Merge branch 'devel' into araq

1 files changed, 1324 insertions, 514 deletions

diff --git a/tinyc/x86_64-gen.c b/tinyc/x86_64-gen.c
index 1227f41f0..f4d86e1b7 100644
--- a/tinyc/x86_64-gen.c
+++ b/tinyc/x86_64-gen.c
@@ -20,10 +20,12 @@
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */
 
-#include <assert.h>
+#ifdef TARGET_DEFS_ONLY
 
 /* number of available registers */
-#define NB_REGS             5
+#define NB_REGS         25
+#define NB_ASM_REGS     16
+#define CONFIG_TCC_ASM
 
 /* a register can belong to several classes. The classes must be
    sorted from more general to more precise (see gv2() code which does
@@ -33,45 +35,60 @@
 #define RC_RAX     0x0004
 #define RC_RCX     0x0008
 #define RC_RDX     0x0010
-#define RC_XMM0    0x0020
-#define RC_ST0     0x0040 /* only for long double */
+#define RC_ST0     0x0080 /* only for long double */
+#define RC_R8      0x0100
+#define RC_R9      0x0200
+#define RC_R10     0x0400
+#define RC_R11     0x0800
+#define RC_XMM0    0x1000
+#define RC_XMM1    0x2000
+#define RC_XMM2    0x4000
+#define RC_XMM3    0x8000
+#define RC_XMM4    0x10000
+#define RC_XMM5    0x20000
+#define RC_XMM6    0x40000
+#define RC_XMM7    0x80000
 #define RC_IRET    RC_RAX /* function return: integer register */
 #define RC_LRET    RC_RDX /* function return: second integer register */
 #define RC_FRET    RC_XMM0 /* function return: float register */
+#define RC_QRET    RC_XMM1 /* function return: second float register */
 
 /* pretty names for the registers */
 enum {
     TREG_RAX = 0,
     TREG_RCX = 1,
     TREG_RDX = 2,
+    TREG_RSP = 4,
     TREG_RSI = 6,
     TREG_RDI = 7,
+
     TREG_R8  = 8,
     TREG_R9  = 9,
     TREG_R10 = 10,
     TREG_R11 = 11,
 
-    TREG_XMM0 = 3,
-    TREG_ST0 = 4,
+    TREG_XMM0 = 16,
+    TREG_XMM1 = 17,
+    TREG_XMM2 = 18,
+    TREG_XMM3 = 19,
+    TREG_XMM4 = 20,
+    TREG_XMM5 = 21,
+    TREG_XMM6 = 22,
+    TREG_XMM7 = 23,
 
-    TREG_MEM = 0x10,
+    TREG_ST0 = 24,
+
+    TREG_MEM = 0x20
 };
 
 #define REX_BASE(reg) (((reg) >> 3) & 1)
 #define REG_VALUE(reg) ((reg) & 7)
 
-int reg_classes[NB_REGS] = {
-    /* eax */ RC_INT | RC_RAX,
-    /* ecx */ RC_INT | RC_RCX,
-    /* edx */ RC_INT | RC_RDX,
-    /* xmm0 */ RC_FLOAT | RC_XMM0,
-    /* st0 */ RC_ST0,
-};
-
 /* return registers for function */
 #define REG_IRET TREG_RAX /* single word int return register */
 #define REG_LRET TREG_RDX /* second word return register (for long long) */
 #define REG_FRET TREG_XMM0 /* float return register */
+#define REG_QRET TREG_XMM1 /* second float return register */
 
 /* defined if function parameters must be evaluated in reverse order */
 #define INVERT_FUNC_PARAMS
@@ -81,32 +98,56 @@ int reg_classes[NB_REGS] = {
 
 /* long double size and alignment, in bytes */
 #define LDOUBLE_SIZE  16
-#define LDOUBLE_ALIGN 8
+#define LDOUBLE_ALIGN 16
 /* maximum alignment (for aligned attribute support) */
-#define MAX_ALIGN     8
+#define MAX_ALIGN     16
 
 /******************************************************/
-/* ELF defines */
-
-#define EM_TCC_TARGET EM_X86_64
-
-/* relocation type for 32 bit data relocation */
-#define R_DATA_32   R_X86_64_64
-#define R_JMP_SLOT  R_X86_64_JUMP_SLOT
-#define R_COPY      R_X86_64_COPY
-
-#define ELF_START_ADDR 0x08048000
-#define ELF_PAGE_SIZE  0x1000
-
+#else /* ! TARGET_DEFS_ONLY */
 /******************************************************/
+#include "tcc.h"
+#include <assert.h>
+
+ST_DATA const int reg_classes[NB_REGS] = {
+    /* eax */ RC_INT | RC_RAX,
+    /* ecx */ RC_INT | RC_RCX,
+    /* edx */ RC_INT | RC_RDX,
+    0,
+    0,
+    0,
+    0,
+    0,
+    RC_R8,
+    RC_R9,
+    RC_R10,
+    RC_R11,
+    0,
+    0,
+    0,
+    0,
+    /* xmm0 */ RC_FLOAT | RC_XMM0,
+    /* xmm1 */ RC_FLOAT | RC_XMM1,
+    /* xmm2 */ RC_FLOAT | RC_XMM2,
+    /* xmm3 */ RC_FLOAT | RC_XMM3,
+    /* xmm4 */ RC_FLOAT | RC_XMM4,
+    /* xmm5 */ RC_FLOAT | RC_XMM5,
+    /* xmm6 an xmm7 are included so gv() can be used on them,
+       but they are not tagged with RC_FLOAT because they are
+       callee saved on Windows */
+    RC_XMM6,
+    RC_XMM7,
+    /* st0 */ RC_ST0
+};
 
 static unsigned long func_sub_sp_offset;
 static int func_ret_sub;
 
 /* XXX: make it faster ? */
-void g(int c)
+ST_FUNC void g(int c)
 {
     int ind1;
+    if (nocode_wanted)
+        return;
     ind1 = ind + 1;
     if (ind1 > cur_text_section->data_allocated)
         section_realloc(cur_text_section, ind1);
@@ -114,7 +155,7 @@ void g(int c)
     ind = ind1;
 }
 
-void o(unsigned int c)
+ST_FUNC void o(unsigned int c)
 {
     while (c) {
         g(c);
@@ -122,7 +163,13 @@ void o(unsigned int c)
     }
 }
 
-void gen_le32(int c)
+ST_FUNC void gen_le16(int v)
+{
+    g(v);
+    g(v >> 8);
+}
+
+ST_FUNC void gen_le32(int c)
 {
     g(c);
     g(c >> 8);
@@ -130,7 +177,7 @@ void gen_le32(int c)
     g(c >> 24);
 }
 
-void gen_le64(int64_t c)
+ST_FUNC void gen_le64(int64_t c)
 {
     g(c);
     g(c >> 8);
@@ -142,14 +189,24 @@ void gen_le64(int64_t c)
     g(c >> 56);
 }
 
+static void orex(int ll, int r, int r2, int b)
+{
+    if ((r & VT_VALMASK) >= VT_CONST)
+        r = 0;
+    if ((r2 & VT_VALMASK) >= VT_CONST)
+        r2 = 0;
+    if (ll || REX_BASE(r) || REX_BASE(r2))
+        o(0x40 | REX_BASE(r) | (REX_BASE(r2) << 2) | (ll << 3));
+    o(b);
+}
+
 /* output a symbol and patch all calls to it */
-void gsym_addr(int t, int a)
+ST_FUNC void gsym_addr(int t, int a)
 {
-    int n, *ptr;
     while (t) {
-        ptr = (int *)(cur_text_section->data + t);
-        n = *ptr; /* next value */
-        *ptr = a - t - 4;
+        unsigned char *ptr = cur_text_section->data + t;
+        uint32_t n = read32le(ptr); /* next value */
+        write32le(ptr, a - t - 4);
         t = n;
     }
 }
@@ -159,9 +216,6 @@ void gsym(int t)
     gsym_addr(t, ind);
 }
 
-/* psym is used to put an instruction with a data field which is a
-   reference to a symbol. It is in fact the same as oad ! */
-#define psym oad
 
 static int is64_type(int t)
 {
@@ -170,58 +224,59 @@ static int is64_type(int t)
             (t & VT_BTYPE) == VT_LLONG);
 }
 
-static int is_sse_float(int t) {
-    int bt;
-    bt = t & VT_BTYPE;
-    return bt == VT_DOUBLE || bt == VT_FLOAT;
-}
-
 /* instruction + 4 bytes data. Return the address of the data */
 static int oad(int c, int s)
 {
-    int ind1;
-
+    int t;
+    if (nocode_wanted)
+        return s;
     o(c);
-    ind1 = ind + 4;
-    if (ind1 > cur_text_section->data_allocated)
-        section_realloc(cur_text_section, ind1);
-    *(int *)(cur_text_section->data + ind) = s;
-    s = ind;
-    ind = ind1;
-    return s;
+    t = ind;
+    gen_le32(s);
+    return t;
+}
+
+/* generate jmp to a label */
+#define gjmp2(instr,lbl) oad(instr,lbl)
+
+ST_FUNC void gen_addr32(int r, Sym *sym, int c)
+{
+    if (r & VT_SYM)
+        greloca(cur_text_section, sym, ind, R_X86_64_32S, c), c=0;
+    gen_le32(c);
 }
 
 /* output constant with relocation if 'r & VT_SYM' is true */
-static void gen_addr64(int r, Sym *sym, int64_t c)
+ST_FUNC void gen_addr64(int r, Sym *sym, int64_t c)
 {
     if (r & VT_SYM)
-        greloc(cur_text_section, sym, ind, R_X86_64_64);
+        greloca(cur_text_section, sym, ind, R_X86_64_64, c), c=0;
     gen_le64(c);
 }
 
 /* output constant with relocation if 'r & VT_SYM' is true */
-static void gen_addrpc32(int r, Sym *sym, int c)
+ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 {
-    if (r & VT_SYM)
-        greloc(cur_text_section, sym, ind, R_X86_64_PC32);
+    greloca(cur_text_section, sym, ind, R_X86_64_PC32, 0);
     gen_le32(c-4);
 }
 
