8037226: compiler/7196199/Test7196199.java fails on 32-bit linux with MaxVectorSize > 16

Summary: verify YMM registers after signal processing and set limit on vector's size.
Reviewed-by: iveresov, twisti

src/cpu/x86/vm/vm_version_x86.cpp

@@ -50,8 +50,13 @@ int VM_Version::_cpuFeatures;
 const char*           VM_Version::_features_str = "";
 VM_Version::CpuidInfo VM_Version::_cpuid_info   = { 0, };
 
+// Address of instruction which causes SEGV
+address VM_Version::_cpuinfo_segv_addr = 0;
+// Address of instruction after the one which causes SEGV
+address VM_Version::_cpuinfo_cont_addr = 0;
+
 static BufferBlob* stub_blob;
-static const int stub_size = 550;
+static const int stub_size = 600;
 
 extern "C" {
   typedef void (*getPsrInfo_stub_t)(void*);
@@ -234,9 +239,9 @@ class VM_Version_StubGenerator: public StubCodeGenerator {
     // Check if OS has enabled XGETBV instruction to access XCR0
     // (OSXSAVE feature flag) and CPU supports AVX
     //
-    __ andl(rcx, 0x18000000);
+    __ andl(rcx, 0x18000000); // cpuid1 bits osxsave | avx
     __ cmpl(rcx, 0x18000000);
-    __ jccb(Assembler::notEqual, sef_cpuid);
+    __ jccb(Assembler::notEqual, sef_cpuid); // jump if AVX is not supported
 
     //
     // XCR0, XFEATURE_ENABLED_MASK register
@@ -247,6 +252,47 @@ class VM_Version_StubGenerator: public StubCodeGenerator {
     __ movl(Address(rsi, 0), rax);
     __ movl(Address(rsi, 4), rdx);
 
+    __ andl(rax, 0x6); // xcr0 bits sse | ymm
+    __ cmpl(rax, 0x6);
+    __ jccb(Assembler::notEqual, sef_cpuid); // jump if AVX is not supported
+
+    //
+    // Some OSs have a bug when upper 128bits of YMM
+    // registers are not restored after a signal processing.
+    // Generate SEGV here (reference through NULL)
+    // and check upper YMM bits after it.
+    //
+    VM_Version::set_avx_cpuFeatures(); // Enable temporary to pass asserts
+
+    // load value into all 32 bytes of ymm7 register
+    __ movl(rcx, VM_Version::ymm_test_value());
+
+    __ movdl(xmm0, rcx);
+    __ pshufd(xmm0, xmm0, 0x00);
+    __ vinsertf128h(xmm0, xmm0, xmm0);
+    __ vmovdqu(xmm7, xmm0);
+#ifdef _LP64
+    __ vmovdqu(xmm8,  xmm0);
+    __ vmovdqu(xmm15, xmm0);
+#endif
+
+    __ xorl(rsi, rsi);
+    VM_Version::set_cpuinfo_segv_addr( __ pc() );
+    // Generate SEGV
+    __ movl(rax, Address(rsi, 0));
+
+    VM_Version::set_cpuinfo_cont_addr( __ pc() );
+    // Returns here after signal. Save xmm0 to check it later.
+    __ lea(rsi, Address(rbp, in_bytes(VM_Version::ymm_save_offset())));
+    __ vmovdqu(Address(rsi,  0), xmm0);
+    __ vmovdqu(Address(rsi, 32), xmm7);
+#ifdef _LP64
+    __ vmovdqu(Address(rsi, 64), xmm8);
+    __ vmovdqu(Address(rsi, 96), xmm15);
+#endif
+
+    VM_Version::clean_cpuFeatures();
+
     //
     // cpuid(0x7) Structured Extended Features
     //
@@ -540,14 +586,28 @@ void VM_Version::get_processor_features() {
     if (MaxVectorSize > 32) {
       FLAG_SET_DEFAULT(MaxVectorSize, 32);
     }
-    if (MaxVectorSize > 16 && UseAVX == 0) {
-      // Only supported with AVX+
+    if (MaxVectorSize > 16 && (UseAVX == 0 || !os_supports_avx_vectors())) {
+      // 32 bytes vectors (in YMM) are only supported with AVX+
       FLAG_SET_DEFAULT(MaxVectorSize, 16);
     }
     if (UseSSE < 2) {
-      // Only supported with SSE2+
+      // Vectors (in XMM) are only supported with SSE2+
       FLAG_SET_DEFAULT(MaxVectorSize, 0);
     }
+#ifdef ASSERT
+    if (supports_avx() && PrintMiscellaneous && Verbose && TraceNewVectors) {
+      tty->print_cr("State of YMM registers after signal handle:");
+      int nreg = 2 LP64_ONLY(+2);
+      const char* ymm_name[4] = {"0", "7", "8", "15"};
+      for (int i = 0; i < nreg; i++) {
+        tty->print("YMM%s:", ymm_name[i]);
+        for (int j = 7; j >=0; j--) {
+          tty->print(" %x", _cpuid_info.ymm_save[i*8 + j]);
+        }
+        tty->cr();
+      }
+    }
+#endif
   }
 #endif
 
