Bug 522314 - Make x86 virtual stack pointer code optional, g=gal.

--HG--
extra : rebase_source : c0feec3b6f94580fab0e8569a817bf36db7ed1ab

js/src/jstracer.cpp

@@ -7475,6 +7475,7 @@ js_InitJIT(JSTraceMonitor *tm)
 #if defined NANOJIT_IA32
         avmplus::AvmCore::config.use_cmov =
             avmplus::AvmCore::config.sse2 = CheckForSSE2();
+        avmplus::AvmCore::config.fixed_esp = true;
 #endif
 #if defined NANOJIT_ARM
 

js/src/nanojit/Nativei386.cpp

@@ -184,14 +184,38 @@ namespace nanojit
             max_regs = iargs;
 
         int32_t istack = iargs-max_regs;  // first 2 4B args are in registers
-        int32_t pushsize = 4*istack + 8*fargs; // actual stack space used
+        int32_t extra = 0;
+        const int32_t pushsize = 4*istack + 8*fargs; // actual stack space used
 
-        // In case of fastcall, stdcall and thiscall the callee cleans up the stack,
-        // and since we reserve max_stk_args words in the prolog to call functions
-        // and don't adjust the stack pointer individually for each call we have
-        // to undo here any changes the callee just did to the stack.
-        if (pushsize && abi != ABI_CDECL)
-            SUBi(SP, pushsize);
+#if _MSC_VER
+        // msc only provides 4-byte alignment, anything more than 4 on windows
+        // x86-32 requires dynamic ESP alignment in prolog/epilog and static
+        // esp-alignment here.
+        uint32_t align = 4;//NJ_ALIGN_STACK;
+#else
+        uint32_t align = NJ_ALIGN_STACK;
+#endif
+
+        if (pushsize) {
+            if (config.fixed_esp) {
+                // In case of fastcall, stdcall and thiscall the callee cleans up the stack,
+                // and since we reserve max_stk_args words in the prolog to call functions
+                // and don't adjust the stack pointer individually for each call we have
+                // to undo here any changes the callee just did to the stack.
+                if (abi != ABI_CDECL)
+                    SUBi(SP, pushsize);
+            } else {
+                // stack re-alignment
+                // only pop our adjustment amount since callee pops args in FASTCALL mode
+                extra = alignUp(pushsize, align) - pushsize;
+                if (call->_abi == ABI_CDECL) {
+                    // with CDECL only, caller pops args
+                    ADDi(SP, extra+pushsize);
+                } else if (extra > 0) {
+                    ADDi(SP, extra);
+                }
+            }
+        }
 
         NanoAssert(ins->isop(LIR_pcall) || ins->isop(LIR_fcall));
         if (!indirect) {
@@ -215,10 +239,12 @@ namespace nanojit
         ArgSize sizes[MAXARGS];
         uint32_t argc = call->get_sizes(sizes);
         int32_t stkd = 0;
-
+        
         if (indirect) {
             argc--;
             asm_arg(ARGSIZE_P, ins->arg(argc), EAX, stkd);
+            if (!config.fixed_esp) 
+                stkd = 0;
         }
 
         for(uint32_t i=0; i < argc; i++)
@@ -230,10 +256,16 @@ namespace nanojit
                 r = argRegs[n++]; // tell asm_arg what reg to use
             }
             asm_arg(sz, ins->arg(j), r, stkd);
+            if (!config.fixed_esp) 
+                stkd = 0;
         }
 
-        if (pushsize > max_stk_args)
-            max_stk_args = pushsize;
+        if (config.fixed_esp) {
+            if (pushsize > max_stk_args)
+                max_stk_args = pushsize;
+        } else if (extra > 0) {
+            SUBi(SP, extra);
+        }
     }
 
     Register Assembler::nRegisterAllocFromSet(RegisterMask set)
@@ -1320,7 +1352,10 @@ namespace nanojit
                 }
             }
             else {
-                asm_stkarg(p, stkd);
+                if (config.fixed_esp)
+                    asm_stkarg(p, stkd);
+                else
+                    asm_pusharg(p);
             }
         }
         else
@@ -1330,6 +1365,29 @@ namespace nanojit
         }
     }
 
+    void Assembler::asm_pusharg(LInsp p)
+    {
+        // arg goes on stack
+        if (!p->isUsed() && p->isconst())
+        {
+            // small const we push directly
+            PUSHi(p->imm32());
+        }
+        else if (!p->isUsed() || p->isop(LIR_alloc))
+        {
+            Register ra = findRegFor(p, GpRegs);
+            PUSHr(ra);
+        }
+        else if (!p->hasKnownReg())
+        {
+            PUSHm(disp(p), FP);
+        }
+        else
+        {
+            PUSHr(p->getReg());
+        }
+    }
+
     void Assembler::asm_stkarg(LInsp p, int32_t& stkd)
     {
         // arg goes on stack
@@ -1364,6 +1422,8 @@ namespace nanojit
              */
             evictIfActive(FST0);
         }
+        if (!config.fixed_esp)
+            SUBi(ESP,8);
 
         stkd += sizeof(double);
     }

js/src/nanojit/Nativei386.h

@@ -174,7 +174,7 @@ namespace nanojit
     #define DECLARE_PLATFORM_REGALLOC()
 
     #define DECLARE_PLATFORM_ASSEMBLER()    \
-        const static Register argRegs[2], retRegs[2];\
+        const static Register argRegs[2], retRegs[2]; \
         int32_t max_stk_args;\
         void nativePageReset();\
         void nativePageSetup();\
@@ -182,8 +182,9 @@ namespace nanojit
         void asm_stkarg(LInsp p, int32_t& stkd);\
         void asm_farg(LInsp, int32_t& stkd);\
         void asm_arg(ArgSize sz, LInsp p, Register r, int32_t& stkd);\
+        void asm_pusharg(LInsp);\
         void asm_fcmp(LIns *cond);\
-        void asm_cmp(LIns *cond);\
+        void asm_cmp(LIns *cond); \
         void asm_div_mod(LIns *cond);
 
     #define swapptrs()  { NIns* _tins = _nIns; _nIns=_nExitIns; _nExitIns=_tins; }
@@ -521,6 +522,11 @@ namespace nanojit
     *(--_nIns) = (uint8_t) ( 0x50 | (r) );  \
     asm_output("push %s",gpn(r)); } while(0)
 
+#define PUSHm(d,b) do { \
+    count_pushld();\
+    ALUm(0xff, 6, d, b);        \
+    asm_output("push %d(%s)",d,gpn(b)); } while(0)
+
 #define POPr(r) do { \
     count_pop();\
     underrunProtect(1);         \

js/src/nanojit/avmplus.h

@@ -202,6 +202,8 @@ namespace avmplus {
     // Whether or not we can use SSE2 instructions and conditional moves.
         bool sse2;
         bool use_cmov;
+        // Whether to use a virtual stack pointer
+        bool fixed_esp;
 #endif
 
 #if defined (AVMPLUS_ARM)