8369435: C2: transform (LShiftX (SubX con0 a), con1) into (SubX con0<<con1 (LShiftX a con1))

Reviewed-by: epeter, qamai

src/hotspot/share/opto/mulnode.cpp

@@ -1080,6 +1080,19 @@ Node* LShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
       }
     }
   }
+  // Check for "(con0 - X) << con1"
+  // Transform is legal, but check for profit.  Avoid breaking 'i2s'
+  // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
+  if (add1_op == Op_Sub(bt) && (bt != T_INT || con < 16)) {    // Left input is a sub?
+    // Left input is a sub from a constant?
+    const TypeInteger* t11 = phase->type(add1->in(1))->isa_integer(bt);
+    if (t11 != nullptr && t11->is_con()) {
+      // Compute X << con0
+      Node* lsh = phase->transform(LShiftNode::make(add1->in(2), in(2), bt));
+      // Compute (con1<<con0) - (X<<con0)
+      return SubNode::make(phase->integercon(java_shift_left(t11->get_con_as_long(bt), con, bt), bt), lsh, bt);
+    }
+  }
 
   // Check for "(x >> C1) << C2"
   if (add1_op == Op_RShift(bt) || add1_op == Op_URShift(bt)) {

test/hotspot/jtreg/compiler/c2/irTests/LShiftINodeIdealizationTests.java

@@ -65,6 +65,7 @@ public class LShiftINodeIdealizationTests {
             "testLShiftOfAndOfRShift",
             "testLShiftOfAndOfURShift",
             "testLShiftOfAndOfCon",
+            "testShiftOfSubConstant",
     })
     public void runMethod() {
         int a = RunInfo.getRandom().nextInt();
@@ -122,6 +123,7 @@ public class LShiftINodeIdealizationTests {
         Asserts.assertEQ((a + a) << 31, testLargeShiftOfAddSameInput(a));
         Asserts.assertEQ(((a + 1) << 1) + 1, testShiftOfAddConstant(a));
         Asserts.assertEQ((a & ((1 << (32 - 10)) -1)) << 10, testLShiftOfAndOfCon(a));
+        Asserts.assertEQ(((1 - a) << 1) + 1, testShiftOfSubConstant(a));
 
         assertDoubleShiftResult(a);
     }
@@ -359,17 +361,45 @@ public class LShiftINodeIdealizationTests {
     @Test
     @IR(counts = { IRNode.ADD_I, "1"} , failOn = { IRNode.LSHIFT_I, IRNode.RSHIFT_I } )
     @Arguments( values = { Argument.NUMBER_42 })
-    public void testStoreShort(int x) {
+    public void testStoreShort1(int x) {
         shortField = (short)(x + x);
     }
 
     @Test
     @IR(counts = { IRNode.ADD_I, "1"} , failOn = { IRNode.LSHIFT_I, IRNode.RSHIFT_I } )
     @Arguments( values = { Argument.NUMBER_42 })
-    public void testStoreByte(int x) {
+    public void testStoreByte1(int x) {
         byteField = (byte)(x + x);
     }
 
+    @Test
+    @IR(counts = { IRNode.ADD_I, "1"} , failOn = { IRNode.LSHIFT_I, IRNode.RSHIFT_I } )
+    @Arguments( values = { Argument.NUMBER_42 })
+    public void testStoreShort2(int x) {
+        shortField = (short)(x + 1);
+    }
+
+    @Test
+    @IR(counts = { IRNode.ADD_I, "1"} , failOn = { IRNode.LSHIFT_I, IRNode.RSHIFT_I } )
+    @Arguments( values = { Argument.NUMBER_42 })
+    public void testStoreByte2(int x) {
+        byteField = (byte)(x + 1);
+    }
+
+    @Test
+    @IR(counts = { IRNode.SUB_I, "1"} , failOn = { IRNode.LSHIFT_I, IRNode.RSHIFT_I } )
+    @Arguments( values = { Argument.NUMBER_42 })
+    public void testStoreShort3(int x) {
+        shortField = (short)(1 - x);
+    }
+
+    @Test
+    @IR(counts = { IRNode.SUB_I, "1"} , failOn = { IRNode.LSHIFT_I, IRNode.RSHIFT_I } )
+    @Arguments( values = { Argument.NUMBER_42 })
+    public void testStoreByte3(int x) {
+        byteField = (byte)(1 - x);
+    }
+
     static int otherInput;
 