 /* output got address with relocation */
 static void gen_gotpcrel(int r, Sym *sym, int c)
 {
-    Section *sr;
-    ElfW(Rela) *rel;
-    greloc(cur_text_section, sym, ind, R_X86_64_GOTPCREL);
-    sr = cur_text_section->reloc;
-    rel = (ElfW(Rela) *)(sr->data + sr->data_offset - sizeof(ElfW(Rela)));
-    rel->r_addend = -4;
+#ifdef TCC_TARGET_PE
+    tcc_error("internal error: no GOT on PE: %s %x %x | %02x %02x %02x\n",
+        get_tok_str(sym->v, NULL), c, r,
+        cur_text_section->data[ind-3],
+        cur_text_section->data[ind-2],
+        cur_text_section->data[ind-1]
+        );
+#endif
+    greloca(cur_text_section, sym, ind, R_X86_64_GOTPCREL, -4);
     gen_le32(0);
-
     if (c) {
         /* we use add c, %xxx for displacement */
-        o(0x48 + REX_BASE(r));
-        o(0x81);
+        orex(1, r, 0, 0x81);
         o(0xc0 + REG_VALUE(r));
         gen_le32(c);
     }
@@ -255,30 +310,24 @@ static void gen_modrm_impl(int op_reg, int r, Sym *sym, int c, int is_got)
             g(0x00 | op_reg | REG_VALUE(r));
         }
     } else {
-        g(0x00 | op_reg | (r & VT_VALMASK));
+        g(0x00 | op_reg | REG_VALUE(r));
     }
 }
 
-/* generate a modrm reference. 'op_reg' contains the addtionnal 3
+/* generate a modrm reference. 'op_reg' contains the additional 3
    opcode bits */
 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 {
     gen_modrm_impl(op_reg, r, sym, c, 0);
 }
 
-/* generate a modrm reference. 'op_reg' contains the addtionnal 3
+/* generate a modrm reference. 'op_reg' contains the additional 3
    opcode bits */
 static void gen_modrm64(int opcode, int op_reg, int r, Sym *sym, int c)
 {
     int is_got;
-    int rex = 0x48 | (REX_BASE(op_reg) << 2);
-    if ((r & VT_VALMASK) != VT_CONST &&
-        (r & VT_VALMASK) != VT_LOCAL) {
-        rex |= REX_BASE(VT_VALMASK & r);
-    }
-    o(rex);
-    o(opcode);
     is_got = (op_reg & TREG_MEM) && !(sym->type.t & VT_STATIC);
+    orex(1, r, op_reg, opcode);
     gen_modrm_impl(op_reg, r, sym, c, is_got);
 }
 
@@ -289,10 +338,20 @@ void load(int r, SValue *sv)
     int v, t, ft, fc, fr;
     SValue v1;
 
+#ifdef TCC_TARGET_PE
+    SValue v2;
+    sv = pe_getimport(sv, &v2);
+#endif
+
     fr = sv->r;
-    ft = sv->type.t;
-    fc = sv->c.ul;
+    ft = sv->type.t & ~VT_DEFSIGN;
+    fc = sv->c.i;
+    if (fc != sv->c.i && (fr & VT_SYM))
+      tcc_error("64 bit addend in load");
 
+    ft &= ~(VT_VOLATILE | VT_CONSTANT);
+
+#ifndef TCC_TARGET_PE
     /* we use indirect access via got */
     if ((fr & VT_VALMASK) == VT_CONST && (fr & VT_SYM) &&
         (fr & VT_LVAL) && !(sv->sym->type.t & VT_STATIC)) {
@@ -307,97 +366,151 @@ void load(int r, SValue *sv)
         /* load from the temporal register */
         fr = tr | VT_LVAL;
     }
+#endif
 
     v = fr & VT_VALMASK;
     if (fr & VT_LVAL) {
+        int b, ll;
         if (v == VT_LLOCAL) {
             v1.type.t = VT_PTR;
             v1.r = VT_LOCAL | VT_LVAL;
-            v1.c.ul = fc;
-            load(r, &v1);
+            v1.c.i = fc;
             fr = r;
+            if (!(reg_classes[fr] & (RC_INT|RC_R11)))
+                fr = get_reg(RC_INT);
+            load(fr, &v1);
         }
+        ll = 0;
+	/* Like GCC we can load from small enough properly sized
+	   structs and unions as well.
+	   XXX maybe move to generic operand handling, but should
+	   occur only with asm, so tccasm.c might also be a better place */
+	if ((ft & VT_BTYPE) == VT_STRUCT) {
+	    int align;
+	    switch (type_size(&sv->type, &align)) {
+		case 1: ft = VT_BYTE; break;
+		case 2: ft = VT_SHORT; break;
+		case 4: ft = VT_INT; break;
+		case 8: ft = VT_LLONG; break;
+		default:
+		    tcc_error("invalid aggregate type for register load");
+		    break;
+	    }
+	}
         if ((ft & VT_BTYPE) == VT_FLOAT) {
-            o(0x6e0f66); /* movd */
-            r = 0;
+            b = 0x6e0f66;
+            r = REG_VALUE(r); /* movd */
         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
-            o(0x7e0ff3); /* movq */
-            r = 0;
+            b = 0x7e0ff3; /* movq */
+            r = REG_VALUE(r);
         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
-            o(0xdb); /* fldt */
-            r = 5;
-        } else if ((ft & VT_TYPE) == VT_BYTE) {
-            o(0xbe0f);   /* movsbl */
+            b = 0xdb, r = 5; /* fldt */
+        } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
+            b = 0xbe0f;   /* movsbl */
         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
-            o(0xb60f);   /* movzbl */
+            b = 0xb60f;   /* movzbl */
         } else if ((ft & VT_TYPE) == VT_SHORT) {
-            o(0xbf0f);   /* movswl */
+            b = 0xbf0f;   /* movswl */
         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
-            o(0xb70f);   /* movzwl */
-        } else if (is64_type(ft)) {
-            gen_modrm64(0x8b, r, fr, sv->sym, fc);
-            return;
+            b = 0xb70f;   /* movzwl */
+        } else {
+            assert(((ft & VT_BTYPE) == VT_INT)
+                   || ((ft & VT_BTYPE) == VT_LLONG)
+                   || ((ft & VT_BTYPE) == VT_PTR)
+                   || ((ft & VT_BTYPE) == VT_FUNC)
+                );
+            ll = is64_type(ft);
+            b = 0x8b;
+        }
+        if (ll) {
+            gen_modrm64(b, r, fr, sv->sym, fc);
         } else {
-            o(0x8b);   /* movl */
+            orex(ll, fr, r, b);
+            gen_modrm(r, fr, sv->sym, fc);
         }
-        gen_modrm(r, fr, sv->sym, fc);
     } else {
         if (v == VT_CONST) {
-            if ((ft & VT_BTYPE) == VT_LLONG) {
-                assert(!(fr & VT_SYM));
-                o(0x48);
-                o(0xb8 + REG_VALUE(r)); /* mov $xx, r */
-                gen_addr64(fr, sv->sym, sv->c.ull);
-            } else {
-                if (fr & VT_SYM) {
-                    if (sv->sym->type.t & VT_STATIC) {
-                        o(0x8d48);
-                        o(0x05 + REG_VALUE(r) * 8); /* lea xx(%rip), r */
-                        gen_addrpc32(fr, sv->sym, fc);
-                    } else {
-                        o(0x8b48);
-                        o(0x05 + REG_VALUE(r) * 8); /* mov xx(%rip), r */
-                        gen_gotpcrel(r, sv->sym, fc);
-                    }
+            if (fr & VT_SYM) {
+#ifdef TCC_TARGET_PE
+                orex(1,0,r,0x8d);
+                o(0x05 + REG_VALUE(r) * 8); /* lea xx(%rip), r */
+                gen_addrpc32(fr, sv->sym, fc);
+#else
+                if (sv->sym->type.t & VT_STATIC) {
+                    orex(1,0,r,0x8d);
+                    o(0x05 + REG_VALUE(r) * 8); /* lea xx(%rip), r */
+                    gen_addrpc32(fr, sv->sym, fc);
                 } else {
-                    o(0xb8 + REG_VALUE(r)); /* mov $xx, r */
-                    gen_le32(fc);
+                    orex(1,0,r,0x8b);
+                    o(0x05 + REG_VALUE(r) * 8); /* mov xx(%rip), r */
+                    gen_gotpcrel(r, sv->sym, fc);
                 }
+#endif
+            } else if (is64_type(ft)) {
+                orex(1,r,0, 0xb8 + REG_VALUE(r)); /* mov $xx, r */
+                gen_le64(sv->c.i);
+            } else {
+                orex(0,r,0, 0xb8 + REG_VALUE(r)); /* mov $xx, r */
+                gen_le32(fc);
             }
         } else if (v == VT_LOCAL) {
-            o(0x48 | REX_BASE(r));
-            o(0x8d); /* lea xxx(%ebp), r */
+            orex(1,0,r,0x8d); /* lea xxx(%ebp), r */
             gen_modrm(r, VT_LOCAL, sv->sym, fc);
         } else if (v == VT_CMP) {
-            oad(0xb8 + r, 0); /* mov $0, r */
-            o(0x0f); /* setxx %br */
+            orex(0,r,0,0);
+	    if ((fc & ~0x100) != TOK_NE)
+              oad(0xb8 + REG_VALUE(r), 0); /* mov $0, r */
+	    else
+              oad(0xb8 + REG_VALUE(r), 1); /* mov $1, r */
+	    if (fc & 0x100)
+	      {
+	        /* This was a float compare.  If the parity bit is
+		   set the result was unordered, meaning false for everything
+		   except TOK_NE, and true for TOK_NE.  */
+		fc &= ~0x100;
+		o(0x037a + (REX_BASE(r) << 8));
+	      }
+            orex(0,r,0, 0x0f); /* setxx %br */
             o(fc);
-            o(0xc0 + r);
+            o(0xc0 + REG_VALUE(r));
         } else if (v == VT_JMP || v == VT_JMPI) {
             t = v & 1;
-            oad(0xb8 + r, t); /* mov $1, r */
-            o(0x05eb); /* jmp after */
+            orex(0,r,0,0);
+            oad(0xb8 + REG_VALUE(r), t); /* mov $1, r */
+            o(0x05eb + (REX_BASE(r) << 8)); /* jmp after */
             gsym(fc);
-            oad(0xb8 + r, t ^ 1); /* mov $0, r */
+            orex(0,r,0,0);
+            oad(0xb8 + REG_VALUE(r), t ^ 1); /* mov $0, r */
         } else if (v != r) {
-            if (r == TREG_XMM0) {
-                assert(v == TREG_ST0);
-                /* gen_cvt_ftof(VT_DOUBLE); */
-                o(0xf0245cdd); /* fstpl -0x10(%rsp) */
-                /* movsd -0x10(%rsp),%xmm0 */
-                o(0x44100ff2);
-                o(0xf024);
+            if ((r >= TREG_XMM0) && (r <= TREG_XMM7)) {
+                if (v == TREG_ST0) {
+                    /* gen_cvt_ftof(VT_DOUBLE); */
+                    o(0xf0245cdd); /* fstpl -0x10(%rsp) */
+                    /* movsd -0x10(%rsp),%xmmN */
+                    o(0x100ff2);
+                    o(0x44 + REG_VALUE(r)*8); /* %xmmN */
+                    o(0xf024);
+                } else {
+                    assert((v >= TREG_XMM0) && (v <= TREG_XMM7));
+                    if ((ft & VT_BTYPE) == VT_FLOAT) {
+                        o(0x100ff3);
+                    } else {
+                        assert((ft & VT_BTYPE) == VT_DOUBLE);
+                        o(0x100ff2);
+                    }
+                    o(0xc0 + REG_VALUE(v) + REG_VALUE(r)*8);
+                }
             } else if (r == TREG_ST0) {
-                assert(v == TREG_XMM0);
+                assert((v >= TREG_XMM0) && (v <= TREG_XMM7));
                 /* gen_cvt_ftof(VT_LDOUBLE); */
-                /* movsd %xmm0,-0x10(%rsp) */
-                o(0x44110ff2);
+                /* movsd %xmmN,-0x10(%rsp) */
+                o(0x110ff2);
+                o(0x44 + REG_VALUE(r)*8); /* %xmmN */
                 o(0xf024);
                 o(0xf02444dd); /* fldl -0x10(%rsp) */
             } else {
-                o(0x48 | REX_BASE(r) | (REX_BASE(v) << 2));
-                o(0x89);
-                o(0xc0 + r + v * 8); /* mov v, r */
+                orex(1,r,v, 0x89);
+                o(0xc0 + REG_VALUE(r) + REG_VALUE(v) * 8); /* mov v, r */
             }
         }
     }
@@ -411,30 +524,40 @@ void store(int r, SValue *v)
     /* store the REX prefix in this variable when PIC is enabled */
     int pic = 0;
 
