This change fixes a potential problem in unwinding on Linux. PUSH and…

… POP instructions were used to preserve RSI/RDI on the stack when stosd instructions are used to initialize the stack vars. These registers are used as the first two parameters to pass parameters to a callee.

The change makes sure there is a FrameRegisterUsed if these PUSH/POPs need to occur. If there is no FrameRegister, the unwinding, if attempted between the first PUSH and the last POP would cause incorrect unwinding of the stack.

SharedCodebaseChange: Yes
SharedCodebaseChangeRisk: Low

[tfs-changeset: 1409112]

src/jit/codegencommon.cpp

@@ -4821,7 +4821,15 @@ void CodeGen::genCheckUseBlockInit()
        we waste all the other slots.  Really need to compute the correct
        and compare that against zeroing the slots individually */
 
-    genUseBlockInit = (genInitStkLclCnt > (largeGcStructs + 4));
+#if defined(UNIX_AMD64_ABI)
+    // For AMD64_UNIX don't use block initialization if there is no FrameRegister.
+    // The RDI and RSI registers are in the block initialization and are also
+    // the first two parameters to a callee. 
+    // Need to push and pop them in the prolog and this will break unwinding.
+    genUseBlockInit = (genInitStkLclCnt > (largeGcStructs + 4)) && isFramePointerUsed();
+#else // UNIX_AMD64_ABI
+	genUseBlockInit = (genInitStkLclCnt > (largeGcStructs + 4)); 
+#endif // UNIX_AMD64_ABI
 
     if  (genUseBlockInit)
     {
@@ -6072,6 +6080,12 @@ void        CodeGen::genZeroInitFrame(int        untrLclHi,
 
     if  (genUseBlockInit)
     {
+#ifdef UNIX_AMD64_ABI
+        // Should not be here for Unix AMD64 if there is no Frame Pointer used.
+        // No block initialization in this case.
+        assert(isFramePointerUsed());
+#endif // UNIX_AMD64_ABI
+
         assert(untrLclHi > untrLclLo);
 #ifdef _TARGET_ARMARCH_
         /*
@@ -6266,21 +6280,12 @@ void        CodeGen::genZeroInitFrame(int        untrLclHi,
         }
 
         noway_assert((intRegState.rsCalleeRegArgMaskLiveIn & RBM_EAX) == 0);
-        int disp = untrLclLo;
-#ifdef UNIX_AMD64_ABI
-        // If there is no frame register the pushes above mess up the
-        // RSP, so adjust by adding 0x10 to the offset.
-        if (!isFramePointerUsed())
-        {
-            disp += 0x10;
-        }
-#endif // UNIX_AMD64_ABI
 
         getEmitter()->emitIns_R_AR(INS_lea,
                                  EA_PTRSIZE,
                                  REG_EDI,
                                  genFramePointerReg(),
-                                 disp);
+                                 untrLclLo);
         regTracker.rsTrackRegTrash(REG_EDI);
 
         inst_RV_IV(INS_mov, REG_ECX, (untrLclHi - untrLclLo) / sizeof(int), EA_4BYTE);
@@ -6294,8 +6299,8 @@ void        CodeGen::genZeroInitFrame(int        untrLclHi,
 
 #ifdef UNIX_AMD64_ABI
         // Restore the RDI and RSI.
-        inst_RV(INS_pop, REG_RSI, TYP_I_IMPL);
-        inst_RV(INS_pop, REG_RDI, TYP_I_IMPL);
+         inst_RV(INS_pop, REG_RSI, TYP_I_IMPL);
+         inst_RV(INS_pop, REG_RDI, TYP_I_IMPL);
 #endif // UNIX_AMD64_ABI
 #else // _TARGET_*
 #error Unsupported or unset target architecture

src/jit/lclvars.cpp

@@ -4040,10 +4040,15 @@ int Compiler::lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize
 #if defined(_TARGET_X86_)
         argOffs += sizeof(void *);
 #elif defined(_TARGET_AMD64_)
-#ifndef UNIX_AMD64_ABI
-        varDsc->lvStkOffs = argOffs;
-        argOffs += sizeof(void *);
-#endif // !UNIX_AMD64_ABI
+#ifdef UNIX_AMD64_ABI
+        // Reserve space on the stack only for OnFrame variables.
+        // No need to do that for OutgoingArg. No such thing on Linux.
+        if (varDsc->lvOnFrame)
+#endif // UNIX_AMD64_ABI
+        {
+            varDsc->lvStkOffs = argOffs;
+            argOffs += sizeof(void *);
+        }
 #elif defined(_TARGET_ARM64_)
         // Register arguments don't take stack space.
 #elif defined(_TARGET_ARM_)

src/jit/lsra.cpp

@@ -1489,6 +1489,37 @@ LinearScan::doLinearScan()
 
     compiler->codeGen->regSet.rsClearRegsModified();
 
+#ifdef UNIX_AMD64_ABI
+    // Count the numbers of must initialize local vars. 
+    // Set the FramePointerRequired if there are more or equal to MAX_VARS_FOR_NO_FRAMEPOINTER MustInit vars.
+    // This way block initialize can be used. On Linux stosd requires RDI and RSI,
+    // which are the first 2 parameters to a callee. They need to be preserved on the stack. PUSH/POP 
+    // without FrameRegister breaks unwinding. 
+    // If more than MAX_VARS_FOR_NO_FRAMEPOINTER vars are used, the code for initializing the vars gets 
+    // big and an instruction group is not enough -
+    // (in emit.cpp there is an assert assert(emitCurIG != emitPrologIG);) multi_IG prologs are not allowed.
+    // So, set the frame to have a frame pointer and use block initialization.
+    unsigned lclNum;
+    unsigned lclMustInitCnt = 0;
+    LclVarDsc *varDsc;
+
+    for (lclNum = 0, varDsc = compiler->lvaTable;
+        lclNum < compiler->lvaCount;
+        lclNum++, varDsc++)
+    {
+        if (varDsc->lvMustInit)
+        {
+            lclMustInitCnt++;
+        }
+    }
+
+    if (lclMustInitCnt >= MAX_VARS_FOR_NO_FRAMEPOINTER)
+    {
+        compiler->codeGen->setFramePointerRequired(true);
+    }
+
+#endif // UNIX_AMD64_ABI
+
     // Figure out if we're going to use an RSP frame or an RBP frame. We need to do this
     // before building the intervals and ref positions, because those objects will embed
     // RBP in various register masks (like preferences) if RBP is allowed to be allocated.

src/jit/target.h

@@ -698,6 +698,10 @@ typedef unsigned short          regPairNoSmall; // arm: need 12 bits
   #define RBM_FLT_CALLEE_SAVED    (0)
   #define RBM_FLT_CALLEE_TRASH    (RBM_XMM0|RBM_XMM1|RBM_XMM2|RBM_XMM3|RBM_XMM4|RBM_XMM5|RBM_XMM6|RBM_XMM7| \
                                    RBM_XMM8|RBM_XMM9|RBM_XMM10|RBM_XMM11|RBM_XMM12|RBM_XMM13|RBM_XMM14|RBM_XMM15)
+  // Use this value to specify how many MustInit vars on Linux would trigger a FramePointer to be used.
+  // If there is no FramePointer blockInit in codegencommon.cpp is not used. There is a limit to the size of the 
+  // prolog (it should not exceed one IG.) Make sure we don't get in such case.
+  #define MAX_VARS_FOR_NO_FRAMEPOINTER 6
 #endif // UNIX_AMD64_ABI
 
   #define REG_FLT_CALLEE_SAVED_FIRST   REG_XMM6