From 032625882990ae843cbdcd526f3c72087ab31888 Mon Sep 17 00:00:00 2001
From: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Date: Sat, 9 Mar 2019 19:59:56 +0100
Subject: [PATCH] [dxbc] Use SDiv instead of ShiftRightLogical for index calc
 on Nvidia

Reportedly fixes Resident Evil 2 and Devil May Cry 5 on Nvidia.
---
 src/dxbc/dxbc_compiler.cpp | 16 +++++++++++-----
 src/dxbc/dxbc_options.cpp  |  2 ++
 src/dxbc/dxbc_options.h    |  4 ++++
 3 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/src/dxbc/dxbc_compiler.cpp b/src/dxbc/dxbc_compiler.cpp
index 9ebdb291..c8dd616e 100644
--- a/src/dxbc/dxbc_compiler.cpp
+++ b/src/dxbc/dxbc_compiler.cpp
@@ -5122,9 +5122,13 @@ namespace dxvk {
     
     const uint32_t typeId = getVectorTypeId(result.type);
     
+    uint32_t offset = m_moduleInfo.options.useSdivForBufferIndex
+      ? m_module.opSDiv             (typeId, structOffset.id, m_module.consti32(4))
+      : m_module.opShiftRightLogical(typeId, structOffset.id, m_module.consti32(2));
+    
     result.id = m_module.opIAdd(typeId,
       m_module.opIMul(typeId, structId.id, m_module.consti32(structStride / 4)),
-      m_module.opShiftRightLogical(typeId, structOffset.id, m_module.consti32(2)));
+      offset);
     return result;
   }
   
@@ -5134,10 +5138,12 @@ namespace dxvk {
     DxbcRegisterValue result;
     result.type.ctype  = DxbcScalarType::Sint32;
     result.type.ccount = 1;
-    result.id = m_module.opShiftRightLogical(
-      getVectorTypeId(result.type),
-      byteOffset.id,
-      m_module.consti32(2));
+    
+    uint32_t typeId = getVectorTypeId(result.type);
+    
+    result.id = m_moduleInfo.options.useSdivForBufferIndex
+      ? m_module.opSDiv             (typeId, byteOffset.id, m_module.consti32(4))
+      : m_module.opShiftRightLogical(typeId, byteOffset.id, m_module.consti32(2));
     return result;
   }
   
diff --git a/src/dxbc/dxbc_options.cpp b/src/dxbc/dxbc_options.cpp
index 3ec58c05..cce13915 100644
--- a/src/dxbc/dxbc_options.cpp
+++ b/src/dxbc/dxbc_options.cpp
@@ -25,6 +25,8 @@ namespace dxvk {
      && (devInfo.coreSubgroup.supportedOperations & VK_SUBGROUP_FEATURE_BALLOT_BIT);
     useRawSsbo
       = (devInfo.core.properties.limits.minStorageBufferOffsetAlignment <= sizeof(uint32_t));
+    useSdivForBufferIndex
+      = adapter->matchesDriver(DxvkGpuVendor::Nvidia, VK_DRIVER_ID_NVIDIA_PROPRIETARY_KHR, 0, 0);
     
     strictDivision          = options.strictDivision;
     zeroInitWorkgroupMemory = options.zeroInitWorkgroupMemory;
diff --git a/src/dxbc/dxbc_options.h b/src/dxbc/dxbc_options.h
index ef332c16..9fdf554c 100644
--- a/src/dxbc/dxbc_options.h
+++ b/src/dxbc/dxbc_options.h
@@ -24,6 +24,10 @@ namespace dxvk {
     /// Use SSBOs instead of texel buffers
     /// for raw and structured buffers.
     bool useRawSsbo = false;
+    
+    /// Use SDiv instead of SHR to converte byte offsets to
+    /// dword offsets. Fixes RE2 and DMC5 on Nvidia drivers.
+    bool useSdivForBufferIndex = false;
 
     /// Enables sm4-compliant division-by-zero behaviour
     bool strictDivision = false;