-    ft = v->type.t;
-    fc = v->c.ul;
+#ifdef TCC_TARGET_PE
+    SValue v2;
+    v = pe_getimport(v, &v2);
+#endif
+
     fr = v->r & VT_VALMASK;
+    ft = v->type.t;
+    fc = v->c.i;
+    if (fc != v->c.i && (fr & VT_SYM))
+      tcc_error("64 bit addend in store");
+    ft &= ~(VT_VOLATILE | VT_CONSTANT);
     bt = ft & VT_BTYPE;
 
+#ifndef TCC_TARGET_PE
     /* we need to access the variable via got */
     if (fr == VT_CONST && (v->r & VT_SYM)) {
         /* mov xx(%rip), %r11 */
         o(0x1d8b4c);
-        gen_gotpcrel(TREG_R11, v->sym, v->c.ul);
+        gen_gotpcrel(TREG_R11, v->sym, v->c.i);
         pic = is64_type(bt) ? 0x49 : 0x41;
     }
+#endif
 
     /* XXX: incorrect if float reg to reg */
     if (bt == VT_FLOAT) {
         o(0x66);
         o(pic);
         o(0x7e0f); /* movd */
-        r = 0;
+        r = REG_VALUE(r);
     } else if (bt == VT_DOUBLE) {
         o(0x66);
         o(pic);
         o(0xd60f); /* movq */
-        r = 0;
+        r = REG_VALUE(r);
     } else if (bt == VT_LDOUBLE) {
         o(0xc0d9); /* fld %st(0) */
         o(pic);
@@ -445,11 +568,11 @@ void store(int r, SValue *v)
             o(0x66);
         o(pic);
         if (bt == VT_BYTE || bt == VT_BOOL)
-            o(0x88);
+            orex(0, 0, r, 0x88);
         else if (is64_type(bt))
             op64 = 0x89;
         else
-            o(0x89);
+            orex(0, 0, r, 0x89);
     }
     if (pic) {
         /* xxx r, (%r11) where xxx is mov, movq, fld, or etc */
@@ -457,9 +580,7 @@ void store(int r, SValue *v)
             o(op64);
         o(3 + (r << 3));
     } else if (op64) {
-        if (fr == VT_CONST ||
-            fr == VT_LOCAL ||
-            (v->r & VT_LVAL)) {
+        if (fr == VT_CONST || fr == VT_LOCAL || (v->r & VT_LVAL)) {
             gen_modrm64(op64, r, v->r, v->sym, fc);
         } else if (fr != r) {
             /* XXX: don't we really come here? */
@@ -467,9 +588,7 @@ void store(int r, SValue *v)
             o(0xc0 + fr + r * 8); /* mov r, fr */
         }
     } else {
-        if (fr == VT_CONST ||
-            fr == VT_LOCAL ||
-            (v->r & VT_LVAL)) {
+        if (fr == VT_CONST || fr == VT_LOCAL || (v->r & VT_LVAL)) {
             gen_modrm(r, v->r, v->sym, fc);
         } else if (fr != r) {
             /* XXX: don't we really come here? */
@@ -479,32 +598,26 @@ void store(int r, SValue *v)
     }
 }
 
-static void gadd_sp(int val)
-{
-    if (val == (char)val) {
-        o(0xc48348);
-        g(val);
-    } else {
-        oad(0xc48148, val); /* add $xxx, %rsp */
-    }
-}
-
 /* 'is_jmp' is '1' if it is a jump */
 static void gcall_or_jmp(int is_jmp)
 {
     int r;
-    if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
+    if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST &&
+	((vtop->r & VT_SYM) || (vtop->c.i-4) == (int)(vtop->c.i-4))) {
         /* constant case */
         if (vtop->r & VT_SYM) {
             /* relocation case */
-            greloc(cur_text_section, vtop->sym,
-                   ind + 1, R_X86_64_PC32);
+#ifdef TCC_TARGET_PE
+            greloca(cur_text_section, vtop->sym, ind + 1, R_X86_64_PC32, (int)(vtop->c.i-4));
+#else
+            greloca(cur_text_section, vtop->sym, ind + 1, R_X86_64_PLT32, (int)(vtop->c.i-4));
+#endif
         } else {
             /* put an empty PC32 relocation */
-            put_elf_reloc(symtab_section, cur_text_section,
-                          ind + 1, R_X86_64_PC32, 0);
+            put_elf_reloca(symtab_section, cur_text_section,
+                          ind + 1, R_X86_64_PC32, 0, (int)(vtop->c.i-4));
         }
-        oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
+        oad(0xe8 + is_jmp, 0); /* call/jmp im */
     } else {
         /* otherwise, indirect call */
         r = TREG_R11;
@@ -515,142 +628,763 @@ static void gcall_or_jmp(int is_jmp)
     }
 }
 
-static uint8_t arg_regs[6] = {
-    TREG_RDI, TREG_RSI, TREG_RDX, TREG_RCX, TREG_R8, TREG_R9
+#if defined(CONFIG_TCC_BCHECK)
+#ifndef TCC_TARGET_PE
+static addr_t func_bound_offset;
+static unsigned long func_bound_ind;
+#endif
+
+static void gen_static_call(int v)
+{
+    Sym *sym = external_global_sym(v, &func_old_type, 0);
+    oad(0xe8, 0);
+    greloca(cur_text_section, sym, ind-4, R_X86_64_PC32, -4);
+}
+
+/* generate a bounded pointer addition */
+ST_FUNC void gen_bounded_ptr_add(void)
+{
+    /* save all temporary registers */
+    save_regs(0);
+
+    /* prepare fast x86_64 function call */
+    gv(RC_RAX);
+    o(0xc68948); // mov  %rax,%rsi ## second arg in %rsi, this must be size
+    vtop--;
+
+    gv(RC_RAX);
+    o(0xc78948); // mov  %rax,%rdi ## first arg in %rdi, this must be ptr
+    vtop--;
+
+    /* do a fast function call */
+    gen_static_call(TOK___bound_ptr_add);
+
+    /* returned pointer is in rax */
+    vtop++;
+    vtop->r = TREG_RAX | VT_BOUNDED;
+
+
+    /* relocation offset of the bounding function call point */
+    vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(ElfW(Rela)));
+}
+
+/* patch pointer addition in vtop so that pointer dereferencing is
+   also tested */
+ST_FUNC void gen_bounded_ptr_deref(void)
+{
+    addr_t func;
+    int size, align;
+    ElfW(Rela) *rel;
+    Sym *sym;
+
+    size = 0;
+    /* XXX: put that code in generic part of tcc */
+    if (!is_float(vtop->type.t)) {
+        if (vtop->r & VT_LVAL_BYTE)
+            size = 1;
+        else if (vtop->r & VT_LVAL_SHORT)
+            size = 2;
+    }
+    if (!size)
+    size = type_size(&vtop->type, &align);
+    switch(size) {
+    case  1: func = TOK___bound_ptr_indir1; break;
+    case  2: func = TOK___bound_ptr_indir2; break;
+    case  4: func = TOK___bound_ptr_indir4; break;
+    case  8: func = TOK___bound_ptr_indir8; break;
+    case 12: func = TOK___bound_ptr_indir12; break;
+    case 16: func = TOK___bound_ptr_indir16; break;
+    default:
+        tcc_error("unhandled size when dereferencing bounded pointer");
+        func = 0;
+        break;
+    }
+
+    sym = external_global_sym(func, &func_old_type, 0);
+    if (!sym->c)
+        put_extern_sym(sym, NULL, 0, 0);
+
+    /* patch relocation */
+    /* XXX: find a better solution ? */
+
+    rel = (ElfW(Rela) *)(cur_text_section->reloc->data + vtop->c.i);
+    rel->r_info = ELF64_R_INFO(sym->c, ELF64_R_TYPE(rel->r_info));
+}
+#endif
+
+#ifdef TCC_TARGET_PE
+
+#define REGN 4
+static const uint8_t arg_regs[REGN] = {
+    TREG_RCX, TREG_RDX, TREG_R8, TREG_R9
 };
+
+/* Prepare arguments in R10 and R11 rather than RCX and RDX
+   because gv() will not ever use these */
+static int arg_prepare_reg(int idx) {
+  if (idx == 0 || idx == 1)
+      /* idx=0: r10, idx=1: r11 */
+      return idx + 10;
+  else
+      return arg_regs[idx];
+}
+
+static int func_scratch, func_alloca;
+
 /* Generate function call. The function address is pushed first, then
    all the parameters in call order. This functions pops all the
    parameters and the function address. */
+
+static void gen_offs_sp(int b, int r, int d)
+{
+    orex(1,0,r & 0x100 ? 0 : r, b);
+    if (d == (char)d) {
+        o(0x2444 | (REG_VALUE(r) << 3));
+        g(d);
+    } else {
+        o(0x2484 | (REG_VALUE(r) << 3));
+        gen_le32(d);
+    }
+}
+
+static int using_regs(int size)
+{
+    return !(size > 8 || (size & (size - 1)));
+}
+
+/* Return the number of registers needed to return the struct, or 0 if
+   returning via struct pointer. */
+ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
+{
+    int size, align;
+    *ret_align = 1; // Never have to re-align return values for x86-64
+    *regsize = 8;
+    size = type_size(vt, &align);
+    if (!using_regs(size))
+        return 0;
+    if (size == 8)
+        ret->t = VT_LLONG;
+    else if (size == 4)
+        ret->t = VT_INT;
+    else if (size == 2)
+        ret->t = VT_SHORT;
+    else
+        ret->t = VT_BYTE;
+    ret->ref = NULL;
+    return 1;
+}
+
+static int is_sse_float(int t) {
+    int bt;
+    bt = t & VT_BTYPE;
+    return bt == VT_DOUBLE || bt == VT_FLOAT;
+}
+
+static int gfunc_arg_size(CType *type) {
+    int align;
+    if (type->t & (VT_ARRAY|VT_BITFIELD))
+        return 8;
+    return type_size(type, &align);
+}
+
 void gfunc_call(int nb_args)
 {
-    int size, align, r, args_size, i, func_call;
-    Sym *func_sym;
-    SValue *orig_vtop;
-    int nb_reg_args = 0;
-    int nb_sse_args = 0;
-    int sse_reg, gen_reg;
+    int size, r, args_size, i, d, bt, struct_size;
+    int arg;
 
