JIT: Speed up floating to integer casts on x86/x64 (#114410)

* speed up floating to integer conversion

* fix linux build

* fix double->uint SSE, don't use fixupimm

Author: saucecontrol

src/coreclr/jit/lower.cpp

@@ -550,13 +550,9 @@ GenTree* Lowering::LowerNode(GenTree* node)
                 return nextNode;
             }
 
-            nextNode = LowerCast(node);
-            if (nextNode != nullptr)
-            {
-                return nextNode;
-            }
+            LowerCast(node);
+            break;
         }
-        break;
 
         case GT_BITCAST:
         {

src/coreclr/jit/lower.h

@@ -427,7 +427,7 @@ class Lowering final : public Phase
                                      GenTree*    switchValue,
                                      weight_t    defaultLikelihood);
 
-    GenTree* LowerCast(GenTree* node);
+    void LowerCast(GenTree* node);
 
 #if !CPU_LOAD_STORE_ARCH
     bool IsRMWIndirCandidate(GenTree* operand, GenTree* storeInd);

src/coreclr/jit/lowerarmarch.cpp

@@ -968,22 +968,11 @@ void Lowering::LowerPutArgStkOrSplit(GenTreePutArgStk* putArgNode)
 //    tree - GT_CAST node to be lowered
 //
 // Return Value:
-//    nextNode to be lowered if tree is modified else returns nullptr
-//
-// Notes:
-//    Casts from float/double to a smaller int type are transformed as follows:
-//    GT_CAST(float/double, byte)     =   GT_CAST(GT_CAST(float/double, int32), byte)
-//    GT_CAST(float/double, sbyte)    =   GT_CAST(GT_CAST(float/double, int32), sbyte)
-//    GT_CAST(float/double, int16)    =   GT_CAST(GT_CAST(double/double, int32), int16)
-//    GT_CAST(float/double, uint16)   =   GT_CAST(GT_CAST(double/double, int32), uint16)
-//
-//    Note that for the overflow conversions we still depend on helper calls and
-//    don't expect to see them here.
-//    i) GT_CAST(float/double, int type with overflow detection)
+//    None.
 //
-GenTree* Lowering::LowerCast(GenTree* tree)
+void Lowering::LowerCast(GenTree* tree)
 {
-    assert(tree->OperGet() == GT_CAST);
+    assert(tree->OperIs(GT_CAST));
 
     JITDUMP("LowerCast for: ");
     DISPNODE(tree);
@@ -995,17 +984,16 @@ GenTree* Lowering::LowerCast(GenTree* tree)
 
     if (varTypeIsFloating(srcType))
     {
+        // Overflow casts should have been converted to helper call in morph.
         noway_assert(!tree->gtOverflow());
-        assert(!varTypeIsSmall(dstType)); // fgMorphCast creates intermediate casts when converting from float to small
-                                          // int.
+        // Small types should have had an intermediate int cast inserted in morph.
+        assert(!varTypeIsSmall(dstType));
     }
 
     assert(!varTypeIsSmall(srcType));
 
     // Now determine if we have operands that should be contained.
     ContainCheckCast(tree->AsCast());
-
-    return nullptr;
 }
 
 //------------------------------------------------------------------------

src/coreclr/jit/lowerloongarch64.cpp

@@ -515,21 +515,9 @@ void Lowering::LowerPutArgStkOrSplit(GenTreePutArgStk* putArgNode)
 // Return Value:
 //    None.
 //
-// Notes:
-//    Casts from float/double to a smaller int type are transformed as follows:
-//    GT_CAST(float/double, byte)     =   GT_CAST(GT_CAST(float/double, int32), byte)
-//    GT_CAST(float/double, sbyte)    =   GT_CAST(GT_CAST(float/double, int32), sbyte)
-//    GT_CAST(float/double, int16)    =   GT_CAST(GT_CAST(double/double, int32), int16)
-//    GT_CAST(float/double, uint16)   =   GT_CAST(GT_CAST(double/double, int32), uint16)
-//
-//    Note that for the overflow conversions we still depend on helper calls and
-//    don't expect to see them here.
-//    i) GT_CAST(float/double, int type with overflow detection)
-//
-
-GenTree* Lowering::LowerCast(GenTree* tree)
+void Lowering::LowerCast(GenTree* tree)
 {
-    assert(tree->OperGet() == GT_CAST);
+    assert(tree->OperIs(GT_CAST));
 
     JITDUMP("LowerCast for: ");
     DISPNODE(tree);
@@ -541,17 +529,16 @@ GenTree* Lowering::LowerCast(GenTree* tree)
 
     if (varTypeIsFloating(srcType))
     {
+        // Overflow casts should have been converted to helper call in morph.
         noway_assert(!tree->gtOverflow());
-        assert(!varTypeIsSmall(dstType)); // fgMorphCast creates intermediate casts when converting from float to small
-                                          // int.
+        // Small types should have had an intermediate int cast inserted in morph.
+        assert(!varTypeIsSmall(dstType));
     }
 
     assert(!varTypeIsSmall(srcType));
 
     // Now determine if we have operands that should be contained.
     ContainCheckCast(tree->AsCast());
-
-    return nullptr;
 }
 
 //------------------------------------------------------------------------

src/coreclr/jit/lowerriscv64.cpp

@@ -504,21 +504,9 @@ void Lowering::LowerPutArgStkOrSplit(GenTreePutArgStk* putArgNode)
 // Return Value:
 //    None.
 //
-// Notes:
-//    Casts from float/double to a smaller int type are transformed as follows:
-//    GT_CAST(float/double, byte)     =   GT_CAST(GT_CAST(float/double, int32), byte)
-//    GT_CAST(float/double, sbyte)    =   GT_CAST(GT_CAST(float/double, int32), sbyte)
-//    GT_CAST(float/double, int16)    =   GT_CAST(GT_CAST(double/double, int32), int16)
-//    GT_CAST(float/double, uint16)   =   GT_CAST(GT_CAST(double/double, int32), uint16)
-//
-//    Note that for the overflow conversions we still depend on helper calls and
-//    don't expect to see them here.
-//    i) GT_CAST(float/double, int type with overflow detection)
-//
-
-GenTree* Lowering::LowerCast(GenTree* tree)
+void Lowering::LowerCast(GenTree* tree)
 {
-    assert(tree->OperGet() == GT_CAST);
+    assert(tree->OperIs(GT_CAST));
 
     JITDUMP("LowerCast for: ");
     DISPNODE(tree);
@@ -530,17 +518,16 @@ GenTree* Lowering::LowerCast(GenTree* tree)
 
     if (varTypeIsFloating(srcType))
     {
+        // Overflow casts should have been converted to helper call in morph.
         noway_assert(!tree->gtOverflow());
-        assert(!varTypeIsSmall(dstType)); // fgMorphCast creates intermediate casts when converting from float to small
-                                          // int.
+        // Small types should have had an intermediate int cast inserted in morph.
+        assert(!varTypeIsSmall(dstType));
     }
 
     assert(!varTypeIsSmall(srcType));
 
     // Now determine if we have operands that should be contained.
     ContainCheckCast(tree->AsCast());
-
-    return nullptr;
 }
 
 //------------------------------------------------------------------------