8052081: Optimize generated by C2 code for Intel's Atom processor

Summary: Allow to execute vectorization and crc32 optimization on Atom. Enable UseFPUForSpilling by default on x86.
Reviewed-by: roland

src/cpu/x86/vm/assembler_x86.cpp

@@ -3853,6 +3853,15 @@ void Assembler::vpbroadcastd(XMMRegister dst, XMMRegister src) {
   emit_int8((unsigned char)(0xC0 | encode));
 }
 
+// Carry-Less Multiplication Quadword
+void Assembler::pclmulqdq(XMMRegister dst, XMMRegister src, int mask) {
+  assert(VM_Version::supports_clmul(), "");
+  int encode = simd_prefix_and_encode(dst, dst, src, VEX_SIMD_66, VEX_OPCODE_0F_3A);
+  emit_int8(0x44);
+  emit_int8((unsigned char)(0xC0 | encode));
+  emit_int8((unsigned char)mask);
+}
+
 // Carry-Less Multiplication Quadword
 void Assembler::vpclmulqdq(XMMRegister dst, XMMRegister nds, XMMRegister src, int mask) {
   assert(VM_Version::supports_avx() && VM_Version::supports_clmul(), "");

src/cpu/x86/vm/assembler_x86.hpp

@@ -1837,6 +1837,7 @@ private:
   void vpbroadcastd(XMMRegister dst, XMMRegister src);
 
   // Carry-Less Multiplication Quadword
+  void pclmulqdq(XMMRegister dst, XMMRegister src, int mask);
   void vpclmulqdq(XMMRegister dst, XMMRegister nds, XMMRegister src, int mask);
 
   // AVX instruction which is used to clear upper 128 bits of YMM registers and

src/cpu/x86/vm/macroAssembler_x86.cpp

@@ -7316,17 +7316,34 @@ void MacroAssembler::update_byte_crc32(Register crc, Register val, Register tabl
  * Fold 128-bit data chunk
  */
 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
-  vpclmulhdq(xtmp, xK, xcrc); // [123:64]
-  vpclmulldq(xcrc, xK, xcrc); // [63:0]
-  vpxor(xcrc, xcrc, Address(buf, offset), false /* vector256 */);
-  pxor(xcrc, xtmp);
+  if (UseAVX > 0) {
+    vpclmulhdq(xtmp, xK, xcrc); // [123:64]
+    vpclmulldq(xcrc, xK, xcrc); // [63:0]
+    vpxor(xcrc, xcrc, Address(buf, offset), false /* vector256 */);
+    pxor(xcrc, xtmp);
+  } else {
+    movdqa(xtmp, xcrc);
+    pclmulhdq(xtmp, xK);   // [123:64]
+    pclmulldq(xcrc, xK);   // [63:0]
+    pxor(xcrc, xtmp);
+    movdqu(xtmp, Address(buf, offset));
+    pxor(xcrc, xtmp);
+  }
 }
 
 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
-  vpclmulhdq(xtmp, xK, xcrc);
-  vpclmulldq(xcrc, xK, xcrc);
-  pxor(xcrc, xbuf);
-  pxor(xcrc, xtmp);
+  if (UseAVX > 0) {
+    vpclmulhdq(xtmp, xK, xcrc);
+    vpclmulldq(xcrc, xK, xcrc);
+    pxor(xcrc, xbuf);
+    pxor(xcrc, xtmp);
+  } else {
+    movdqa(xtmp, xcrc);
+    pclmulhdq(xtmp, xK);
+    pclmulldq(xcrc, xK);
+    pxor(xcrc, xbuf);
+    pxor(xcrc, xtmp);
+  }
 }
 
 /**
@@ -7444,9 +7461,17 @@ void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Regi
   // Fold 128 bits in xmm1 down into 32 bits in crc register.
   BIND(L_fold_128b);
   movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
-  vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
-  vpand(xmm3, xmm0, xmm2, false /* vector256 */);
-  vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
+  if (UseAVX > 0) {
+    vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
+    vpand(xmm3, xmm0, xmm2, false /* vector256 */);
+    vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
+  } else {
+    movdqa(xmm2, xmm0);
+    pclmulqdq(xmm2, xmm1, 0x1);
+    movdqa(xmm3, xmm0);
+    pand(xmm3, xmm2);
+    pclmulqdq(xmm0, xmm3, 0x1);
+  }
   psrldq(xmm1, 8);
   psrldq(xmm2, 4);
   pxor(xmm0, xmm1);

src/cpu/x86/vm/macroAssembler_x86.hpp

@@ -966,6 +966,16 @@ public:
   void mulss(XMMRegister dst, Address src)        { Assembler::mulss(dst, src); }
   void mulss(XMMRegister dst, AddressLiteral src);
 