-    /* calculate the number of integer/float arguments */
-    args_size = 0;
-    for(i = 0; i < nb_args; i++) {
-        if ((vtop[-i].type.t & VT_BTYPE) == VT_STRUCT) {
-            args_size += type_size(&vtop->type, &align);
-        } else if ((vtop[-i].type.t & VT_BTYPE) == VT_LDOUBLE) {
-            args_size += 16;
-        } else if (is_sse_float(vtop[-i].type.t)) {
-            nb_sse_args++;
-            if (nb_sse_args > 8) args_size += 8;
-        } else {
-            nb_reg_args++;
-            if (nb_reg_args > 6) args_size += 8;
-        }
-    }
+    args_size = (nb_args < REGN ? REGN : nb_args) * PTR_SIZE;
+    arg = nb_args;
 
     /* for struct arguments, we need to call memcpy and the function
        call breaks register passing arguments we are preparing.
        So, we process arguments which will be passed by stack first. */
-    orig_vtop = vtop;
-    gen_reg = nb_reg_args;
-    sse_reg = nb_sse_args;
-    /* adjust stack to align SSE boundary */
-    if (args_size &= 8) {
-        o(0x50); /* push $rax */
-    }
+    struct_size = args_size;
     for(i = 0; i < nb_args; i++) {
-        if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
-            size = type_size(&vtop->type, &align);
+        SValue *sv;
+        
+        --arg;
+        sv = &vtop[-i];
+        bt = (sv->type.t & VT_BTYPE);
+        size = gfunc_arg_size(&sv->type);
+
+        if (using_regs(size))
+            continue; /* arguments smaller than 8 bytes passed in registers or on stack */
+
+        if (bt == VT_STRUCT) {
             /* align to stack align size */
-            size = (size + 3) & ~3;
-            /* allocate the necessary size on stack */
-            o(0x48);
-            oad(0xec81, size); /* sub $xxx, %rsp */
+            size = (size + 15) & ~15;
             /* generate structure store */
             r = get_reg(RC_INT);
-            o(0x48 + REX_BASE(r));
-            o(0x89); /* mov %rsp, r */
-            o(0xe0 + r);
-            {
-                /* following code breaks vtop[1] */
-                SValue tmp = vtop[1];
-                vset(&vtop->type, r | VT_LVAL, 0);
-                vswap();
-                vstore();
-                vtop[1] = tmp;
-            }
-            args_size += size;
-        } else if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
+            gen_offs_sp(0x8d, r, struct_size);
+            struct_size += size;
+
+            /* generate memcpy call */
+            vset(&sv->type, r | VT_LVAL, 0);
+            vpushv(sv);
+            vstore();
+            --vtop;
+        } else if (bt == VT_LDOUBLE) {
             gv(RC_ST0);
-            size = LDOUBLE_SIZE;
-            oad(0xec8148, size); /* sub $xxx, %rsp */
-            o(0x7cdb); /* fstpt 0(%rsp) */
-            g(0x24);
-            g(0x00);
-            args_size += size;
-        } else if (is_sse_float(vtop->type.t)) {
-            int j = --sse_reg;
-            if (j >= 8) {
-                gv(RC_FLOAT);
-                o(0x50); /* push $rax */
-                /* movq %xmm0, (%rsp) */
-                o(0x04d60f66);
-                o(0x24);
-                args_size += 8;
+            gen_offs_sp(0xdb, 0x107, struct_size);
+            struct_size += 16;
+        }
+    }
+
+    if (func_scratch < struct_size)
+        func_scratch = struct_size;
+
+    arg = nb_args;
+    struct_size = args_size;
+
+    for(i = 0; i < nb_args; i++) {
+        --arg;
+        bt = (vtop->type.t & VT_BTYPE);
+
+        size = gfunc_arg_size(&vtop->type);
+        if (!using_regs(size)) {
+            /* align to stack align size */
+            size = (size + 15) & ~15;
+            if (arg >= REGN) {
+                d = get_reg(RC_INT);
+                gen_offs_sp(0x8d, d, struct_size);
+                gen_offs_sp(0x89, d, arg*8);
+            } else {
+                d = arg_prepare_reg(arg);
+                gen_offs_sp(0x8d, d, struct_size);
             }
+            struct_size += size;
         } else {
-            int j = --gen_reg;
-            /* simple type */
-            /* XXX: implicit cast ? */
-            if (j >= 6) {
+            if (is_sse_float(vtop->type.t)) {
+		if (tcc_state->nosse)
+		  tcc_error("SSE disabled");
+                gv(RC_XMM0); /* only use one float register */
+                if (arg >= REGN) {
+                    /* movq %xmm0, j*8(%rsp) */
+                    gen_offs_sp(0xd60f66, 0x100, arg*8);
+                } else {
+                    /* movaps %xmm0, %xmmN */
+                    o(0x280f);
+                    o(0xc0 + (arg << 3));
+                    d = arg_prepare_reg(arg);
+                    /* mov %xmm0, %rxx */
+                    o(0x66);
+                    orex(1,d,0, 0x7e0f);
+                    o(0xc0 + REG_VALUE(d));
+                }
+            } else {
+                if (bt == VT_STRUCT) {
+                    vtop->type.ref = NULL;
+                    vtop->type.t = size > 4 ? VT_LLONG : size > 2 ? VT_INT
+                        : size > 1 ? VT_SHORT : VT_BYTE;
+                }
+                
                 r = gv(RC_INT);
-                o(0x50 + r); /* push r */
-                args_size += 8;
+                if (arg >= REGN) {
+                    gen_offs_sp(0x89, r, arg*8);
+                } else {
+                    d = arg_prepare_reg(arg);
+                    orex(1,d,r,0x89); /* mov */
+                    o(0xc0 + REG_VALUE(r) * 8 + REG_VALUE(d));
+                }
             }
         }
         vtop--;
     }
-    vtop = orig_vtop;
+    save_regs(0);
+    
+    /* Copy R10 and R11 into RCX and RDX, respectively */
+    if (nb_args > 0) {
+        o(0xd1894c); /* mov %r10, %rcx */
+        if (nb_args > 1) {
+            o(0xda894c); /* mov %r11, %rdx */
+        }
+    }
+    
+    gcall_or_jmp(0);
+
+    if ((vtop->r & VT_SYM) && vtop->sym->v == TOK_alloca) {
+        /* need to add the "func_scratch" area after alloca */
+        o(0x0548), gen_le32(func_alloca), func_alloca = ind - 4;
+    }
+
+    /* other compilers don't clear the upper bits when returning char/short */
+    bt = vtop->type.ref->type.t & (VT_BTYPE | VT_UNSIGNED);
+    if (bt == (VT_BYTE | VT_UNSIGNED))
+        o(0xc0b60f);  /* movzbl %al, %eax */
+    else if (bt == VT_BYTE)
+        o(0xc0be0f); /* movsbl %al, %eax */
+    else if (bt == VT_SHORT)
+        o(0x98); /* cwtl */
+    else if (bt == (VT_SHORT | VT_UNSIGNED))
+        o(0xc0b70f);  /* movzbl %al, %eax */
+#if 0 /* handled in gen_cast() */
+    else if (bt == VT_INT)
+        o(0x9848); /* cltq */
+    else if (bt == (VT_INT | VT_UNSIGNED))
+        o(0xc089); /* mov %eax,%eax */
+#endif
+    vtop--;
+}
+
+
+#define FUNC_PROLOG_SIZE 11
+
+/* generate function prolog of type 't' */
+void gfunc_prolog(CType *func_type)
+{
+    int addr, reg_param_index, bt, size;
+    Sym *sym;
+    CType *type;
+
+    func_ret_sub = 0;
+    func_scratch = 0;
+    func_alloca = 0;
+    loc = 0;
+
+    addr = PTR_SIZE * 2;
+    ind += FUNC_PROLOG_SIZE;
+    func_sub_sp_offset = ind;
+    reg_param_index = 0;
+
+    sym = func_type->ref;
+
+    /* if the function returns a structure, then add an
+       implicit pointer parameter */
+    func_vt = sym->type;
+    func_var = (sym->f.func_type == FUNC_ELLIPSIS);
+    size = gfunc_arg_size(&func_vt);
+    if (!using_regs(size)) {
+        gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
+        func_vc = addr;
+        reg_param_index++;
+        addr += 8;
+    }
+
+    /* define parameters */
+    while ((sym = sym->next) != NULL) {
+        type = &sym->type;
+        bt = type->t & VT_BTYPE;
+        size = gfunc_arg_size(type);
+        if (!using_regs(size)) {
+            if (reg_param_index < REGN) {
+                gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
+            }
+            sym_push(sym->v & ~SYM_FIELD, type, VT_LLOCAL | VT_LVAL, addr);
+        } else {
+            if (reg_param_index < REGN) {
+                /* save arguments passed by register */
+                if ((bt == VT_FLOAT) || (bt == VT_DOUBLE)) {
+		    if (tcc_state->nosse)
+		      tcc_error("SSE disabled");
+                    o(0xd60f66); /* movq */
+                    gen_modrm(reg_param_index, VT_LOCAL, NULL, addr);
+                } else {
+                    gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
+                }
+            }
+            sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | VT_LVAL, addr);
+        }
+        addr += 8;
+        reg_param_index++;
+    }
+
+    while (reg_param_index < REGN) {
+        if (func_type->ref->f.func_type == FUNC_ELLIPSIS) {
+            gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, addr);
+            addr += 8;
+        }
+        reg_param_index++;
+    }
+}
+
+/* generate function epilog */
+void gfunc_epilog(void)
+{
+    int v, saved_ind;
+
+    o(0xc9); /* leave */
+    if (func_ret_sub == 0) {
+        o(0xc3); /* ret */
+    } else {
+        o(0xc2); /* ret n */
+        g(func_ret_sub);
+        g(func_ret_sub >> 8);
+    }
+
+    saved_ind = ind;
+    ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
