ir_passes: Integrate DS barriers in block

shadps4-emu · Jan 2, 2025 · 827ea02 · 827ea02
1 parent c977011
commit 827ea02
Show file tree

Hide file tree

Showing 3 changed files with 52 additions and 26 deletions.
diff --git a/src/shader_recompiler/frontend/translate/data_share.cpp b/src/shader_recompiler/frontend/translate/data_share.cpp
@@ -205,7 +205,6 @@ void Translator::DS_WRITE(int bit_size, bool is_signed, bool is_pair, bool strid
             addr, ir.Imm32((u32(inst.control.ds.offset1) << 8u) + u32(inst.control.ds.offset0)));
         ir.WriteShared(bit_size, ir.GetVectorReg(data0), addr0);
     }
-    emit_ds_read_barrier = true;
 }
 
 void Translator::DS_SWIZZLE_B32(const GcnInst& inst) {
@@ -222,11 +221,6 @@ void Translator::DS_SWIZZLE_B32(const GcnInst& inst) {
 
 void Translator::DS_READ(int bit_size, bool is_signed, bool is_pair, bool stride64,
                          const GcnInst& inst) {
-    if (emit_ds_read_barrier && profile.needs_lds_barriers) {
-        ir.Barrier();
-        emit_ds_read_barrier = false;
-    }
-
     const IR::U32 addr{ir.GetVectorReg(IR::VectorReg(inst.src[0].code))};
     IR::VectorReg dst_reg{inst.dst[0].code};
     if (is_pair) {

diff --git a/src/shader_recompiler/frontend/translate/translate.h b/src/shader_recompiler/frontend/translate/translate.h
@@ -308,7 +308,6 @@ class Translator {
     const RuntimeInfo& runtime_info;
     const Profile& profile;
     bool opcode_missing = false;
-    bool emit_ds_read_barrier = false;
 };
 
 void Translate(IR::Block* block, u32 block_base, std::span<const GcnInst> inst_list, Info& info,

diff --git a/src/shader_recompiler/ir/passes/shared_memory_barrier_pass.cpp b/src/shader_recompiler/ir/passes/shared_memory_barrier_pass.cpp
@@ -8,6 +8,54 @@
 
 namespace Shader::Optimization {
 
+static void EmitBarrierInBlock(IR::Block* block) {
+    // This is inteded to insert a barrier when shared memory write and read
+    // occur in the same basic block. Also checks if branch depth is zero as
+    // we don't want to insert barrier in potentially divergent code.
+    bool emit_barrier_on_write = false;
+    bool emit_barrier_on_read = false;
+    const auto emit_barrier = [block](bool& emit_cond, IR::Inst& inst) {
+        if (emit_cond) {
+            IR::IREmitter ir{*block, IR::Block::InstructionList::s_iterator_to(inst)};
+            ir.Barrier();
+            emit_cond = false;
+        }
+    };
+    for (IR::Inst& inst : block->Instructions()) {
+        if (inst.GetOpcode() == IR::Opcode::LoadSharedU32 ||
+            inst.GetOpcode() == IR::Opcode::LoadSharedU64) {
+            emit_barrier(emit_barrier_on_read, inst);
+            emit_barrier_on_write = true;
+        }
+        if (inst.GetOpcode() == IR::Opcode::WriteSharedU32 ||
+            inst.GetOpcode() == IR::Opcode::WriteSharedU64) {
+            emit_barrier(emit_barrier_on_write, inst);
+            emit_barrier_on_read = true;
+        }
+    }
+}
+
+static void EmitBarrierInMergeBlock(const IR::AbstractSyntaxNode::Data& data) {
+    // Insert a barrier after divergent conditional blocks.
+    // This avoids potential softlocks and crashes when some threads
+    // initialize shared memory and others read from it.
+    const IR::U1 cond = data.if_node.cond;
+    const auto insert_barrier =
+        IR::BreadthFirstSearch(cond, [](IR::Inst* inst) -> std::optional<bool> {
+            if (inst->GetOpcode() == IR::Opcode::GetAttributeU32 &&
+                inst->Arg(0).Attribute() == IR::Attribute::LocalInvocationId) {
+                return true;
+            }
+            return std::nullopt;
+        });
+    if (insert_barrier) {
+        IR::Block* const merge = data.if_node.merge;
+        auto insert_point = std::ranges::find_if_not(merge->Instructions(), IR::IsPhi);
+        IR::IREmitter ir{*merge, insert_point};
+        ir.Barrier();
+    }
+}
+
 void SharedMemoryBarrierPass(IR::Program& program, const Profile& profile) {
     if (!program.info.uses_shared || !profile.needs_lds_barriers) {
         return;
@@ -19,27 +67,12 @@ void SharedMemoryBarrierPass(IR::Program& program, const Profile& profile) {
             --branch_depth;
             continue;
         }
-        if (node.type != Type::If) {
+        if (node.type == Type::If && branch_depth++ == 0) {
+            EmitBarrierInMergeBlock(node.data);
             continue;
         }
-        u32 curr_depth = branch_depth++;
-        if (curr_depth != 0) {
-            continue;
-        }
-        const IR::U1 cond = node.data.if_node.cond;
-        const auto insert_barrier =
-            IR::BreadthFirstSearch(cond, [](IR::Inst* inst) -> std::optional<bool> {
-                if (inst->GetOpcode() == IR::Opcode::GetAttributeU32 &&
-                    inst->Arg(0).Attribute() == IR::Attribute::LocalInvocationId) {
-                    return true;
-                }
-                return std::nullopt;
-            });
-        if (insert_barrier) {
-            IR::Block* const merge = node.data.if_node.merge;
-            auto insert_point = std::ranges::find_if_not(merge->Instructions(), IR::IsPhi);
-            IR::IREmitter ir{*merge, insert_point};
-            ir.Barrier();
+        if (node.type == Type::Block && branch_depth == 0) {
+            EmitBarrierInBlock(node.data.block);
         }
     }
 }