From a12072868d3ff71cfa7d124d1e0984d5723a00b4 Mon Sep 17 00:00:00 2001
From: Gert Wollny <gert.wollny@collabora.com>
Date: Tue, 30 May 2023 11:46:10 +0200
Subject: [PATCH] r600/sfn: assert that group barrier is not emitted in
 divergent code flow

Also rename emit_barrier to emit_group_barrier

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23272>
---
 src/gallium/drivers/r600/sfn/sfn_shader.cpp | 33 ++++++++++++++++++++-
 src/gallium/drivers/r600/sfn/sfn_shader.h   |  3 +-
 2 files changed, 34 insertions(+), 2 deletions(-)

diff --git a/src/gallium/drivers/r600/sfn/sfn_shader.cpp b/src/gallium/drivers/r600/sfn/sfn_shader.cpp
index fe21c03ccc1..fabb9594c10 100644
--- a/src/gallium/drivers/r600/sfn/sfn_shader.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_shader.cpp
@@ -724,6 +724,29 @@ child_block_empty(const exec_list& list)
    return result;
 }
 
+static bool value_has_non_const_source(VirtualValue *value)
+{
+   auto reg = value->as_register();
+   if (reg) {
+      // Non-ssa registers are probably the result of some control flow
+      // that makes the values non-uniform across the work group
+      if (!reg->has_flag(Register::ssa))
+         return true;
+
+      for (const auto& p : reg->parents()) {
+         auto alu = p->as_alu();
+         if (alu) {
+            for (auto& s : p->as_alu()->sources()) {
+               return value_has_non_const_source(s);
+            }
+         } else {
+            return true;
+         }
+      }
+   }
+   return false;
+}
+
 bool
 Shader::process_if(nir_if *if_stmt)
 {
@@ -731,6 +754,8 @@ Shader::process_if(nir_if *if_stmt)
 
    auto value = value_factory().src(if_stmt->condition, 0);
 
+   bool non_const_cond = value_has_non_const_source(value);
+
    EAluOp op = child_block_empty(if_stmt->then_list) ? op2_prede_int :
                                                        op2_pred_setne_int;
 
@@ -745,6 +770,8 @@ Shader::process_if(nir_if *if_stmt)
 
    IfInstr *ir = new IfInstr(pred);
    emit_instruction(ir);
+   if (non_const_cond)
+      ++m_control_flow_depth;
    start_new_block(1);
 
    if (!child_block_empty(if_stmt->then_list)) {
@@ -775,6 +802,9 @@ Shader::process_if(nir_if *if_stmt)
    if (!emit_control_flow(ControlFlowInstr::cf_endif))
       return false;
 
+   if (non_const_cond)
+      --m_control_flow_depth;
+
    return true;
 }
 
@@ -1268,8 +1298,9 @@ Shader::emit_shader_clock(nir_intrinsic_instr *instr)
 }
 
 bool
-Shader::emit_barrier(nir_intrinsic_instr *intr)
+Shader::emit_group_barrier(nir_intrinsic_instr *intr)
 {
+   assert(m_control_flow_depth == 0);
    (void)intr;
    /* Put barrier into it's own block, so that optimizers and the
     * scheduler don't move code */
diff --git a/src/gallium/drivers/r600/sfn/sfn_shader.h b/src/gallium/drivers/r600/sfn/sfn_shader.h
index 8969a2b05f2..135a1179418 100644
--- a/src/gallium/drivers/r600/sfn/sfn_shader.h
+++ b/src/gallium/drivers/r600/sfn/sfn_shader.h
@@ -317,7 +317,7 @@ private:
    bool emit_local_store(nir_intrinsic_instr *intr);
    bool emit_local_load(nir_intrinsic_instr *instr);
    bool emit_load_tcs_param_base(nir_intrinsic_instr *instr, int offset);
-   bool emit_barrier(nir_intrinsic_instr *intr);
+   bool emit_group_barrier(nir_intrinsic_instr *intr);
    bool emit_shader_clock(nir_intrinsic_instr *instr);
    bool emit_wait_ack();
 
@@ -396,6 +396,7 @@ private:
 
    InstructionChain m_chain_instr;
    std::list<Instr *, Allocator<Instr *>> m_loops;
+   int m_control_flow_depth{0};
 };
 
 std::pair<unsigned, unsigned>