+    /* align local size to word & save local variables */
+    v = (func_scratch + -loc + 15) & -16;
+
+    if (v >= 4096) {
+        Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
+        oad(0xb8, v); /* mov stacksize, %eax */
+        oad(0xe8, 0); /* call __chkstk, (does the stackframe too) */
+        greloca(cur_text_section, sym, ind-4, R_X86_64_PC32, -4);
+        o(0x90); /* fill for FUNC_PROLOG_SIZE = 11 bytes */
+    } else {
+        o(0xe5894855);  /* push %rbp, mov %rsp, %rbp */
+        o(0xec8148);  /* sub rsp, stacksize */
+        gen_le32(v);
+    }
+
+    /* add the "func_scratch" area after each alloca seen */
+    while (func_alloca) {
+        unsigned char *ptr = cur_text_section->data + func_alloca;
+        func_alloca = read32le(ptr);
+        write32le(ptr, func_scratch);
+    }
+
+    cur_text_section->data_offset = saved_ind;
+    pe_add_unwind_data(ind, saved_ind, v);
+    ind = cur_text_section->data_offset;
+}
+
+#else
+
+static void gadd_sp(int val)
+{
+    if (val == (char)val) {
+        o(0xc48348);
+        g(val);
+    } else {
+        oad(0xc48148, val); /* add $xxx, %rsp */
+    }
+}
+
+typedef enum X86_64_Mode {
+  x86_64_mode_none,
+  x86_64_mode_memory,
+  x86_64_mode_integer,
+  x86_64_mode_sse,
+  x86_64_mode_x87
+} X86_64_Mode;
+
+static X86_64_Mode classify_x86_64_merge(X86_64_Mode a, X86_64_Mode b)
+{
+    if (a == b)
+        return a;
+    else if (a == x86_64_mode_none)
+        return b;
+    else if (b == x86_64_mode_none)
+        return a;
+    else if ((a == x86_64_mode_memory) || (b == x86_64_mode_memory))
+        return x86_64_mode_memory;
+    else if ((a == x86_64_mode_integer) || (b == x86_64_mode_integer))
+        return x86_64_mode_integer;
+    else if ((a == x86_64_mode_x87) || (b == x86_64_mode_x87))
+        return x86_64_mode_memory;
+    else
+        return x86_64_mode_sse;
+}
+
+static X86_64_Mode classify_x86_64_inner(CType *ty)
+{
+    X86_64_Mode mode;
+    Sym *f;
+    
+    switch (ty->t & VT_BTYPE) {
+    case VT_VOID: return x86_64_mode_none;
+    
+    case VT_INT:
+    case VT_BYTE:
+    case VT_SHORT:
+    case VT_LLONG:
+    case VT_BOOL:
+    case VT_PTR:
+    case VT_FUNC:
+        return x86_64_mode_integer;
+    
+    case VT_FLOAT:
+    case VT_DOUBLE: return x86_64_mode_sse;
+    
+    case VT_LDOUBLE: return x86_64_mode_x87;
+      
+    case VT_STRUCT:
+        f = ty->ref;
+
+        mode = x86_64_mode_none;
+        for (f = f->next; f; f = f->next)
+            mode = classify_x86_64_merge(mode, classify_x86_64_inner(&f->type));
+        
+        return mode;
+    }
+    assert(0);
+    return 0;
+}
+
+static X86_64_Mode classify_x86_64_arg(CType *ty, CType *ret, int *psize, int *palign, int *reg_count)
+{
+    X86_64_Mode mode;
+    int size, align, ret_t = 0;
+    
+    if (ty->t & (VT_BITFIELD|VT_ARRAY)) {
+        *psize = 8;
+        *palign = 8;
+        *reg_count = 1;
+        ret_t = ty->t;
+        mode = x86_64_mode_integer;
+    } else {
+        size = type_size(ty, &align);
+        *psize = (size + 7) & ~7;
+        *palign = (align + 7) & ~7;
+    
+        if (size > 16) {
+            mode = x86_64_mode_memory;
+        } else {
+            mode = classify_x86_64_inner(ty);
+            switch (mode) {
+            case x86_64_mode_integer:
+                if (size > 8) {
+                    *reg_count = 2;
+                    ret_t = VT_QLONG;
+                } else {
+                    *reg_count = 1;
+                    ret_t = (size > 4) ? VT_LLONG : VT_INT;
+                }
+                break;
+                
+            case x86_64_mode_x87:
+                *reg_count = 1;
+                ret_t = VT_LDOUBLE;
+                break;
+
+            case x86_64_mode_sse:
+                if (size > 8) {
+                    *reg_count = 2;
+                    ret_t = VT_QFLOAT;
+                } else {
+                    *reg_count = 1;
+                    ret_t = (size > 4) ? VT_DOUBLE : VT_FLOAT;
+                }
+                break;
+            default: break; /* nothing to be done for x86_64_mode_memory and x86_64_mode_none*/
+            }
+        }
+    }
+    
+    if (ret) {
+        ret->ref = NULL;
+        ret->t = ret_t;
+    }
+    
+    return mode;
+}
+
+ST_FUNC int classify_x86_64_va_arg(CType *ty)
+{
+    /* This definition must be synced with stdarg.h */
+    enum __va_arg_type {
+        __va_gen_reg, __va_float_reg, __va_stack
+    };
+    int size, align, reg_count;
+    X86_64_Mode mode = classify_x86_64_arg(ty, NULL, &size, &align, &reg_count);
+    switch (mode) {
+    default: return __va_stack;
+    case x86_64_mode_integer: return __va_gen_reg;
+    case x86_64_mode_sse: return __va_float_reg;
+    }
+}
+
+/* Return the number of registers needed to return the struct, or 0 if
+   returning via struct pointer. */
+ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
+{
+    int size, align, reg_count;
+    *ret_align = 1; // Never have to re-align return values for x86-64
+    *regsize = 8;
+    return (classify_x86_64_arg(vt, ret, &size, &align, &reg_count) != x86_64_mode_memory);
+}
+
+#define REGN 6
+static const uint8_t arg_regs[REGN] = {
+    TREG_RDI, TREG_RSI, TREG_RDX, TREG_RCX, TREG_R8, TREG_R9
+};
+
+static int arg_prepare_reg(int idx) {
+  if (idx == 2 || idx == 3)
+      /* idx=2: r10, idx=3: r11 */
+      return idx + 8;
+  else
+      return arg_regs[idx];
+}
+
+/* Generate function call. The function address is pushed first, then
+   all the parameters in call order. This functions pops all the
+   parameters and the function address. */
+void gfunc_call(int nb_args)
+{
+    X86_64_Mode mode;
+    CType type;
+    int size, align, r, args_size, stack_adjust, i, reg_count;
+    int nb_reg_args = 0;
+    int nb_sse_args = 0;
+    int sse_reg, gen_reg;
+    char _onstack[nb_args], *onstack = _onstack;
+
+    /* calculate the number of integer/float register arguments, remember
+       arguments to be passed via stack (in onstack[]), and also remember
+       if we have to align the stack pointer to 16 (onstack[i] == 2).  Needs
+       to be done in a left-to-right pass over arguments.  */
+    stack_adjust = 0;
+    for(i = nb_args - 1; i >= 0; i--) {
+        mode = classify_x86_64_arg(&vtop[-i].type, NULL, &size, &align, &reg_count);
+        if (mode == x86_64_mode_sse && nb_sse_args + reg_count <= 8) {
+            nb_sse_args += reg_count;
+	    onstack[i] = 0;
+	} else if (mode == x86_64_mode_integer && nb_reg_args + reg_count <= REGN) {
+            nb_reg_args += reg_count;
+	    onstack[i] = 0;
+	} else if (mode == x86_64_mode_none) {
+	    onstack[i] = 0;
+	} else {
+	    if (align == 16 && (stack_adjust &= 15)) {
+		onstack[i] = 2;
+		stack_adjust = 0;
+	    } else
+	      onstack[i] = 1;
+	    stack_adjust += size;
+	}
+    }
+
+    if (nb_sse_args && tcc_state->nosse)
+      tcc_error("SSE disabled but floating point arguments passed");
+
+    /* fetch cpu flag before generating any code */
+    if (vtop >= vstack && (vtop->r & VT_VALMASK) == VT_CMP)
+      gv(RC_INT);
+
+    /* for struct arguments, we need to call memcpy and the function
+       call breaks register passing arguments we are preparing.
+       So, we process arguments which will be passed by stack first. */
+    gen_reg = nb_reg_args;
+    sse_reg = nb_sse_args;
+    args_size = 0;
+    stack_adjust &= 15;
+    for (i = 0; i < nb_args;) {
+	mode = classify_x86_64_arg(&vtop[-i].type, NULL, &size, &align, &reg_count);
+	if (!onstack[i]) {
+	    ++i;
+	    continue;
+	}
+        /* Possibly adjust stack to align SSE boundary.  We're processing
+	   args from right to left while allocating happens left to right
+	   (stack grows down), so the adjustment needs to happen _after_
+	   an argument that requires it.  */
+        if (stack_adjust) {
+	    o(0x50); /* push %rax; aka sub $8,%rsp */
+            args_size += 8;
+	    stack_adjust = 0;
+        }
+	if (onstack[i] == 2)
+	  stack_adjust = 1;
+
+	vrotb(i+1);
+
+	switch (vtop->type.t & VT_BTYPE) {
+	    case VT_STRUCT:
+		/* allocate the necessary size on stack */
+		o(0x48);
+		oad(0xec81, size); /* sub $xxx, %rsp */
+		/* generate structure store */
+		r = get_reg(RC_INT);
+		orex(1, r, 0, 0x89); /* mov %rsp, r */
+		o(0xe0 + REG_VALUE(r));
+		vset(&vtop->type, r | VT_LVAL, 0);
+		vswap();
+		vstore();
+		break;
+
+	    case VT_LDOUBLE:
+                gv(RC_ST0);
+                oad(0xec8148, size); /* sub $xxx, %rsp */
+                o(0x7cdb); /* fstpt 0(%rsp) */
+                g(0x24);
+                g(0x00);
+		break;
+
+	    case VT_FLOAT:
+	    case VT_DOUBLE:
+		assert(mode == x86_64_mode_sse);
+		r = gv(RC_FLOAT);
+		o(0x50); /* push $rax */
+		/* movq %xmmN, (%rsp) */
+		o(0xd60f66);
+		o(0x04 + REG_VALUE(r)*8);
+		o(0x24);
+		break;
+
+	    default:
+		assert(mode == x86_64_mode_integer);
+		/* simple type */
+		/* XXX: implicit cast ? */
+		r = gv(RC_INT);
+		orex(0,r,0,0x50 + REG_VALUE(r)); /* push r */
+		break;
+	}
+	args_size += size;
+
+	vpop();
+	--nb_args;
+	onstack++;
+    }
+
+    /* XXX This should be superfluous.  */
+    save_regs(0); /* save used temporary registers */
 