     @Test
@@ -409,4 +439,10 @@ public class LShiftINodeIdealizationTests {
     public int testLShiftOfAndOfCon(int x) {
         return (x & ((1 << (32 - 10)) -1)) << 10;
     }
+
+    @Test
+    @IR(counts = { IRNode.LSHIFT_I, "1",  IRNode.SUB_I, "1" }, failOn =  { IRNode.ADD_I })
+    public int testShiftOfSubConstant(int x) {
+        return ((1 - x) << 1) + 1;
+    }
 }

test/hotspot/jtreg/compiler/c2/irTests/LShiftLNodeIdealizationTests.java

@@ -63,6 +63,7 @@ public class LShiftLNodeIdealizationTests {
             "testLShiftOfAndOfRShift",
             "testLShiftOfAndOfURShift",
             "testLShiftOfAndOfCon",
+            "testShiftOfSubConstant",
     })
     public void runMethod() {
         long a = RunInfo.getRandom().nextLong();
@@ -118,6 +119,7 @@ public class LShiftLNodeIdealizationTests {
         Asserts.assertEQ((a + a) << 63, testLargeShiftOfAddSameInput(a));
         Asserts.assertEQ(((a + 1) << 1) + 1, testShiftOfAddConstant(a));
         Asserts.assertEQ((a & ((1L << (64 - 10)) -1)) << 10, testLShiftOfAndOfCon(a));
+        Asserts.assertEQ(((1L - a) << 1) + 1, testShiftOfSubConstant(a));
 
         assertDoubleShiftResult(a);
     }
@@ -358,4 +360,10 @@ public class LShiftLNodeIdealizationTests {
     public long testLShiftOfAndOfCon(long x) {
         return (x & ((1L << (64 - 10)) -1)) << 10;
     }
+
+    @Test
+    @IR(counts = { IRNode.LSHIFT_L, "1",  IRNode.SUB_L, "1" }, failOn =  { IRNode.ADD_L })
+    public long testShiftOfSubConstant(long x) {
+        return ((1 - x) << 1) + 1;
+    }
 }

test/hotspot/jtreg/compiler/loopopts/superword/TestMemorySegment_SubOfShift.java

@@ -30,17 +30,17 @@ import compiler.lib.ir_framework.*;
 
 /*
  * @test
- * @bug 8324751
+ * @bug 8324751 8369435
  * @summary Reported issue: JDK-8359688: C2 SuperWord: missing RCE with MemorySegment
  *          The examples are generated from TestAliasingFuzzer.java
  *          So if you see something change here, you may want to investigate if we
  *          can also tighten up the IR rules there.
  * @library /test/lib /
- * @run driver compiler.loopopts.superword.TestMemorySegment_8359688
+ * @run driver compiler.loopopts.superword.TestMemorySegment_SubOfShift
  */
 
 
-public class TestMemorySegment_8359688 {
+public class TestMemorySegment_SubOfShift {
 
     public static MemorySegment b = MemorySegment.ofArray(new long[4 * 30_000]);
 
@@ -61,17 +61,13 @@ public class TestMemorySegment_8359688 {
 
     @Test
     @Arguments(setup = "setup")
-    @IR(counts = {IRNode.STORE_VECTOR, "= 0",
-                  IRNode.REPLICATE_L,  "= 0",
+    @IR(counts = {IRNode.STORE_VECTOR, "> 0",
+                  IRNode.REPLICATE_L,  "= 1",
                   ".*multiversion.*",  "= 0"}, // AutoVectorization Predicate SUFFICES, there is no aliasing
         phase = CompilePhase.PRINT_IDEAL,
         applyIfPlatform = {"64-bit", "true"},
         applyIf = {"AlignVector", "false"},
         applyIfCPUFeatureOr = {"avx", "true", "asimd", "true"})
-    // Does not manage to remove all RangeChecks -> no vectorization
-    // If you see this IR rule fail: investigate JDK-8359688, possibly close it and fix this IR rule!
-    // Also: consider renaming the file to something more descriptive: what have you fixed with this?
-    // And: you may now be able to tighten IR rules in TestAliasingFuzzer.java
     public static void test1(MemorySegment b, int ivLo, int ivHi, int invar) {
         for (int i = ivLo; i < ivHi; i++) {
             b.setAtIndex(ValueLayout.JAVA_LONG_UNALIGNED, 30_000L - (long)i + (long)invar, 42);