@@ -678,14 +738,6 @@ void VM_Version::get_processor_features() {
       }
     }
   }
-#if defined(COMPILER2) && defined(_ALLBSD_SOURCE)
-    if (MaxVectorSize > 16) {
-      // Limit vectors size to 16 bytes on BSD until it fixes
-      // restoring upper 128bit of YMM registers on return
-      // from signal handler.
-      FLAG_SET_DEFAULT(MaxVectorSize, 16);
-    }
-#endif // COMPILER2
 
   // Use count leading zeros count instruction if available.
   if (supports_lzcnt()) {
@@ -814,6 +866,11 @@ void VM_Version::get_processor_features() {
     if (UseAES) {
       tty->print("  UseAES=1");
     }
+#ifdef COMPILER2
+    if (MaxVectorSize > 0) {
+      tty->print("  MaxVectorSize=%d", MaxVectorSize);
+    }
+#endif
     tty->cr();
     tty->print("Allocation");
     if (AllocatePrefetchStyle <= 0 || UseSSE == 0 && !supports_3dnow_prefetch()) {

src/cpu/x86/vm/vm_version_x86.hpp

@@ -229,6 +229,9 @@ protected:
                                // 0 if this instruction is not available
   static const char* _features_str;
 
+  static address   _cpuinfo_segv_addr; // address of instruction which causes SEGV
+  static address   _cpuinfo_cont_addr; // address of instruction after the one which causes SEGV
+
   enum {
     CPU_CX8    = (1 << 0), // next bits are from cpuid 1 (EDX)
     CPU_CMOV   = (1 << 1),
@@ -361,6 +364,9 @@ protected:
     // extended control register XCR0 (the XFEATURE_ENABLED_MASK register)
     XemXcr0Eax   xem_xcr0_eax;
     uint32_t     xem_xcr0_edx; // reserved
+
+    // Space to save ymm registers after signal handle
+    int          ymm_save[8*4]; // Save ymm0, ymm7, ymm8, ymm15
   };
 
   // The actual cpuid info block
@@ -460,6 +466,21 @@ protected:
     return result;
   }
 
