diff src/cpu/x86/vm/macroAssembler_x86.cpp @ 7482:989155e2d07a

Merge with hs25-b15.
author Thomas Wuerthinger <thomas.wuerthinger@oracle.com>
date Wed, 16 Jan 2013 01:34:24 +0100
parents 038dd2875b94
children b30b3c2a0cf2 db9981fd3124
line wrap: on
line diff
--- a/src/cpu/x86/vm/macroAssembler_x86.cpp	Tue Jan 15 18:54:02 2013 +0100
+++ b/src/cpu/x86/vm/macroAssembler_x86.cpp	Wed Jan 16 01:34:24 2013 +0100
@@ -1023,7 +1023,7 @@
 
 void MacroAssembler::leave() {
   // %%% is this really better? Why not on 32bit too?
-  emit_byte(0xC9); // LEAVE
+  emit_int8((unsigned char)0xC9); // LEAVE
 }
 
 void MacroAssembler::lneg(Register hi, Register lo) {
@@ -2112,11 +2112,11 @@
   if (UseAddressNop) {
     addr_nop_5();
   } else {
-    emit_byte(0x26); // es:
-    emit_byte(0x2e); // cs:
-    emit_byte(0x64); // fs:
-    emit_byte(0x65); // gs:
-    emit_byte(0x90);
+    emit_int8(0x26); // es:
+    emit_int8(0x2e); // cs:
+    emit_int8(0x64); // fs:
+    emit_int8(0x65); // gs:
+    emit_int8((unsigned char)0x90);
   }
 }
 
@@ -2534,13 +2534,13 @@
     int offs = (intptr_t)dst.target() - ((intptr_t)pc());
     if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
       // 0111 tttn #8-bit disp
-      emit_byte(0x70 | cc);
-      emit_byte((offs - short_size) & 0xFF);
+      emit_int8(0x70 | cc);
+      emit_int8((offs - short_size) & 0xFF);
     } else {
       // 0000 1111 1000 tttn #32-bit disp
-      emit_byte(0x0F);
-      emit_byte(0x80 | cc);
-      emit_long(offs - long_size);
+      emit_int8(0x0F);
+      emit_int8((unsigned char)(0x80 | cc));
+      emit_int32(offs - long_size);
     }
   } else {
 #ifdef ASSERT
@@ -3085,7 +3085,8 @@
 
 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
   // Used in sign-bit flipping with aligned address.
-  assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
+  bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
+  assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
   if (reachable(src)) {
     Assembler::pshufb(dst, as_Address(src));
   } else {
@@ -5223,6 +5224,22 @@
 
 }
 
+void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp) {
+  // cnt - number of qwords (8-byte words).
+  // base - start address, qword aligned.
+  assert(base==rdi, "base register must be edi for rep stos");
+  assert(tmp==rax,   "tmp register must be eax for rep stos");
+  assert(cnt==rcx,   "cnt register must be ecx for rep stos");
+
+  xorptr(tmp, tmp);
+  if (UseFastStosb) {
+    shlptr(cnt,3); // convert to number of bytes
+    rep_stosb();
+  } else {
+    NOT_LP64(shlptr(cnt,1);) // convert to number of dwords for 32-bit VM
+    rep_stos();
+  }
+}
 
 // IndexOf for constant substrings with size >= 8 chars
 // which don't need to be loaded through stack.
@@ -5658,42 +5675,114 @@
   testl(cnt2, cnt2);
   jcc(Assembler::zero, LENGTH_DIFF_LABEL);
 
-  // Load first characters
+  // Compare first characters
   load_unsigned_short(result, Address(str1, 0));
   load_unsigned_short(cnt1, Address(str2, 0));
-
-  // Compare first characters
   subl(result, cnt1);
   jcc(Assembler::notZero,  POP_LABEL);
-  decrementl(cnt2);
-  jcc(Assembler::zero, LENGTH_DIFF_LABEL);
-
-  {
-    // Check after comparing first character to see if strings are equivalent
-    Label LSkip2;
-    // Check if the strings start at same location
-    cmpptr(str1, str2);
-    jccb(Assembler::notEqual, LSkip2);
-
-    // Check if the length difference is zero (from stack)
-    cmpl(Address(rsp, 0), 0x0);
-    jcc(Assembler::equal,  LENGTH_DIFF_LABEL);
-
-    // Strings might not be equivalent
-    bind(LSkip2);
-  }
+  cmpl(cnt2, 1);
+  jcc(Assembler::equal, LENGTH_DIFF_LABEL);
+
+  // Check if the strings start at the same location.
+  cmpptr(str1, str2);
+  jcc(Assembler::equal, LENGTH_DIFF_LABEL);
 
   Address::ScaleFactor scale = Address::times_2;
   int stride = 8;
 
