8302358: Behavior of adler32 changes after JDK-8300208

Reviewed-by: kvn, jbhateja

Author: Sandhya Viswanathan

src/hotspot/cpu/x86/stubGenerator_x86_64_adler.cpp

@@ -110,7 +110,7 @@ address StubGenerator::generate_updateBytesAdler32() {
   const XMMRegister xtmp4 = xmm9;
   const XMMRegister xtmp5 = xmm10;
 
-  Label SLOOP1, SPRELOOP1A_AVX2, SLOOP1A_AVX2, SLOOP1A_AVX3, AVX3_REDUCE, SKIP_LOOP_1A;
+  Label SLOOP1, SLOOP1A_AVX2, SLOOP1A_AVX3, AVX3_REDUCE, SKIP_LOOP_1A;
   Label SKIP_LOOP_1A_AVX3, FINISH, LT64, DO_FINAL, FINAL_LOOP, ZERO_SIZE, END;
 
   __ enter(); // required for proper stackwalking of RuntimeStub frame
@@ -133,6 +133,7 @@ address StubGenerator::generate_updateBytesAdler32() {
   __ movdl(xa, init_d); //vmovd - 32bit
 
   __ bind(SLOOP1);
+  __ vpxor(yb, yb, yb, VM_Version::supports_avx512vl() ? Assembler::AVX_512bit : Assembler::AVX_256bit);
   __ movl(s, LIMIT);
   __ cmpl(s, size);
   __ cmovl(Assembler::above, s, size); // s = min(size, LIMIT)
@@ -144,10 +145,8 @@ address StubGenerator::generate_updateBytesAdler32() {
   if (VM_Version::supports_avx512vl()) {
     // AVX2 performs better for smaller inputs because of leaner post loop reduction sequence..
     __ cmpl(s, MAX2(128, VM_Version::avx3_threshold()));
-    __ jcc(Assembler::belowEqual, SPRELOOP1A_AVX2);
-
+    __ jcc(Assembler::belowEqual, SLOOP1A_AVX2);
     __ lea(end, Address(s, data, Address::times_1, - (2*CHUNKSIZE -1)));
-    __ vpxor(yb, yb, yb, Assembler::AVX_512bit);
 
     // Some notes on vectorized main loop algorithm.
     // Additions are performed in slices of 16 bytes in the main loop.
@@ -164,8 +163,8 @@ address StubGenerator::generate_updateBytesAdler32() {
     // Since addition was performed in chunks of 16 bytes, thus to match the scalar implementation
     // Oth lane element must be repeatedly added 16 times, 1st element 15 times and so on so forth.
     // Thus we first multiply yb by 16 followed by subtracting appropriately scaled ya value.
-    // yb = 16 x yb  - [0 - 15] x ya
-    //    = 64 x [0 - 15] + 48 x [16 - 31] + 32 x [32 - 47] + 16 x [48 - 63]  -  [0 - 15] x ya
+    // yb = 16 x yb  - [a0 - a15] x ya
+    //    = 64 x [a0 - a15] + 48 x [a16 - a31] + 32 x [a32 - a47] + 16 x [a48 - a63]  -  [a0 - a15] x ya
     //    = 64 x a0 + 63 x a1 + 62 x a2 ...... + a63
     __ bind(SLOOP1A_AVX3);
       __ evpmovzxbd(ydata0, Address(data, 0), Assembler::AVX_512bit);
@@ -220,8 +219,6 @@ address StubGenerator::generate_updateBytesAdler32() {
   }
 
   __ align32();
-  __ bind(SPRELOOP1A_AVX2);
-  __ vpxor(yb, yb, yb, Assembler::AVX_256bit);
   __ bind(SLOOP1A_AVX2);
     __ vbroadcastf128(ydata, Address(data, 0), Assembler::AVX_256bit);
     __ addptr(data, CHUNKSIZE);

test/hotspot/jtreg/compiler/intrinsics/zip/TestAdler32.java

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2015, 2023, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -146,7 +146,7 @@ private static boolean check(Checksum adler0, Checksum adler1) {
         if (adler0.getValue() != adler1.getValue()) {
             System.err.printf("ERROR: adler0 = %08x, adler1 = %08x\n",
                               adler0.getValue(), adler1.getValue());
-            throw new AssertionError("TEST FAILED", null);
+            return false;
         }
         return true;
     }
@@ -166,6 +166,7 @@ private static void test_multi(int iters) {
         int len1 = 8;    // the  8B/iteration loop
         int len2 = 32;   // the 32B/iteration loop
         int len3 = 4096; // the 4KB/iteration loop
+        int len4 = 5552; // the adler limit
 
         byte[] b = initializedBytes(len3*16, 0);
         int[] offsets = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 16, 32, 64, 128, 256, 512 };
@@ -176,6 +177,8 @@ private static void test_multi(int iters) {
                         len2*2, len2*4, len2*8, len2*16, len2*32, len2*64,
                         len3, len3+1, len3+3, len3+5, len3+7,
                         len3*2, len3*4, len3*8,
+                        len4, len4+1, len4+3, len4+5, len4+7, len4+len1, len4+len2, len4+len3,
+                        len4*2, len4*4, len4*2+1, len4*4+4,
                         len1+len2, len1+len2+1, len1+len2+3, len1+len2+5, len1+len2+7,
                         len1+len3, len1+len3+1, len1+len3+3, len1+len3+5, len1+len3+7,
                         len2+len3, len2+len3+1, len2+len3+3, len2+len3+5, len2+len3+7,
@@ -214,8 +217,9 @@ private static void test_multi(int iters) {
         for (i = 0; i < offsets.length; i++) {
             for (j = 0; j < sizes.length; j++) {
                 if (!check(adler0[i*sizes.length + j], adler1[i*sizes.length + j])) {
-                    System.out.printf("offsets[%d] = %d", i, offsets[i]);
-                    System.out.printf("\tsizes[%d] = %d\n", j, sizes[j]);
+                    System.out.printf("Failed at: offsets[%d] = %d", i, offsets[i]);
+                    System.out.printf(", sizes[%d] = %d\n", j, sizes[j]);
+                    throw new AssertionError("TEST FAILED", null);
                 }
             }
         }