# HG changeset patch
# User never
# Date 1297292543 28800
# Node ID e9a2b8e0572a86b7b2e8ec0f3eb257cef2618976
# Parent  908c5092d72a8e7697995d4016035d28892f45d1
7016474: string compare intrinsic improvements
Reviewed-by: kvn

diff -r 908c5092d72a -r e9a2b8e0572a src/cpu/x86/vm/assembler_x86.cpp
--- a/src/cpu/x86/vm/assembler_x86.cpp	Mon Jan 31 16:28:40 2011 -0500
+++ b/src/cpu/x86/vm/assembler_x86.cpp	Wed Feb 09 15:02:23 2011 -0800
@@ -2349,6 +2349,17 @@
   a_byte(p);
 }
 
+void Assembler::por(XMMRegister dst, XMMRegister src) {
+  NOT_LP64(assert(VM_Version::supports_sse2(), ""));
+
+  emit_byte(0x66);
+  int  encode = prefix_and_encode(dst->encoding(), src->encoding());
+  emit_byte(0x0F);
+
+  emit_byte(0xEB);
+  emit_byte(0xC0 | encode);
+}
+
 void Assembler::pshufd(XMMRegister dst, XMMRegister src, int mode) {
   assert(isByte(mode), "invalid value");
   NOT_LP64(assert(VM_Version::supports_sse2(), ""));
@@ -8636,7 +8647,7 @@
 // Compare strings.
 void MacroAssembler::string_compare(Register str1, Register str2,
                                     Register cnt1, Register cnt2, Register result,
-                                    XMMRegister vec1, XMMRegister vec2) {
+                                    XMMRegister vec1) {
   Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
 
   // Compute the minimum of the string lengths and the
@@ -8683,62 +8694,85 @@
     bind(LSkip2);
   }
 
-  // Advance to next character
-  addptr(str1, 2);
-  addptr(str2, 2);
+  Address::ScaleFactor scale = Address::times_2;
+  int stride = 8;
+
+  // Advance to next element
+  addptr(str1, 16/stride);
+  addptr(str2, 16/stride);
 
   if (UseSSE42Intrinsics) {
-    // With SSE4.2, use double quad vector compare
-    Label COMPARE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
+    Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
+    int pcmpmask = 0x19;
     // Setup to compare 16-byte vectors
-    movl(cnt1, cnt2);
-    andl(cnt2, 0xfffffff8); // cnt2 holds the vector count
-    andl(cnt1, 0x00000007); // cnt1 holds the tail count
-    testl(cnt2, cnt2);
+    movl(result, cnt2);
+    andl(cnt2, ~(stride - 1));   // cnt2 holds the vector count
     jccb(Assembler::zero, COMPARE_TAIL);
 
-    lea(str2, Address(str2, cnt2, Address::times_2));
-    lea(str1, Address(str1, cnt2, Address::times_2));
-    negptr(cnt2);
-
-    bind(COMPARE_VECTORS);
-    movdqu(vec1, Address(str1, cnt2, Address::times_2));
-    movdqu(vec2, Address(str2, cnt2, Address::times_2));
-    pxor(vec1, vec2);
-    ptest(vec1, vec1);
-    jccb(Assembler::notZero, VECTOR_NOT_EQUAL);
-    addptr(cnt2, 8);
-    jcc(Assembler::notZero, COMPARE_VECTORS);
-    jmpb(COMPARE_TAIL);
+    lea(str1, Address(str1, result, scale));
+    lea(str2, Address(str2, result, scale));
+    negptr(result);
+
+    // pcmpestri
+    //   inputs:
+    //     vec1- substring
+    //     rax - negative string length (elements count)
+    //     mem - scaned string
+    //     rdx - string length (elements count)
+    //     pcmpmask - cmp mode: 11000 (string compare with negated result)
+    //               + 00 (unsigned bytes) or  + 01 (unsigned shorts)
+    //   outputs:
+    //     rcx - first mismatched element index
+    assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
+
+    bind(COMPARE_WIDE_VECTORS);
+    movdqu(vec1, Address(str1, result, scale));
+    pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
+    // After pcmpestri cnt1(rcx) contains mismatched element index
+
+    jccb(Assembler::below, VECTOR_NOT_EQUAL);  // CF==1
+    addptr(result, stride);
+    subptr(cnt2, stride);
+    jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
+
+    // compare wide vectors tail
+    testl(result, result);
+    jccb(Assembler::zero, LENGTH_DIFF_LABEL);
+
+    movl(cnt2, stride);
+    movl(result, stride);
+    negptr(result);
+    movdqu(vec1, Address(str1, result, scale));
+    pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
+    jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
 