-  // Advance to next element
-  addptr(str1, 16/stride);
-  addptr(str2, 16/stride);
-
-  if (UseSSE42Intrinsics) {
+  if (UseAVX >= 2) {
+    Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
+    Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
+    Label COMPARE_TAIL_LONG;
+    int pcmpmask = 0x19;
+
+    // Setup to compare 16-chars (32-bytes) vectors,
+    // start from first character again because it has aligned address.
+    int stride2 = 16;
+    int adr_stride  = stride  << scale;
+    int adr_stride2 = stride2 << scale;
+
+    assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
+    // rax and rdx are used by pcmpestri as elements counters
+    movl(result, cnt2);
+    andl(cnt2, ~(stride2-1));   // cnt2 holds the vector count
+    jcc(Assembler::zero, COMPARE_TAIL_LONG);
+
+    // fast path : compare first 2 8-char vectors.
+    bind(COMPARE_16_CHARS);
+    movdqu(vec1, Address(str1, 0));
+    pcmpestri(vec1, Address(str2, 0), pcmpmask);
+    jccb(Assembler::below, COMPARE_INDEX_CHAR);
+
+    movdqu(vec1, Address(str1, adr_stride));
+    pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
+    jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
+    addl(cnt1, stride);
+
+    // Compare the characters at index in cnt1
+    bind(COMPARE_INDEX_CHAR); //cnt1 has the offset of the mismatching character
+    load_unsigned_short(result, Address(str1, cnt1, scale));
+    load_unsigned_short(cnt2, Address(str2, cnt1, scale));
+    subl(result, cnt2);
+    jmp(POP_LABEL);
+
+    // Setup the registers to start vector comparison loop
+    bind(COMPARE_WIDE_VECTORS);
+    lea(str1, Address(str1, result, scale));
+    lea(str2, Address(str2, result, scale));
+    subl(result, stride2);
+    subl(cnt2, stride2);
+    jccb(Assembler::zero, COMPARE_WIDE_TAIL);
+    negptr(result);
+
+    //  In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
+    bind(COMPARE_WIDE_VECTORS_LOOP);
+    vmovdqu(vec1, Address(str1, result, scale));
+    vpxor(vec1, Address(str2, result, scale));
+    vptest(vec1, vec1);
+    jccb(Assembler::notZero, VECTOR_NOT_EQUAL);
+    addptr(result, stride2);
+    subl(cnt2, stride2);
+    jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
+
+    // compare wide vectors tail
+    bind(COMPARE_WIDE_TAIL);
+    testptr(result, result);
+    jccb(Assembler::zero, LENGTH_DIFF_LABEL);
+
+    movl(result, stride2);
+    movl(cnt2, result);
+    negptr(result);
+    jmpb(COMPARE_WIDE_VECTORS_LOOP);
+
+    // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
+    bind(VECTOR_NOT_EQUAL);
+    lea(str1, Address(str1, result, scale));
+    lea(str2, Address(str2, result, scale));
+    jmp(COMPARE_16_CHARS);
+
+    // Compare tail chars, length between 1 to 15 chars
+    bind(COMPARE_TAIL_LONG);
+    movl(cnt2, result);
+    cmpl(cnt2, stride);
+    jccb(Assembler::less, COMPARE_SMALL_STR);
+
+    movdqu(vec1, Address(str1, 0));
+    pcmpestri(vec1, Address(str2, 0), pcmpmask);
+    jcc(Assembler::below, COMPARE_INDEX_CHAR);
+    subptr(cnt2, stride);
+    jccb(Assembler::zero, LENGTH_DIFF_LABEL);
+    lea(str1, Address(str1, result, scale));
+    lea(str2, Address(str2, result, scale));
+    negptr(cnt2);
+    jmpb(WHILE_HEAD_LABEL);
+
+    bind(COMPARE_SMALL_STR);
+  } else if (UseSSE42Intrinsics) {
     Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
     int pcmpmask = 0x19;
-    // Setup to compare 16-byte vectors
+    // Setup to compare 8-char (16-byte) vectors,
+    // start from first character again because it has aligned address.
     movl(result, cnt2);
     andl(cnt2, ~(stride - 1));   // cnt2 holds the vector count
     jccb(Assembler::zero, COMPARE_TAIL);
@@ -5725,7 +5814,7 @@
     jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
 
