+         bi_nop(b)->scheduling_barrier = true;
          break;
 
       case SCOPE_WORKGROUP:

src/panfrost/compiler/bir.c

@@ -188,6 +188,9 @@ bi_side_effects(const bi_instr *I)
    if (bi_get_opcode_props(I)->last)
       return true;
 
+   if (bi_is_scheduling_barrier(I))
+      return true;
+
    switch (I->op) {
    case BI_OPCODE_DISCARD_F32:
    case BI_OPCODE_DISCARD_B32:

src/panfrost/compiler/compiler.h

@@ -520,6 +520,7 @@ typedef struct {
       bool combine;              /* BRANCHC */
       bool format;               /* LEA_TEX */
       bool z_stencil;            /* LD_TILE */
+      bool scheduling_barrier;   /* NOP */
 
       struct {
          enum bi_special special;   /* FADD_RSCALE, FMA_RSCALE */
@@ -633,6 +634,12 @@ bi_is_staging_src(const bi_instr *I, unsigned s)
    return (s == 0 || s == 4) && bi_get_opcode_props(I)->sr_read;
 }
 
+static inline bool
+bi_is_scheduling_barrier(const bi_instr *I)
+{
+   return I->op == BI_OPCODE_NOP && I->scheduling_barrier;
+}
+
 /*
  * Safe helpers to remove destinations/sources at the end of the
  * destination/source array when changing opcodes. Unlike adding