     /* then, we prepare register passing arguments.
        Note that we cannot set RDX and RCX in this loop because gv()
        may break these temporary registers. Let's use R10 and R11
        instead of them */
-    gen_reg = nb_reg_args;
-    sse_reg = nb_sse_args;
+    assert(gen_reg <= REGN);
+    assert(sse_reg <= 8);
     for(i = 0; i < nb_args; i++) {
-        if ((vtop->type.t & VT_BTYPE) == VT_STRUCT ||
-            (vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
-        } else if (is_sse_float(vtop->type.t)) {
-            int j = --sse_reg;
-            if (j < 8) {
-                gv(RC_FLOAT); /* only one float register */
-                /* movaps %xmm0, %xmmN */
-                o(0x280f);
-                o(0xc0 + (sse_reg << 3));
+        mode = classify_x86_64_arg(&vtop->type, &type, &size, &align, &reg_count);
+        /* Alter stack entry type so that gv() knows how to treat it */
+        vtop->type = type;
+        if (mode == x86_64_mode_sse) {
+            if (reg_count == 2) {
+                sse_reg -= 2;
+                gv(RC_FRET); /* Use pair load into xmm0 & xmm1 */
+                if (sse_reg) { /* avoid redundant movaps %xmm0, %xmm0 */
+                    /* movaps %xmm0, %xmmN */
+                    o(0x280f);
+                    o(0xc0 + (sse_reg << 3));
+                    /* movaps %xmm1, %xmmN */
+                    o(0x280f);
+                    o(0xc1 + ((sse_reg+1) << 3));
+                }
+            } else {
+                assert(reg_count == 1);
+                --sse_reg;
+                /* Load directly to register */
+                gv(RC_XMM0 << sse_reg);
             }
-        } else {
-            int j = --gen_reg;
+        } else if (mode == x86_64_mode_integer) {
             /* simple type */
             /* XXX: implicit cast ? */
-            if (j < 6) {
-                r = gv(RC_INT);
-                if (j < 2) {
-                    o(0x8948); /* mov */
-                    o(0xc0 + r * 8 + arg_regs[j]);
-                } else if (j < 4) {
-                    o(0x8949); /* mov */
-                    /* j=2: r10, j=3: r11 */
-                    o(0xc0 + r * 8 + j);
-                } else {
-                    o(0x8949); /* mov */
-                    /* j=4: r8, j=5: r9 */
-                    o(0xc0 + r * 8 + j - 4);
-                }
+            int d;
+            gen_reg -= reg_count;
+            r = gv(RC_INT);
+            d = arg_prepare_reg(gen_reg);
+            orex(1,d,r,0x89); /* mov */
+            o(0xc0 + REG_VALUE(r) * 8 + REG_VALUE(d));
+            if (reg_count == 2) {
+                d = arg_prepare_reg(gen_reg+1);
+                orex(1,d,vtop->r2,0x89); /* mov */
+                o(0xc0 + REG_VALUE(vtop->r2) * 8 + REG_VALUE(d));
             }
         }
         vtop--;
     }
+    assert(gen_reg == 0);
+    assert(sse_reg == 0);
 
-    save_regs(0); /* save used temporary registers */
+    /* We shouldn't have many operands on the stack anymore, but the
+       call address itself is still there, and it might be in %eax
+       (or edx/ecx) currently, which the below writes would clobber.
+       So evict all remaining operands here.  */
+    save_regs(0);
 
     /* Copy R10 and R11 into RDX and RCX, respectively */
     if (nb_reg_args > 2) {
@@ -660,22 +1394,16 @@ void gfunc_call(int nb_args)
         }
     }
 
-    func_sym = vtop->type.ref;
-    func_call = FUNC_CALL(func_sym->r);
-    oad(0xb8, nb_sse_args < 8 ? nb_sse_args : 8); /* mov nb_sse_args, %eax */
+    if (vtop->type.ref->f.func_type != FUNC_NEW) /* implies FUNC_OLD or FUNC_ELLIPSIS */
+        oad(0xb8, nb_sse_args < 8 ? nb_sse_args : 8); /* mov nb_sse_args, %eax */
     gcall_or_jmp(0);
     if (args_size)
         gadd_sp(args_size);
     vtop--;
 }
 
-#ifdef TCC_TARGET_PE
-/* XXX: support PE? */
-#warning "PE isn't tested at all"
-#define FUNC_PROLOG_SIZE 12
-#else
+
 #define FUNC_PROLOG_SIZE 11
-#endif
 
 static void push_arg_reg(int i) {
     loc -= 8;
@@ -685,21 +1413,20 @@ static void push_arg_reg(int i) {
 /* generate function prolog of type 't' */
 void gfunc_prolog(CType *func_type)
 {
-    int i, addr, align, size, func_call;
-    int param_index, param_addr, reg_param_index, sse_param_index;
+    X86_64_Mode mode;
+    int i, addr, align, size, reg_count;
+    int param_addr = 0, reg_param_index, sse_param_index;
     Sym *sym;
     CType *type;
 
-    func_ret_sub = 0;
-
     sym = func_type->ref;
-    func_call = FUNC_CALL(sym->r);
     addr = PTR_SIZE * 2;
     loc = 0;
     ind += FUNC_PROLOG_SIZE;
     func_sub_sp_offset = ind;
+    func_ret_sub = 0;
 
-    if (func_type->ref->c == FUNC_ELLIPSIS) {
+    if (sym->f.func_type == FUNC_ELLIPSIS) {
         int seen_reg_num, seen_sse_num, seen_stack_size;
         seen_reg_num = seen_sse_num = 0;
         /* frame pointer and return address */
@@ -708,24 +1435,24 @@ void gfunc_prolog(CType *func_type)
         sym = func_type->ref;
         while ((sym = sym->next) != NULL) {
             type = &sym->type;
-            if (is_sse_float(type->t)) {
-                if (seen_sse_num < 8) {
-                    seen_sse_num++;
-                } else {
-                    seen_stack_size += 8;
-                }
-            } else if ((type->t & VT_BTYPE) == VT_STRUCT) {
-                size = type_size(type, &align);
-                size = (size + 3) & ~3;
-                seen_stack_size += size;
-            } else if ((type->t & VT_BTYPE) == VT_LDOUBLE) {
-                seen_stack_size += LDOUBLE_SIZE;
-            } else {
-                if (seen_reg_num < 6) {
-                    seen_reg_num++;
-                } else {
-                    seen_stack_size += 8;
-                }
+            mode = classify_x86_64_arg(type, NULL, &size, &align, &reg_count);
+            switch (mode) {
+            default:
+            stack_arg:
+                seen_stack_size = ((seen_stack_size + align - 1) & -align) + size;
+                break;
+                
+            case x86_64_mode_integer:
+                if (seen_reg_num + reg_count > REGN)
+		    goto stack_arg;
+		seen_reg_num += reg_count;
+                break;
+                
+            case x86_64_mode_sse:
+                if (seen_sse_num + reg_count > 8)
+		    goto stack_arg;
+		seen_sse_num += reg_count;
+                break;
             }
         }
 
@@ -743,70 +1470,96 @@ void gfunc_prolog(CType *func_type)
         /* save all register passing arguments */
         for (i = 0; i < 8; i++) {
             loc -= 16;
-            o(0xd60f66); /* movq */
-            gen_modrm(7 - i, VT_LOCAL, NULL, loc);
+	    if (!tcc_state->nosse) {
+		o(0xd60f66); /* movq */
+		gen_modrm(7 - i, VT_LOCAL, NULL, loc);
+	    }
             /* movq $0, loc+8(%rbp) */
             o(0x85c748);
             gen_le32(loc + 8);
             gen_le32(0);
         }
-        for (i = 0; i < 6; i++) {
-            push_arg_reg(5 - i);
+        for (i = 0; i < REGN; i++) {
+            push_arg_reg(REGN-1-i);
         }
     }
 
     sym = func_type->ref;
-    param_index = 0;
     reg_param_index = 0;
     sse_param_index = 0;
 
     /* if the function returns a structure, then add an
        implicit pointer parameter */
     func_vt = sym->type;
-    if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
+    mode = classify_x86_64_arg(&func_vt, NULL, &size, &align, &reg_count);
+    if (mode == x86_64_mode_memory) {
         push_arg_reg(reg_param_index);
-        param_addr = loc;
-
         func_vc = loc;
-        param_index++;
         reg_param_index++;
     }
     /* define parameters */
     while ((sym = sym->next) != NULL) {
         type = &sym->type;
-        size = type_size(type, &align);
-        size = (size + 3) & ~3;
-        if (is_sse_float(type->t)) {
-            if (sse_param_index < 8) {
+        mode = classify_x86_64_arg(type, NULL, &size, &align, &reg_count);
+        switch (mode) {
+        case x86_64_mode_sse:
+	    if (tcc_state->nosse)
+	        tcc_error("SSE disabled but floating point arguments used");
+            if (sse_param_index + reg_count <= 8) {
                 /* save arguments passed by register */
-                loc -= 8;
-                o(0xd60f66); /* movq */
-                gen_modrm(sse_param_index, VT_LOCAL, NULL, loc);
+                loc -= reg_count * 8;
                 param_addr = loc;
+                for (i = 0; i < reg_count; ++i) {
+                    o(0xd60f66); /* movq */
+                    gen_modrm(sse_param_index, VT_LOCAL, NULL, param_addr + i*8);
+                    ++sse_param_index;
+                }
             } else {
+                addr = (addr + align - 1) & -align;
                 param_addr = addr;
                 addr += size;
             }
-            sse_param_index++;
-        } else if ((type->t & VT_BTYPE) == VT_STRUCT ||
-                   (type->t & VT_BTYPE) == VT_LDOUBLE) {
+            break;
+            
+        case x86_64_mode_memory:
+        case x86_64_mode_x87:
+            addr = (addr + align - 1) & -align;
             param_addr = addr;
             addr += size;
-        } else {
-            if (reg_param_index < 6) {
+            break;
+            
+        case x86_64_mode_integer: {
+            if (reg_param_index + reg_count <= REGN) {
                 /* save arguments passed by register */
-                push_arg_reg(reg_param_index);
+                loc -= reg_count * 8;
                 param_addr = loc;
+                for (i = 0; i < reg_count; ++i) {
+                    gen_modrm64(0x89, arg_regs[reg_param_index], VT_LOCAL, NULL, param_addr + i*8);
+                    ++reg_param_index;
+                }
             } else {
+                addr = (addr + align - 1) & -align;
                 param_addr = addr;
-                addr += 8;
+                addr += size;
             }
-            reg_param_index++;
+            break;
+        }
+	default: break; /* nothing to be done for x86_64_mode_none */
         }
         sym_push(sym->v & ~SYM_FIELD, type,
                  VT_LOCAL | VT_LVAL, param_addr);
-        param_index++;
     }
+
+#ifdef CONFIG_TCC_BCHECK
+    /* leave some room for bound checking code */
+    if (tcc_state->do_bounds_check) {
+        func_bound_offset = lbounds_section->data_offset;
+        func_bound_ind = ind;
+        oad(0xb8, 0); /* lbound section pointer */
+	o(0xc78948);  /* mov  %rax,%rdi ## first arg in %rdi, this must be ptr */
+	oad(0xb8, 0); /* call to function */
+    }
+#endif
 }
 
