From 55c4a1e4cb0b9e72fbd7e8302413cc3f981e3a23 Mon Sep 17 00:00:00 2001
From: gibbed <rick@gibbed.us>
Date: Sat, 30 Nov 2019 18:45:25 -0600
Subject: [PATCH] [x64] Fix corruption of src1 in calls to
 EmulateVectorRotateLeft.

To mitigate this mistake in the future, implemented new
StashConstantXmm functions.
---
 src/xenia/cpu/backend/x64/x64_emitter.cc    | 30 +++++++++++++++++++++
 src/xenia/cpu/backend/x64/x64_emitter.h     |  3 +++
 src/xenia/cpu/backend/x64/x64_seq_vector.cc | 16 +++++------
 3 files changed, 40 insertions(+), 9 deletions(-)

diff --git a/src/xenia/cpu/backend/x64/x64_emitter.cc b/src/xenia/cpu/backend/x64/x64_emitter.cc
index 29b29b805..91de2ef9b 100644
--- a/src/xenia/cpu/backend/x64/x64_emitter.cc
+++ b/src/xenia/cpu/backend/x64/x64_emitter.cc
@@ -784,6 +784,7 @@ Xbyak::Address X64Emitter::GetXmmConstPtr(XmmConst id) {
                                      sizeof(vec128_t) * id)];
 }
 
+// Implies possible StashXmm(0, ...)!
 void X64Emitter::LoadConstantXmm(Xbyak::Xmm dest, const vec128_t& v) {
   // https://www.agner.org/optimize/optimizing_assembly.pdf
   // 13.4 Generating constants
@@ -846,6 +847,35 @@ Xbyak::Address X64Emitter::StashXmm(int index, const Xbyak::Xmm& r) {
   return addr;
 }
 
+Xbyak::Address X64Emitter::StashConstantXmm(int index, float v) {
+  union {
+    float f;
+    uint32_t i;
+  } x = {v};
+  auto addr = rsp + kStashOffset + (index * 16);
+  MovMem64(addr, x.i);
+  MovMem64(addr + 8, 0);
+  return ptr[addr];
+}
+
+Xbyak::Address X64Emitter::StashConstantXmm(int index, double v) {
+  union {
+    double d;
+    uint64_t i;
+  } x = {v};
+  auto addr = rsp + kStashOffset + (index * 16);
+  MovMem64(addr, x.i);
+  MovMem64(addr + 8, 0);
+  return ptr[addr];
+}
+
+Xbyak::Address X64Emitter::StashConstantXmm(int index, const vec128_t& v) {
+  auto addr = rsp + kStashOffset + (index * 16);
+  MovMem64(addr, v.low);
+  MovMem64(addr + 8, v.high);
+  return ptr[addr];
+}
+
 }  // namespace x64
 }  // namespace backend
 }  // namespace cpu
diff --git a/src/xenia/cpu/backend/x64/x64_emitter.h b/src/xenia/cpu/backend/x64/x64_emitter.h
index af38fb739..4f661a331 100644
--- a/src/xenia/cpu/backend/x64/x64_emitter.h
+++ b/src/xenia/cpu/backend/x64/x64_emitter.h
@@ -214,6 +214,9 @@ class X64Emitter : public Xbyak::CodeGenerator {
   void LoadConstantXmm(Xbyak::Xmm dest, double v);
   void LoadConstantXmm(Xbyak::Xmm dest, const vec128_t& v);
   Xbyak::Address StashXmm(int index, const Xbyak::Xmm& r);
+  Xbyak::Address StashConstantXmm(int index, float v);
+  Xbyak::Address StashConstantXmm(int index, double v);
+  Xbyak::Address StashConstantXmm(int index, const vec128_t& v);
 
   bool IsFeatureEnabled(uint32_t feature_flag) const {
     return (feature_flags_ & feature_flag) != 0;
diff --git a/src/xenia/cpu/backend/x64/x64_seq_vector.cc b/src/xenia/cpu/backend/x64/x64_seq_vector.cc
index dc9aa7186..8cc36eb8d 100644
--- a/src/xenia/cpu/backend/x64/x64_seq_vector.cc
+++ b/src/xenia/cpu/backend/x64/x64_seq_vector.cc
@@ -1223,26 +1223,24 @@ struct VECTOR_ROTATE_LEFT_V128
     switch (i.instr->flags) {
       case INT8_TYPE:
         // TODO(benvanik): native version (with shift magic).
-        e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
         if (i.src2.is_constant) {
-          e.LoadConstantXmm(e.xmm0, i.src2.constant());
-          e.lea(e.GetNativeParam(1), e.StashXmm(1, e.xmm0));
+          e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
         } else {
           e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
         }
+        e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
         e.CallNativeSafe(
             reinterpret_cast<void*>(EmulateVectorRotateLeft<uint8_t>));
         e.vmovaps(i.dest, e.xmm0);
         break;
       case INT16_TYPE:
         // TODO(benvanik): native version (with shift magic).
-        e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
         if (i.src2.is_constant) {
-          e.LoadConstantXmm(e.xmm0, i.src2.constant());
-          e.lea(e.GetNativeParam(1), e.StashXmm(1, e.xmm0));
+          e.lea(e.GetNativeParam(1), e.StashConstantXmm(1, i.src2.constant()));
         } else {
           e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
         }
+        e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
         e.CallNativeSafe(
             reinterpret_cast<void*>(EmulateVectorRotateLeft<uint16_t>));
         e.vmovaps(i.dest, e.xmm0);
@@ -1264,13 +1262,13 @@ struct VECTOR_ROTATE_LEFT_V128
           e.vpor(i.dest, e.xmm1);
         } else {
           // TODO(benvanik): non-AVX2 native version.
-          e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
           if (i.src2.is_constant) {
-            e.LoadConstantXmm(e.xmm0, i.src2.constant());
-            e.lea(e.GetNativeParam(1), e.StashXmm(1, e.xmm0));
+            e.lea(e.GetNativeParam(1),
+                  e.StashConstantXmm(1, i.src2.constant()));
           } else {
             e.lea(e.GetNativeParam(1), e.StashXmm(1, i.src2));
           }
+          e.lea(e.GetNativeParam(0), e.StashXmm(0, i.src1));
           e.CallNativeSafe(
               reinterpret_cast<void*>(EmulateVectorRotateLeft<uint32_t>));
           e.vmovaps(i.dest, e.xmm0);