3 files changed, 13 insertions, 4 deletions
diff --git a/src/video_core/shader/shader.cpp b/src/video_core/shader/shader.cpp
index c860375a1..67ed19ba8 100644
--- a/src/video_core/shader/shader.cpp
+++ b/src/video_core/shader/shader.cpp
@@ -8,7 +8,8 @@
 #include "common/logging/log.h"
 #include "common/microprofile.h"
 #include "video_core/pica_state.h"
-#include "video_core/regs.h"
+#include "video_core/regs_rasterizer.h"
+#include "video_core/regs_shader.h"
 #include "video_core/shader/shader.h"
 #include "video_core/shader/shader_interpreter.h"
 #ifdef ARCHITECTURE_x86_64
@@ -39,9 +40,8 @@ OutputVertex OutputVertex::FromAttributeBuffer(const RasterizerRegs& regs, Attri
 
         for (unsigned comp = 0; comp < 4; ++comp) {
             RasterizerRegs::VSOutputAttributes::Semantic semantic = semantics[comp];
-            float24* out = &vertex_slots[semantic];
             if (semantic < vertex_slots.size()) {
-                *out = input.attr[i][comp];
+                vertex_slots[semantic] = input.attr[i][comp];
             } else if (semantic != RasterizerRegs::VSOutputAttributes::INVALID) {
                 LOG_ERROR(HW_GPU, "Invalid/unknown semantic id: %u", (unsigned int)semantic);
             }
diff --git a/src/video_core/shader/shader.h b/src/video_core/shader/shader.h
index d52682479..38ea717ab 100644
--- a/src/video_core/shader/shader.h
+++ b/src/video_core/shader/shader.h
@@ -13,7 +13,8 @@
 #include "common/common_types.h"
 #include "common/vector_math.h"
 #include "video_core/pica_types.h"
-#include "video_core/regs.h"
+#include "video_core/regs_rasterizer.h"
+#include "video_core/regs_shader.h"
 
 using nihstro::RegisterType;
 using nihstro::SourceRegister;
diff --git a/src/video_core/shader/shader_jit_x64_compiler.cpp b/src/video_core/shader/shader_jit_x64_compiler.cpp
index 92b35dbc0..2dbc8b147 100644
--- a/src/video_core/shader/shader_jit_x64_compiler.cpp
+++ b/src/video_core/shader/shader_jit_x64_compiler.cpp
@@ -295,14 +295,22 @@ void JitShader::Compile_DestEnable(Instruction instr, Xmm src) {
 }
 
 void JitShader::Compile_SanitizedMul(Xmm src1, Xmm src2, Xmm scratch) {
+    // 0 * inf and inf * 0 in the PICA should return 0 instead of NaN. This can be implemented by
+    // checking for NaNs before and after the multiplication.  If the multiplication result is NaN
+    // where neither source was, this NaN was generated by a 0 * inf multiplication, and so the
+    // result should be transformed to 0 to match PICA fp rules.
+
+    // Set scratch to mask of (src1 != NaN and src2 != NaN)
     movaps(scratch, src1);
     cmpordps(scratch, src2);
 
     mulps(src1, src2);
 
+    // Set src2 to mask of (result == NaN)
     movaps(src2, src1);
     cmpunordps(src2, src2);
 
+    // Clear components where scratch != src2 (i.e. if result is NaN where neither source was NaN)
     xorps(scratch, src2);
     andps(src1, scratch);
 }