 /* generate function epilog */
@@ -814,6 +1567,37 @@ void gfunc_epilog(void)
 {
     int v, saved_ind;
 
+#ifdef CONFIG_TCC_BCHECK
+    if (tcc_state->do_bounds_check
+	&& func_bound_offset != lbounds_section->data_offset)
+    {
+        addr_t saved_ind;
+        addr_t *bounds_ptr;
+        Sym *sym_data;
+
+        /* add end of table info */
+        bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
+        *bounds_ptr = 0;
+
+        /* generate bound local allocation */
+        sym_data = get_sym_ref(&char_pointer_type, lbounds_section, 
+                               func_bound_offset, lbounds_section->data_offset);
+        saved_ind = ind;
+        ind = func_bound_ind;
+        greloca(cur_text_section, sym_data, ind + 1, R_X86_64_64, 0);
+        ind = ind + 5 + 3;
+        gen_static_call(TOK___bound_local_new);
+        ind = saved_ind;
+
+        /* generate bound check local freeing */
+        o(0x5250); /* save returned value, if any */
+        greloca(cur_text_section, sym_data, ind + 1, R_X86_64_64, 0);
+        oad(0xb8, 0); /* mov xxx, %rax */
+        o(0xc78948);  /* mov %rax,%rdi # first arg in %rdi, this must be ptr */
+        gen_static_call(TOK___bound_local_delete);
+        o(0x585a); /* restore returned value, if any */
+    }
+#endif
     o(0xc9); /* leave */
     if (func_ret_sub == 0) {
         o(0xc3); /* ret */
@@ -826,29 +1610,18 @@ void gfunc_epilog(void)
     v = (-loc + 15) & -16;
     saved_ind = ind;
     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
-#ifdef TCC_TARGET_PE
-    if (v >= 4096) {
-        Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
-        oad(0xb8, v); /* mov stacksize, %eax */
-        oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
-        greloc(cur_text_section, sym, ind-4, R_X86_64_PC32);
-    } else
-#endif
-    {
-        o(0xe5894855);  /* push %rbp, mov %rsp, %rbp */
-        o(0xec8148);  /* sub rsp, stacksize */
-        gen_le32(v);
-#if FUNC_PROLOG_SIZE == 12
-        o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
-#endif
-    }
+    o(0xe5894855);  /* push %rbp, mov %rsp, %rbp */
+    o(0xec8148);  /* sub rsp, stacksize */
+    gen_le32(v);
     ind = saved_ind;
 }
 
+#endif /* not PE */
+
 /* generate a jump to a label */
 int gjmp(int t)
 {
-    return psym(0xe9, t);
+    return gjmp2(0xe9, t);
 }
 
 /* generate a jump to a fixed address */
@@ -864,46 +1637,76 @@ void gjmp_addr(int a)
     }
 }
 
+ST_FUNC void gtst_addr(int inv, int a)
+{
+    int v = vtop->r & VT_VALMASK;
+    if (v == VT_CMP) {
+	inv ^= (vtop--)->c.i;
+	a -= ind + 2;
+	if (a == (char)a) {
+	    g(inv - 32);
+	    g(a);
+	} else {
+	    g(0x0f);
+	    oad(inv - 16, a - 4);
+	}
+    } else if ((v & ~1) == VT_JMP) {
+	if ((v & 1) != inv) {
+	    gjmp_addr(a);
+	    gsym(vtop->c.i);
+	} else {
+	    gsym(vtop->c.i);
+	    o(0x05eb);
+	    gjmp_addr(a);
+	}
+	vtop--;
+    }
+}
+
 /* generate a test. set 'inv' to invert test. Stack entry is popped */
-int gtst(int inv, int t)
+ST_FUNC int gtst(int inv, int t)
 {
-    int v, *p;
+    int v = vtop->r & VT_VALMASK;
 
-    v = vtop->r & VT_VALMASK;
-    if (v == VT_CMP) {
+    if (nocode_wanted) {
+        ;
+    } else if (v == VT_CMP) {
         /* fast case : can jump directly since flags are set */
+	if (vtop->c.i & 0x100)
+	  {
+	    /* This was a float compare.  If the parity flag is set
+	       the result was unordered.  For anything except != this
+	       means false and we don't jump (anding both conditions).
+	       For != this means true (oring both).
+	       Take care about inverting the test.  We need to jump
+	       to our target if the result was unordered and test wasn't NE,
+	       otherwise if unordered we don't want to jump.  */
+	    vtop->c.i &= ~0x100;
+            if (inv == (vtop->c.i == TOK_NE))
+	      o(0x067a);  /* jp +6 */
+	    else
+	      {
+	        g(0x0f);
+		t = gjmp2(0x8a, t); /* jp t */
+	      }
+	  }
         g(0x0f);
-        t = psym((vtop->c.i - 16) ^ inv, t);
+        t = gjmp2((vtop->c.i - 16) ^ inv, t);
     } else if (v == VT_JMP || v == VT_JMPI) {
         /* && or || optimization */
         if ((v & 1) == inv) {
             /* insert vtop->c jump list in t */
-            p = &vtop->c.i;
-            while (*p != 0)
-                p = (int *)(cur_text_section->data + *p);
-            *p = t;
-            t = vtop->c.i;
+            uint32_t n1, n = vtop->c.i;
+            if (n) {
+                while ((n1 = read32le(cur_text_section->data + n)))
+                    n = n1;
+                write32le(cur_text_section->data + n, t);
+                t = vtop->c.i;
+            }
         } else {
             t = gjmp(t);
             gsym(vtop->c.i);
         }
-    } else {
-        if (is_float(vtop->type.t) ||
-            (vtop->type.t & VT_BTYPE) == VT_LLONG) {
-            vpushi(0);
-            gen_op(TOK_NE);
-        }
-        if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
-            /* constant jmp optimization */
-            if ((vtop->c.i != 0) != inv)
-                t = gjmp(t);
-        } else {
-            v = gv(RC_INT);
-            o(0x85);
-            o(0xc0 + v * 9);
-            g(0x0f);
-            t = psym(0x85 ^ inv, t);
-        }
     }
     vtop--;
     return t;
@@ -913,41 +1716,37 @@ int gtst(int inv, int t)
 void gen_opi(int op)
 {
     int r, fr, opc, c;
+    int ll, uu, cc;
+
+    ll = is64_type(vtop[-1].type.t);
+    uu = (vtop[-1].type.t & VT_UNSIGNED) != 0;
+    cc = (vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST;
 
     switch(op) {
     case '+':
     case TOK_ADDC1: /* add with carry generation */
         opc = 0;
     gen_op8:
-        if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST &&
-            !is64_type(vtop->type.t)) {
+        if (cc && (!ll || (int)vtop->c.i == vtop->c.i)) {
             /* constant case */
             vswap();
             r = gv(RC_INT);
-            if (is64_type(vtop->type.t)) {
-                o(0x48 | REX_BASE(r));
-            }
             vswap();
             c = vtop->c.i;
             if (c == (char)c) {
                 /* XXX: generate inc and dec for smaller code ? */
-                o(0x83);
+                orex(ll, r, 0, 0x83);
                 o(0xc0 | (opc << 3) | REG_VALUE(r));
                 g(c);
             } else {
-                o(0x81);
+                orex(ll, r, 0, 0x81);
                 oad(0xc0 | (opc << 3) | REG_VALUE(r), c);
             }
         } else {
             gv2(RC_INT, RC_INT);
             r = vtop[-1].r;
             fr = vtop[0].r;
-            if (opc != 7 ||
-                is64_type(vtop[0].type.t) || (vtop[0].type.t & VT_UNSIGNED) ||
-                is64_type(vtop[-1].type.t) || (vtop[-1].type.t & VT_UNSIGNED)) {
-                o(0x48 | REX_BASE(r) | (REX_BASE(fr) << 2));
-            }
-            o((opc << 3) | 0x01);
+            orex(ll, r, fr, (opc << 3) | 0x01);
             o(0xc0 + REG_VALUE(r) + REG_VALUE(fr) * 8);
         }
         vtop--;
@@ -979,13 +1778,9 @@ void gen_opi(int op)
         gv2(RC_INT, RC_INT);
         r = vtop[-1].r;
         fr = vtop[0].r;
-        if (is64_type(vtop[0].type.t) || (vtop[0].type.t & VT_UNSIGNED) ||
-            is64_type(vtop[-1].type.t) || (vtop[-1].type.t & VT_UNSIGNED)) {
-            o(0x48 | REX_BASE(fr) | (REX_BASE(r) << 2));
-        }
+        orex(ll, fr, r, 0xaf0f); /* imul fr, r */
+        o(0xc0 + REG_VALUE(fr) + REG_VALUE(r) * 8);
         vtop--;
-        o(0xaf0f); /* imul fr, r */
-        o(0xc0 + fr + r * 8);
         break;
     case TOK_SHL:
         opc = 4;
@@ -997,39 +1792,32 @@ void gen_opi(int op)
         opc = 7;
     gen_shift:
         opc = 0xc0 | (opc << 3);
-        if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
+        if (cc) {
             /* constant case */
             vswap();
             r = gv(RC_INT);
-            if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
-                o(0x48 | REX_BASE(r));
-                c = 0x3f;
-            } else {
-                c = 0x1f;
-            }
             vswap();
-            c &= vtop->c.i;
-            o(0xc1); /* shl/shr/sar $xxx, r */
-            o(opc | r);
-            g(c);
+            orex(ll, r, 0, 0xc1); /* shl/shr/sar $xxx, r */
+            o(opc | REG_VALUE(r));
+            g(vtop->c.i & (ll ? 63 : 31));
         } else {
             /* we generate the shift in ecx */
             gv2(RC_INT, RC_RCX);
             r = vtop[-1].r;
-            if ((vtop[-1].type.t & VT_BTYPE) == VT_LLONG) {
-                o(0x48 | REX_BASE(r));
-            }
-            o(0xd3); /* shl/shr/sar %cl, r */
-            o(opc | r);
+            orex(ll, r, 0, 0xd3); /* shl/shr/sar %cl, r */
+            o(opc | REG_VALUE(r));
         }
         vtop--;
         break;
-    case '/':
     case TOK_UDIV:
-    case TOK_PDIV:
-    case '%':
     case TOK_UMOD:
-    case TOK_UMULL:
+        uu = 1;
+        goto divmod;
+    case '/':
+    case '%':
+    case TOK_PDIV:
+        uu = 0;
+    divmod:
         /* first operand must be in eax */
         /* XXX: need better constraint for second operand */
         gv2(RC_RAX, RC_RCX);
@@ -1037,30 +1825,13 @@ void gen_opi(int op)
         fr = vtop[0].r;
         vtop--;
         save_reg(TREG_RDX);
-        if (op == TOK_UMULL) {
-            o(0xf7); /* mul fr */
-            o(0xe0 + fr);
-            vtop->r2 = TREG_RDX;
+        orex(ll, 0, 0, uu ? 0xd231 : 0x99); /* xor %edx,%edx : cqto */
+        orex(ll, fr, 0, 0xf7); /* div fr, %eax */
+        o((uu ? 0xf0 : 0xf8) + REG_VALUE(fr));
+        if (op == '%' || op == TOK_UMOD)
+            r = TREG_RDX;
+        else
             r = TREG_RAX;
-        } else {
-            if (op == TOK_UDIV || op == TOK_UMOD) {
-                o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
-                o(0xf0 + fr);
-            } else {
-                if ((vtop->type.t & VT_BTYPE) & VT_LLONG) {
-                    o(0x9948); /* cqto */
-                    o(0x48 + REX_BASE(fr));
-                } else {
-                    o(0x99); /* cltd */
-                }
-                o(0xf7); /* idiv fr, %eax */
-                o(0xf8 + fr);
-            }
-            if (op == '%' || op == TOK_UMOD)
-                r = TREG_RDX;
-            else
-                r = TREG_RAX;
-        }
         vtop->r = r;
         break;
     default:
@@ -1075,7 +1846,7 @@ void gen_opl(int op)
 }
 