     // Mismatched characters in the vectors
     bind(VECTOR_NOT_EQUAL);
-    lea(str1, Address(str1, cnt2, Address::times_2));
-    lea(str2, Address(str2, cnt2, Address::times_2));
-    movl(cnt1, 8);
-
-    // Compare tail (< 8 chars), or rescan last vectors to
-    // find 1st mismatched characters
-    bind(COMPARE_TAIL);
-    testl(cnt1, cnt1);
-    jccb(Assembler::zero, LENGTH_DIFF_LABEL);
-    movl(cnt2, cnt1);
+    addptr(result, cnt1);
+    movptr(cnt2, result);
+    load_unsigned_short(result, Address(str1, cnt2, scale));
+    load_unsigned_short(cnt1, Address(str2, cnt2, scale));
+    subl(result, cnt1);
+    jmpb(POP_LABEL);
+
+    bind(COMPARE_TAIL); // limit is zero
+    movl(cnt2, result);
     // Fallthru to tail compare
   }
 
   // Shift str2 and str1 to the end of the arrays, negate min
-  lea(str1, Address(str1, cnt2, Address::times_2, 0));
-  lea(str2, Address(str2, cnt2, Address::times_2, 0));
+  lea(str1, Address(str1, cnt2, scale, 0));
+  lea(str2, Address(str2, cnt2, scale, 0));
   negptr(cnt2);
 
-    // Compare the rest of the characters
+  // Compare the rest of the elements
   bind(WHILE_HEAD_LABEL);
-  load_unsigned_short(result, Address(str1, cnt2, Address::times_2, 0));
-  load_unsigned_short(cnt1, Address(str2, cnt2, Address::times_2, 0));
+  load_unsigned_short(result, Address(str1, cnt2, scale, 0));
+  load_unsigned_short(cnt1, Address(str2, cnt2, scale, 0));
   subl(result, cnt1);
   jccb(Assembler::notZero, POP_LABEL);
   increment(cnt2);
-  jcc(Assembler::notZero, WHILE_HEAD_LABEL);
+  jccb(Assembler::notZero, WHILE_HEAD_LABEL);
 
   // Strings are equal up to min length.  Return the length difference.
   bind(LENGTH_DIFF_LABEL);
@@ -8747,7 +8781,7 @@
 
   // Discard the stored length difference
   bind(POP_LABEL);
-  addptr(rsp, wordSize);
+  pop(cnt1);
 
   // That's it
   bind(DONE_LABEL);
@@ -8795,6 +8829,7 @@
   if (UseSSE42Intrinsics) {
     // With SSE4.2, use double quad vector compare
     Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
+
     // Compare 16-byte vectors
     andl(result, 0x0000000e);  //   tail count (in bytes)
     andl(limit, 0xfffffff0);   // vector count (in bytes)
@@ -8808,11 +8843,23 @@
     movdqu(vec1, Address(ary1, limit, Address::times_1));
     movdqu(vec2, Address(ary2, limit, Address::times_1));
     pxor(vec1, vec2);
+
     ptest(vec1, vec1);
     jccb(Assembler::notZero, FALSE_LABEL);
     addptr(limit, 16);
     jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
 
+    testl(result, result);
+    jccb(Assembler::zero, TRUE_LABEL);
+
+    movdqu(vec1, Address(ary1, result, Address::times_1, -16));
+    movdqu(vec2, Address(ary2, result, Address::times_1, -16));
+    pxor(vec1, vec2);
+
+    ptest(vec1, vec1);
+    jccb(Assembler::notZero, FALSE_LABEL);
+    jmpb(TRUE_LABEL);
+
     bind(COMPARE_TAIL); // limit is zero
     movl(limit, result);
     // Fallthru to tail compare
diff -r 908c5092d72a -r e9a2b8e0572a src/cpu/x86/vm/assembler_x86.hpp
--- a/src/cpu/x86/vm/assembler_x86.hpp	Mon Jan 31 16:28:40 2011 -0500
+++ b/src/cpu/x86/vm/assembler_x86.hpp	Wed Feb 09 15:02:23 2011 -0800
@@ -1277,6 +1277,9 @@
   void prefetcht2(Address src);
   void prefetchw(Address src);
 
