"Merge Tagged PR 12066"

src/shader_recompiler/CMakeLists.txt

@@ -231,6 +231,7 @@ add_library(shader_recompiler STATIC
     ir_opt/rescaling_pass.cpp
     ir_opt/ssa_rewrite_pass.cpp
     ir_opt/texture_pass.cpp
+    ir_opt/vendor_workaround_pass.cpp
     ir_opt/verification_pass.cpp
     object_pool.h
     precompiled_headers.h

src/shader_recompiler/frontend/maxwell/translate_program.cpp

@@ -310,6 +310,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
     }
     Optimization::CollectShaderInfoPass(env, program);
     Optimization::LayerPass(program, host_info);
+    Optimization::VendorWorkaroundPass(program);
 
     CollectInterpolationInfo(env, program);
     AddNVNStorageBuffers(program);

src/shader_recompiler/ir_opt/passes.h

@@ -26,6 +26,7 @@ void SsaRewritePass(IR::Program& program);
 void PositionPass(Environment& env, IR::Program& program);
 void TexturePass(Environment& env, IR::Program& program, const HostTranslateInfo& host_info);
 void LayerPass(IR::Program& program, const HostTranslateInfo& host_info);
+void VendorWorkaroundPass(IR::Program& program);
 void VerificationPass(const IR::Program& program);
 
 // Dual Vertex

src/shader_recompiler/ir_opt/vendor_workaround_pass.cpp

@@ -0,0 +1,79 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "shader_recompiler/frontend/ir/basic_block.h"
+#include "shader_recompiler/frontend/ir/ir_emitter.h"
+#include "shader_recompiler/frontend/ir/value.h"
+#include "shader_recompiler/ir_opt/passes.h"
+
+namespace Shader::Optimization {
+
+namespace {
+void AddingByteSwapsWorkaround(IR::Block& block, IR::Inst& inst) {
+    /*
+     * Workaround for an NVIDIA bug seen in Super Mario RPG
+     *
+     * We are looking for this pattern:
+     *   %lhs_bfe = BitFieldUExtract %factor_a, #0, #16
+     *   %lhs_mul = IMul32 %lhs_bfe, %factor_b           // potentially optional?
+     *   %lhs_shl = ShiftLeftLogical32 %lhs_mul, #16
+     *   %rhs_bfe = BitFieldUExtract %factor_a, #16, #16
+     *   %result  = IAdd32 %lhs_shl, %rhs_bfe
+     *
+     * And replacing the IAdd32 with a BitwiseOr32
+     *   %result  = BitwiseOr32 %lhs_shl, %rhs_bfe
+     *
+     */
+    IR::Inst* const lhs_shl{inst.Arg(0).TryInstRecursive()};
+    IR::Inst* const rhs_bfe{inst.Arg(1).TryInstRecursive()};
+    if (!lhs_shl || !rhs_bfe) {
+        return;
+    }
+    if (lhs_shl->GetOpcode() != IR::Opcode::ShiftLeftLogical32 ||
+        lhs_shl->Arg(1) != IR::Value{16U}) {
+        return;
+    }
+    if (rhs_bfe->GetOpcode() != IR::Opcode::BitFieldUExtract || rhs_bfe->Arg(1) != IR::Value{16U} ||
+        rhs_bfe->Arg(2) != IR::Value{16U}) {
+        return;
+    }
+    IR::Inst* const lhs_mul{lhs_shl->Arg(0).TryInstRecursive()};
+    if (!lhs_mul) {
+        return;
+    }
+    const bool lhs_mul_optional{lhs_mul->GetOpcode() == IR::Opcode::BitFieldUExtract};
+    if (lhs_mul->GetOpcode() != IR::Opcode::IMul32 &&
+        lhs_mul->GetOpcode() != IR::Opcode::BitFieldUExtract) {
+        return;
+    }
+    IR::Inst* const lhs_bfe{lhs_mul_optional ? lhs_mul : lhs_mul->Arg(0).TryInstRecursive()};
+    if (!lhs_bfe) {
+        return;
+    }
+    if (lhs_bfe->GetOpcode() != IR::Opcode::BitFieldUExtract) {
+        return;
+    }
+    if (lhs_bfe->Arg(1) != IR::Value{0U} || lhs_bfe->Arg(2) != IR::Value{16U}) {
+        return;
+    }
+    IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
+    inst.ReplaceUsesWith(ir.BitwiseOr(IR::U32{inst.Arg(0)}, IR::U32{inst.Arg(1)}));
+}
+
+} // Anonymous namespace
+
+void VendorWorkaroundPass(IR::Program& program) {
+    for (IR::Block* const block : program.post_order_blocks) {
+        for (IR::Inst& inst : block->Instructions()) {
+            switch (inst.GetOpcode()) {
+            case IR::Opcode::IAdd32:
+                AddingByteSwapsWorkaround(*block, inst);
+                break;
+            default:
+                break;
+            }
+        }
+    }
+}
+
+} // namespace Shader::Optimization

src/video_core/buffer_cache/buffer_cache.h

@@ -1192,11 +1192,6 @@ void BufferCache<P>::UpdateDrawIndirect() {
             .size = static_cast<u32>(size),
             .buffer_id = FindBuffer(*cpu_addr, static_cast<u32>(size)),
         };
-        VAddr cpu_addr_start = Common::AlignDown(*cpu_addr, 64);
-        VAddr cpu_addr_end = Common::AlignUp(*cpu_addr + size, 64);
-        IntervalType interval{cpu_addr_start, cpu_addr_end};
-        ClearDownload(interval);
-        common_ranges.subtract(interval);
     };
     if (current_draw_indirect->include_count) {
         update(current_draw_indirect->count_start_address, sizeof(u32),

src/video_core/engines/maxwell_3d.cpp

@@ -268,7 +268,7 @@ size_t Maxwell3D::EstimateIndexBufferSize() {
                                                         std::numeric_limits<u32>::max()};
     const size_t byte_size = regs.index_buffer.FormatSizeInBytes();
     const size_t log2_byte_size = Common::Log2Ceil64(byte_size);
-    const size_t cap{GetMaxCurrentVertices() * 3 * byte_size};
+    const size_t cap{GetMaxCurrentVertices() * 4 * byte_size};
     const size_t lower_cap =
         std::min<size_t>(static_cast<size_t>(end_address - start_address), cap);
     return std::min<size_t>(