+  static bool os_supports_avx_vectors() {
+    if (!supports_avx()) {
+      return false;
+    }
+    // Verify that OS save/restore all bits of AVX registers
+    // during signal processing.
+    int nreg = 2 LP64_ONLY(+2);
+    for (int i = 0; i < 8 * nreg; i++) { // 32 bytes per ymm register
+      if (_cpuid_info.ymm_save[i] != ymm_test_value()) {
+        return false;
+      }
+    }
+    return true;
+  }
+
   static void get_processor_features();
 
 public:
@@ -476,6 +497,19 @@ public:
   static ByteSize tpl_cpuidB1_offset() { return byte_offset_of(CpuidInfo, tpl_cpuidB1_eax); }
   static ByteSize tpl_cpuidB2_offset() { return byte_offset_of(CpuidInfo, tpl_cpuidB2_eax); }
   static ByteSize xem_xcr0_offset() { return byte_offset_of(CpuidInfo, xem_xcr0_eax); }
+  static ByteSize ymm_save_offset() { return byte_offset_of(CpuidInfo, ymm_save); }
+
+  // The value used to check ymm register after signal handle
+  static int ymm_test_value()    { return 0xCAFEBABE; }
+
+  static void set_cpuinfo_segv_addr(address pc) { _cpuinfo_segv_addr = pc; }
+  static bool  is_cpuinfo_segv_addr(address pc) { return _cpuinfo_segv_addr == pc; }
+  static void set_cpuinfo_cont_addr(address pc) { _cpuinfo_cont_addr = pc; }
+  static address  cpuinfo_cont_addr()           { return _cpuinfo_cont_addr; }
+
+  static void clean_cpuFeatures()   { _cpuFeatures = 0; }
+  static void set_avx_cpuFeatures() { _cpuFeatures = (CPU_SSE | CPU_SSE2 | CPU_AVX); }
+
 
   // Initialization
   static void initialize();

src/os/windows/vm/os_windows.cpp

@@ -2437,6 +2437,12 @@ LONG WINAPI topLevelExceptionFilter(struct _EXCEPTION_POINTERS* exceptionInfo) {
     }
   }
 
+  if ((exception_code == EXCEPTION_ACCESS_VIOLATION) &&
+      VM_Version::is_cpuinfo_segv_addr(pc)) {
+    // Verify that OS save/restore AVX registers.
+    return Handle_Exception(exceptionInfo, VM_Version::cpuinfo_cont_addr());
+  }
+
   if (t != NULL && t->is_Java_thread()) {
     JavaThread* thread = (JavaThread*) t;
     bool in_java = thread->thread_state() == _thread_in_Java;

src/os_cpu/bsd_x86/vm/os_bsd_x86.cpp

@@ -492,6 +492,11 @@ JVM_handle_bsd_signal(int sig,
       }
     }
 
+    if ((sig == SIGSEGV || sig == SIGBUS) && VM_Version::is_cpuinfo_segv_addr(pc)) {
+      // Verify that OS save/restore AVX registers.
+      stub = VM_Version::cpuinfo_cont_addr();
+    }
+
     // We test if stub is already set (by the stack overflow code
     // above) so it is not overwritten by the code that follows. This
     // check is not required on other platforms, because on other

src/os_cpu/linux_x86/vm/os_linux_x86.cpp

@@ -337,6 +337,11 @@ JVM_handle_linux_signal(int sig,
       }
     }
 
+    if ((sig == SIGSEGV) && VM_Version::is_cpuinfo_segv_addr(pc)) {
+      // Verify that OS save/restore AVX registers.
+      stub = VM_Version::cpuinfo_cont_addr();
+    }
+
     if (thread->thread_state() == _thread_in_Java) {
       // Java thread running in Java code => find exception handler if any
       // a fault inside compiled code, the interpreter, or a stub

src/os_cpu/solaris_x86/vm/os_solaris_x86.cpp

@@ -459,6 +459,11 @@ JVM_handle_solaris_signal(int sig, siginfo_t* info, void* ucVoid,
       }
     }
 
+    if ((sig == SIGSEGV) && VM_Version::is_cpuinfo_segv_addr(pc)) {
+      // Verify that OS save/restore AVX registers.
+      stub = VM_Version::cpuinfo_cont_addr();
+    }
+
     if (thread->thread_state() == _thread_in_vm) {
       if (sig == SIGBUS && info->si_code == BUS_OBJERR && thread->doing_unsafe_access()) {
         stub = StubRoutines::handler_for_unsafe_access();