+  // POR - Bitwise logical OR
+  void por(XMMRegister dst, XMMRegister src);
+
   // Shuffle Packed Doublewords
   void pshufd(XMMRegister dst, XMMRegister src, int mode);
   void pshufd(XMMRegister dst, Address src,     int mode);
@@ -2293,7 +2296,7 @@
   // Compare strings.
   void string_compare(Register str1, Register str2,
                       Register cnt1, Register cnt2, Register result,
-                      XMMRegister vec1, XMMRegister vec2);
+                      XMMRegister vec1);
 
   // Compare char[] arrays.
   void char_arrays_equals(bool is_array_equ, Register ary1, Register ary2,
diff -r 908c5092d72a -r e9a2b8e0572a src/cpu/x86/vm/x86_32.ad
--- a/src/cpu/x86/vm/x86_32.ad	Mon Jan 31 16:28:40 2011 -0500
+++ b/src/cpu/x86/vm/x86_32.ad	Wed Feb 09 15:02:23 2011 -0800
@@ -12629,16 +12629,16 @@
   ins_pipe( pipe_slow );
 %}
 
-instruct string_compare(eDIRegP str1, eCXRegI cnt1, eSIRegP str2, eBXRegI cnt2,
-                        eAXRegI result, regXD tmp1, regXD tmp2, eFlagsReg cr) %{
+instruct string_compare(eDIRegP str1, eCXRegI cnt1, eSIRegP str2, eDXRegI cnt2,
+                        eAXRegI result, regXD tmp1, eFlagsReg cr) %{
   match(Set result (StrComp (Binary str1 cnt1) (Binary str2 cnt2)));
-  effect(TEMP tmp1, TEMP tmp2, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, USE_KILL cnt2, KILL cr);
-
-  format %{ "String Compare $str1,$cnt1,$str2,$cnt2 -> $result   // KILL $tmp1, $tmp2" %}
+  effect(TEMP tmp1, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, USE_KILL cnt2, KILL cr);
+
+  format %{ "String Compare $str1,$cnt1,$str2,$cnt2 -> $result   // KILL $tmp1" %}
   ins_encode %{
     __ string_compare($str1$$Register, $str2$$Register,
                       $cnt1$$Register, $cnt2$$Register, $result$$Register,
-                      $tmp1$$XMMRegister, $tmp2$$XMMRegister);
+                      $tmp1$$XMMRegister);
   %}
   ins_pipe( pipe_slow );
 %}
diff -r 908c5092d72a -r e9a2b8e0572a src/cpu/x86/vm/x86_64.ad
--- a/src/cpu/x86/vm/x86_64.ad	Mon Jan 31 16:28:40 2011 -0500
+++ b/src/cpu/x86/vm/x86_64.ad	Wed Feb 09 15:02:23 2011 -0800
@@ -11583,17 +11583,17 @@
   ins_pipe(pipe_slow);
 %}
 
-instruct string_compare(rdi_RegP str1, rcx_RegI cnt1, rsi_RegP str2, rbx_RegI cnt2,
-                        rax_RegI result, regD tmp1, regD tmp2, rFlagsReg cr)
+instruct string_compare(rdi_RegP str1, rcx_RegI cnt1, rsi_RegP str2, rdx_RegI cnt2,
+                        rax_RegI result, regD tmp1, rFlagsReg cr)
 %{
   match(Set result (StrComp (Binary str1 cnt1) (Binary str2 cnt2)));
-  effect(TEMP tmp1, TEMP tmp2, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, USE_KILL cnt2, KILL cr);
-
-  format %{ "String Compare $str1,$cnt1,$str2,$cnt2 -> $result   // KILL $tmp1, $tmp2" %}
+  effect(TEMP tmp1, USE_KILL str1, USE_KILL str2, USE_KILL cnt1, USE_KILL cnt2, KILL cr);
+
+  format %{ "String Compare $str1,$cnt1,$str2,$cnt2 -> $result   // KILL $tmp1" %}
   ins_encode %{
     __ string_compare($str1$$Register, $str2$$Register,
                       $cnt1$$Register, $cnt2$$Register, $result$$Register,
-                      $tmp1$$XMMRegister, $tmp2$$XMMRegister);
+                      $tmp1$$XMMRegister);
   %}
   ins_pipe( pipe_slow );
 %}