4 files changed, 51 insertions, 20 deletions

diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_composite.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_composite.cpp
index 94dc5019d..22321f386 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_composite.cpp
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_composite.cpp
@@ -41,10 +41,23 @@ template <typename ObjectType>
 void CompositeInsert(EmitContext& ctx, IR::Inst& inst, Register composite, ObjectType object,
                      u32 index, char type) {
     const Register ret{ctx.reg_alloc.Define(inst)};
-    if (ret != composite) {
-        ctx.Add("MOV.{} {},{};", type, ret, composite);
+    const char swizzle{"xyzw"[index]};
+    if (ret != composite && ret == object) {
+        // The object is aliased with the return value, so we have to use a temporary to insert
+        ctx.Add("MOV.{} RC,{};"
+                "MOV.{} RC.{},{};"
+                "MOV.{} {},RC;",
+                type, composite, type, swizzle, object, type, ret);
+    } else if (ret != composite) {
+        // The input composite is not aliased with the return value so we have to copy it before
+        // hand. But the insert object is not aliased with the return value, so we don't have to
+        // worry about that
+        ctx.Add("MOV.{} {},{};MOV.{},{}.{},{};", type, ret, composite, type, ret, swizzle, object);
+    } else {
+        // The return value is alised so we can just insert the object, it doesn't matter if it's
+        // aliased
+        ctx.Add("MOV.{} {}.{},{};", type, ret, swizzle, object);
     }
-    ctx.Add("MOV.{} {}.{},{};", type, ret, "xyzw"[index], object);
 }
 } // Anonymous namespace
 
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_floating_point.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_floating_point.cpp
index 15db6618f..d2c324ad6 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_floating_point.cpp
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_floating_point.cpp
@@ -38,9 +38,9 @@ template <typename InputType>
 void Clamp(EmitContext& ctx, Register ret, InputType value, InputType min_value,
            InputType max_value, std::string_view type) {
     // Call MAX first to properly clamp nan to min_value instead
-    ctx.Add("MAX.{} {}.x,{},{};"
-            "MIN.{} {}.x,{}.x,{};",
-            type, ret, min_value, value, type, ret, ret, max_value);
+    ctx.Add("MAX.{} RC.x,{},{};"
+            "MIN.{} {}.x,RC.x,{};",
+            type, min_value, value, type, ret, max_value);
 }
 } // Anonymous namespace
 
@@ -159,7 +159,7 @@ void EmitFPRecipSqrt64([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] Regis
 
 void EmitFPSqrt(EmitContext& ctx, IR::Inst& inst, ScalarF32 value) {
     const Register ret{ctx.reg_alloc.Define(inst)};
-    ctx.Add("RSQ {}.x,{};RCP {}.x,{}.x;", ret, value, ret, ret);
+    ctx.Add("RSQ RC.x,{};RCP {}.x,RC.x;", value, ret);
 }
 
 void EmitFPSaturate16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] Register value) {
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp
index 2be91ccfd..15fd23356 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp
@@ -87,20 +87,38 @@ void EmitBitwiseXor32(EmitContext& ctx, IR::Inst& inst, ScalarS32 a, ScalarS32 b
 
 void EmitBitFieldInsert(EmitContext& ctx, IR::Inst& inst, ScalarS32 base, ScalarS32 insert,
                         ScalarS32 offset, ScalarS32 count) {
-    ctx.Add("MOV.U RC.x,{};MOV.U RC.y,{};", count, offset);
-    ctx.Add("BFI.S {},RC,{},{};", inst, insert, base);
+    const Register ret{ctx.reg_alloc.Define(inst)};
+    if (count.type != Type::Register && offset.type != Type::Register) {
+        ctx.Add("BFI.S {},{{{},{},0,0}},{},{};", ret, count, offset, insert, base);
+    } else {
+        ctx.Add("MOV.S RC.x,{};MOV.U RC.y,{};"
+                "BFI.S {},RC,{},{};",
+                count, offset, ret, insert, base);
+    }
 }
 
 void EmitBitFieldSExtract(EmitContext& ctx, IR::Inst& inst, ScalarS32 base, ScalarS32 offset,
                           ScalarS32 count) {
-    ctx.Add("MOV.U RC.x,{};MOV.U RC.y,{};", count, offset);
-    ctx.Add("BFE.S {},RC,{};", inst, base);
+    const Register ret{ctx.reg_alloc.Define(inst)};
+    if (count.type != Type::Register && offset.type != Type::Register) {
+        ctx.Add("BFE.S {},{{{},{},0,0}},{};", ret, count, offset, base);
+    } else {
+        ctx.Add("MOV.S RC.x,{};MOV.U RC.y,{};"
+                "BFE.S {},RC,{};",
+                count, offset, ret, base);
+    }
 }
 
 void EmitBitFieldUExtract(EmitContext& ctx, IR::Inst& inst, ScalarU32 base, ScalarU32 offset,
                           ScalarU32 count) {
-    ctx.Add("MOV.U RC.x,{};MOV.U RC.y,{};", count, offset);
-    ctx.Add("BFE.U {},RC,{};", inst, base);
+    const Register ret{ctx.reg_alloc.Define(inst)};
+    if (count.type != Type::Register && offset.type != Type::Register) {
+        ctx.Add("BFE.U {},{{{},{},0,0}},{};", ret, count, offset, base);
+    } else {
+        ctx.Add("MOV.U RC.x,{};MOV.U RC.y,{};"
+                "BFE.U {},RC,{};",
+                count, offset, ret, base);
+    }
 }
 
 void EmitBitReverse32(EmitContext& ctx, IR::Inst& inst, ScalarS32 value) {
@@ -141,16 +159,16 @@ void EmitUMax32(EmitContext& ctx, IR::Inst& inst, ScalarU32 a, ScalarU32 b) {
 
 void EmitSClamp32(EmitContext& ctx, IR::Inst& inst, ScalarS32 value, ScalarS32 min, ScalarS32 max) {
     const Register ret{ctx.reg_alloc.Define(inst)};
-    ctx.Add("MIN.S {}.x,{},{};"
-            "MAX.S {}.x,{},{};",
-            ret, max, value, ret, ret, min);
+    ctx.Add("MIN.S RC.x,{},{};"
+            "MAX.S {}.x,RC.x,{};",
+            max, value, ret, min);
 }
 
 void EmitUClamp32(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU32 min, ScalarU32 max) {
     const Register ret{ctx.reg_alloc.Define(inst)};
-    ctx.Add("MIN.U {}.x,{},{};"
-            "MAX.U {}.x,{},{};",
-            ret, max, value, ret, ret, min);
+    ctx.Add("MIN.U RC.x,{},{};"
+            "MAX.U {}.x,RC.x,{};",
+            max, value, ret, min);
 }
 
 void EmitSLessThan(EmitContext& ctx, IR::Inst& inst, ScalarS32 lhs, ScalarS32 rhs) {
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_select.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_select.cpp
index cfde86047..b9e5cbbbe 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_select.cpp
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_select.cpp
@@ -43,7 +43,7 @@ void EmitSelectU64(EmitContext& ctx, IR::Inst& inst, ScalarS32 cond, Register tr
                 cond, ret, true_value);
     } else {
         ctx.Add("MOV.S.CC RC.x,{};"
-                "MOV.U64 {}.x(EQ.x),{};"
+                "MOV.U64 {}.x,{};"
                 "MOV.U64 {}.x(NE.x),{};",
                 cond, ret, false_value, ret, true_value);
     }