+  // Carry-Less Multiplication Quadword
+  void pclmulldq(XMMRegister dst, XMMRegister src) {
+    // 0x00 - multiply lower 64 bits [0:63]
+    Assembler::pclmulqdq(dst, src, 0x00);
+  }
+  void pclmulhdq(XMMRegister dst, XMMRegister src) {
+    // 0x11 - multiply upper 64 bits [64:127]
+    Assembler::pclmulqdq(dst, src, 0x11);
+  }
+
   void sqrtsd(XMMRegister dst, XMMRegister src)    { Assembler::sqrtsd(dst, src); }
   void sqrtsd(XMMRegister dst, Address src)        { Assembler::sqrtsd(dst, src); }
   void sqrtsd(XMMRegister dst, AddressLiteral src);

src/cpu/x86/vm/vm_version_x86.cpp

@@ -568,7 +568,7 @@ void VM_Version::get_processor_features() {
     FLAG_SET_DEFAULT(UseCLMUL, false);
   }
 
-  if (UseCLMUL && (UseAVX > 0) && (UseSSE > 2)) {
+  if (UseCLMUL && (UseSSE > 2)) {
     if (FLAG_IS_DEFAULT(UseCRC32Intrinsics)) {
       UseCRC32Intrinsics = true;
     }
@@ -803,6 +803,21 @@ void VM_Version::get_processor_features() {
         }
       }
     }
+    if ((cpu_family() == 0x06) &&
+        ((extended_cpu_model() == 0x36) || // Centerton
+         (extended_cpu_model() == 0x37) || // Silvermont
+         (extended_cpu_model() == 0x4D))) {
+#ifdef COMPILER2
+      if (FLAG_IS_DEFAULT(OptoScheduling)) {
+        OptoScheduling = true;
+      }
+#endif
+      if (supports_sse4_2()) { // Silvermont
+        if (FLAG_IS_DEFAULT(UseUnalignedLoadStores)) {
+          UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus
+        }
+      }
+    }
   }
 
   // Use count leading zeros count instruction if available.
@@ -890,23 +905,25 @@ void VM_Version::get_processor_features() {
   AllocatePrefetchDistance = allocate_prefetch_distance();
   AllocatePrefetchStyle    = allocate_prefetch_style();
 
-  if( is_intel() && cpu_family() == 6 && supports_sse3() ) {
-    if( AllocatePrefetchStyle == 2 ) { // watermark prefetching on Core
+  if (is_intel() && cpu_family() == 6 && supports_sse3()) {
+    if (AllocatePrefetchStyle == 2) { // watermark prefetching on Core
 #ifdef _LP64
       AllocatePrefetchDistance = 384;
 #else
       AllocatePrefetchDistance = 320;
 #endif
     }
-    if( supports_sse4_2() && supports_ht() ) { // Nehalem based cpus
+    if (supports_sse4_2() && supports_ht()) { // Nehalem based cpus
       AllocatePrefetchDistance = 192;
       AllocatePrefetchLines = 4;
+    }
 #ifdef COMPILER2
-      if (AggressiveOpts && FLAG_IS_DEFAULT(UseFPUForSpilling)) {
+    if (supports_sse4_2()) {
+      if (FLAG_IS_DEFAULT(UseFPUForSpilling)) {
         FLAG_SET_DEFAULT(UseFPUForSpilling, true);
       }
-#endif
     }
+#endif
   }
   assert(AllocatePrefetchDistance % AllocatePrefetchStepSize == 0, "invalid value");
 

src/share/vm/opto/lcm.cpp

@@ -484,7 +484,9 @@ Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &re
         iop == Op_CreateEx ||   // Create-exception must start block
         iop == Op_CheckCastPP
         ) {
-      worklist.map(i,worklist.pop());
+      // select the node n
+      // remove n from worklist and retain the order of remaining nodes
+      worklist.remove((uint)i);
       return n;
     }
 
@@ -570,7 +572,9 @@ Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &re
   assert(idx >= 0, "index should be set");
   Node *n = worklist[(uint)idx];      // Get the winner
 
-  worklist.map((uint)idx, worklist.pop());     // Compress worklist
+  // select the node n
+  // remove n from worklist and retain the order of remaining nodes
+  worklist.remove((uint)idx);
   return n;
 }
 

src/share/vm/opto/superword.cpp

@@ -1374,6 +1374,20 @@ void SuperWord::output() {
       if (n->is_Load()) {
         Node* ctl = n->in(MemNode::Control);
         Node* mem = first->in(MemNode::Memory);
+        SWPointer p1(n->as_Mem(), this);
+        // Identify the memory dependency for the new loadVector node by
+        // walking up through memory chain.
+        // This is done to give flexibility to the new loadVector node so that
+        // it can move above independent storeVector nodes.
+        while (mem->is_StoreVector()) {
+          SWPointer p2(mem->as_Mem(), this);
+          int cmp = p1.cmp(p2);
+          if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
+            mem = mem->in(MemNode::Memory);
+          } else {
+            break; // dependent memory
+          }
+        }
         Node* adr = low_adr->in(MemNode::Address);
         const TypePtr* atyp = n->adr_type();
         vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));