     // compare wide vectors tail
-    testl(result, result);
+    testptr(result, result);
     jccb(Assembler::zero, LENGTH_DIFF_LABEL);
 
     movl(cnt2, stride);
@@ -5737,21 +5826,20 @@
 
     // Mismatched characters in the vectors
     bind(VECTOR_NOT_EQUAL);
-    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);
+    addptr(cnt1, result);
+    load_unsigned_short(result, Address(str1, cnt1, scale));
+    load_unsigned_short(cnt2, Address(str2, cnt1, scale));
+    subl(result, cnt2);
     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, scale, 0));
-  lea(str2, Address(str2, cnt2, scale, 0));
+  lea(str1, Address(str1, cnt2, scale));
+  lea(str2, Address(str2, cnt2, scale));
+  decrementl(cnt2);  // first character was compared already
   negptr(cnt2);
 
   // Compare the rest of the elements
@@ -5816,7 +5904,44 @@
   shll(limit, 1);      // byte count != 0
   movl(result, limit); // copy
 
-  if (UseSSE42Intrinsics) {
+  if (UseAVX >= 2) {
+    // With AVX2, use 32-byte vector compare
+    Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
+
+    // Compare 32-byte vectors
+    andl(result, 0x0000001e);  //   tail count (in bytes)
+    andl(limit, 0xffffffe0);   // vector count (in bytes)
+    jccb(Assembler::zero, COMPARE_TAIL);
+
+    lea(ary1, Address(ary1, limit, Address::times_1));
+    lea(ary2, Address(ary2, limit, Address::times_1));
+    negptr(limit);
+
+    bind(COMPARE_WIDE_VECTORS);
+    vmovdqu(vec1, Address(ary1, limit, Address::times_1));
+    vmovdqu(vec2, Address(ary2, limit, Address::times_1));
+    vpxor(vec1, vec2);
+
+    vptest(vec1, vec1);
+    jccb(Assembler::notZero, FALSE_LABEL);
+    addptr(limit, 32);
+    jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
+
+    testl(result, result);
+    jccb(Assembler::zero, TRUE_LABEL);
+
+    vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
+    vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
+    vpxor(vec1, vec2);
+
+    vptest(vec1, vec1);
+    jccb(Assembler::notZero, FALSE_LABEL);
+    jmpb(TRUE_LABEL);
+
+    bind(COMPARE_TAIL); // limit is zero
+    movl(limit, result);
+    // Fallthru to tail compare
+  } else if (UseSSE42Intrinsics) {
     // With SSE4.2, use double quad vector compare
     Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
 
@@ -5994,29 +6119,53 @@
     {
       assert( UseSSE >= 2, "supported cpu only" );
       Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
-      // Fill 32-byte chunks
       movdl(xtmp, value);
-      pshufd(xtmp, xtmp, 0);
-
-      subl(count, 8 << shift);
-      jcc(Assembler::less, L_check_fill_8_bytes);
-      align(16);
-
-      BIND(L_fill_32_bytes_loop);
-
-      if (UseUnalignedLoadStores) {
-        movdqu(Address(to, 0), xtmp);
-        movdqu(Address(to, 16), xtmp);
+      if (UseAVX >= 2 && UseUnalignedLoadStores) {
+        // Fill 64-byte chunks
+        Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
+        vpbroadcastd(xtmp, xtmp);
+
+        subl(count, 16 << shift);
+        jcc(Assembler::less, L_check_fill_32_bytes);
+        align(16);
+
+        BIND(L_fill_64_bytes_loop);
+        vmovdqu(Address(to, 0), xtmp);
+        vmovdqu(Address(to, 32), xtmp);
+        addptr(to, 64);
+        subl(count, 16 << shift);
+        jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
+
+        BIND(L_check_fill_32_bytes);
+        addl(count, 8 << shift);
+        jccb(Assembler::less, L_check_fill_8_bytes);
+        vmovdqu(Address(to, 0), xtmp);
+        addptr(to, 32);
+        subl(count, 8 << shift);
       } else {
-        movq(Address(to, 0), xtmp);
-        movq(Address(to, 8), xtmp);
-        movq(Address(to, 16), xtmp);
-        movq(Address(to, 24), xtmp);
+        // Fill 32-byte chunks
+        pshufd(xtmp, xtmp, 0);
+
+        subl(count, 8 << shift);
+        jcc(Assembler::less, L_check_fill_8_bytes);
+        align(16);
+
+        BIND(L_fill_32_bytes_loop);
+
+        if (UseUnalignedLoadStores) {
+          movdqu(Address(to, 0), xtmp);
+          movdqu(Address(to, 16), xtmp);
+        } else {
+          movq(Address(to, 0), xtmp);
+          movq(Address(to, 8), xtmp);
+          movq(Address(to, 16), xtmp);
+          movq(Address(to, 24), xtmp);
+        }
+
+        addptr(to, 32);
+        subl(count, 8 << shift);
+        jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
       }
-
-      addptr(to, 32);
-      subl(count, 8 << shift);
-      jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
       BIND(L_check_fill_8_bytes);
       addl(count, 8 << shift);
       jccb(Assembler::zero, L_exit);