 /* generate a floating point operation 'v = t1 op t2' instruction. The
-   two operands are guaranted to have the same floating point type */
+   two operands are guaranteed to have the same floating point type */
 /* XXX: need to use ST1 too */
 void gen_opf(int op)
 {
@@ -1117,7 +1888,10 @@ void gen_opf(int op)
                 swapped = 0;
             if (swapped)
                 o(0xc9d9); /* fxch %st(1) */
-            o(0xe9da); /* fucompp */
+            if (op == TOK_EQ || op == TOK_NE)
+                o(0xe9da); /* fucompp */
+            else
+                o(0xd9de); /* fcompp */
             o(0xe0df); /* fnstsw %ax */
             if (op == TOK_EQ) {
                 o(0x45e480); /* and $0x45, %ah */
@@ -1161,7 +1935,7 @@ void gen_opf(int op)
                 break;
             }
             ft = vtop->type.t;
-            fc = vtop->c.ul;
+            fc = vtop->c.i;
             o(0xde); /* fxxxp %st, %st(1) */
             o(0xc1 + (a << 3));
             vtop--;
@@ -1170,13 +1944,13 @@ void gen_opf(int op)
         if (op >= TOK_ULT && op <= TOK_GT) {
             /* if saved lvalue, then we must reload it */
             r = vtop->r;
-            fc = vtop->c.ul;
+            fc = vtop->c.i;
             if ((r & VT_VALMASK) == VT_LLOCAL) {
                 SValue v1;
                 r = get_reg(RC_INT);
-                v1.type.t = VT_INT;
+                v1.type.t = VT_PTR;
                 v1.r = VT_LOCAL | VT_LVAL;
-                v1.c.ul = fc;
+                v1.c.i = fc;
                 load(r, &v1);
                 fc = 0;
             }
@@ -1194,31 +1968,29 @@ void gen_opf(int op)
             }
 
             if (swapped) {
-                o(0x7e0ff3); /* movq */
-                gen_modrm(1, r, vtop->sym, fc);
+                gv(RC_FLOAT);
+                vswap();
+            }
+            assert(!(vtop[-1].r & VT_LVAL));
+            
+            if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
+                o(0x66);
+            if (op == TOK_EQ || op == TOK_NE)
+                o(0x2e0f); /* ucomisd */
+            else
+                o(0x2f0f); /* comisd */
 
-                if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE) {
-                    o(0x66);
-                }
-                o(0x2e0f); /* ucomisd %xmm0, %xmm1 */
-                o(0xc8);
+            if (vtop->r & VT_LVAL) {
+                gen_modrm(vtop[-1].r, r, vtop->sym, fc);
             } else {
-                if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE) {
-                    o(0x66);
-                }
-                o(0x2e0f); /* ucomisd */
-                gen_modrm(0, r, vtop->sym, fc);
+                o(0xc0 + REG_VALUE(vtop[0].r) + REG_VALUE(vtop[-1].r)*8);
             }
 
             vtop--;
             vtop->r = VT_CMP;
-            vtop->c.i = op;
+            vtop->c.i = op | 0x100;
         } else {
-            /* no memory reference possible for long double operations */
-            if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
-                load(TREG_XMM0, vtop);
-                swapped = !swapped;
-            }
+            assert((vtop->type.t & VT_BTYPE) != VT_LDOUBLE);
             switch(op) {
             default:
             case '+':
@@ -1235,47 +2007,42 @@ void gen_opf(int op)
                 break;
             }
             ft = vtop->type.t;
-            fc = vtop->c.ul;
-            if ((ft & VT_BTYPE) == VT_LDOUBLE) {
-                o(0xde); /* fxxxp %st, %st(1) */
-                o(0xc1 + (a << 3));
+            fc = vtop->c.i;
+            assert((ft & VT_BTYPE) != VT_LDOUBLE);
+            
+            r = vtop->r;
+            /* if saved lvalue, then we must reload it */
+            if ((vtop->r & VT_VALMASK) == VT_LLOCAL) {
+                SValue v1;
+                r = get_reg(RC_INT);
+                v1.type.t = VT_PTR;
+                v1.r = VT_LOCAL | VT_LVAL;
+                v1.c.i = fc;
+                load(r, &v1);
+                fc = 0;
+            }
+            
+            assert(!(vtop[-1].r & VT_LVAL));
+            if (swapped) {
+                assert(vtop->r & VT_LVAL);
+                gv(RC_FLOAT);
+                vswap();
+            }
+            
+            if ((ft & VT_BTYPE) == VT_DOUBLE) {
+                o(0xf2);
             } else {
-                /* if saved lvalue, then we must reload it */
-                r = vtop->r;
-                if ((r & VT_VALMASK) == VT_LLOCAL) {
-                    SValue v1;
-                    r = get_reg(RC_INT);
-                    v1.type.t = VT_INT;
-                    v1.r = VT_LOCAL | VT_LVAL;
-                    v1.c.ul = fc;
-                    load(r, &v1);
-                    fc = 0;
-                }
-                if (swapped) {
-                    /* movq %xmm0,%xmm1 */
-                    o(0x7e0ff3);
-                    o(0xc8);
-                    load(TREG_XMM0, vtop);
-                    /* subsd  %xmm1,%xmm0 (f2 0f 5c c1) */
-                    if ((ft & VT_BTYPE) == VT_DOUBLE) {
-                        o(0xf2);
-                    } else {
-                        o(0xf3);
-                    }
-                    o(0x0f);
-                    o(0x58 + a);
-                    o(0xc1);
-                } else {
-                    if ((ft & VT_BTYPE) == VT_DOUBLE) {
-                        o(0xf2);
-                    } else {
-                        o(0xf3);
-                    }
-                    o(0x0f);
-                    o(0x58 + a);
-                    gen_modrm(0, r, vtop->sym, fc);
-                }
+                o(0xf3);
             }
+            o(0x0f);
+            o(0x58 + a);
+            
+            if (vtop->r & VT_LVAL) {
+                gen_modrm(vtop[-1].r, r, vtop->sym, fc);
+            } else {
+                o(0xc0 + REG_VALUE(vtop[0].r) + REG_VALUE(vtop[-1].r)*8);
+            }
+
             vtop--;
         }
     }
@@ -1310,17 +2077,17 @@ void gen_cvt_itof(int t)
         }
         vtop->r = TREG_ST0;
     } else {
-        save_reg(TREG_XMM0);
+        int r = get_reg(RC_FLOAT);
         gv(RC_INT);
-        o(0xf2 + ((t & VT_BTYPE) == VT_FLOAT));
+        o(0xf2 + ((t & VT_BTYPE) == VT_FLOAT?1:0));
         if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
             (VT_INT | VT_UNSIGNED) ||
             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
             o(0x48); /* REX */
         }
         o(0x2a0f);
-        o(0xc0 + (vtop->r & VT_VALMASK)); /* cvtsi2sd */
-        vtop->r = TREG_XMM0;
+        o(0xc0 + (vtop->r & VT_VALMASK) + REG_VALUE(r)*8); /* cvtsi2sd */
+        vtop->r = r;
     }
 }
 
@@ -1332,15 +2099,19 @@ void gen_cvt_ftof(int t)
     ft = vtop->type.t;
     bt = ft & VT_BTYPE;
     tbt = t & VT_BTYPE;
-
+    
     if (bt == VT_FLOAT) {
         gv(RC_FLOAT);
         if (tbt == VT_DOUBLE) {
-            o(0xc0140f); /* unpcklps */
-            o(0xc05a0f); /* cvtps2pd */
+            o(0x140f); /* unpcklps */
+            o(0xc0 + REG_VALUE(vtop->r)*9);
+            o(0x5a0f); /* cvtps2pd */
+            o(0xc0 + REG_VALUE(vtop->r)*9);
         } else if (tbt == VT_LDOUBLE) {
+            save_reg(RC_ST0);
             /* movss %xmm0,-0x10(%rsp) */
-            o(0x44110ff3);
+            o(0x110ff3);
+            o(0x44 + REG_VALUE(vtop->r)*8);
             o(0xf024);
             o(0xf02444d9); /* flds -0x10(%rsp) */
             vtop->r = TREG_ST0;
@@ -1348,29 +2119,37 @@ void gen_cvt_ftof(int t)
     } else if (bt == VT_DOUBLE) {
         gv(RC_FLOAT);
         if (tbt == VT_FLOAT) {
-            o(0xc0140f66); /* unpcklpd */
-            o(0xc05a0f66); /* cvtpd2ps */
+            o(0x140f66); /* unpcklpd */
+            o(0xc0 + REG_VALUE(vtop->r)*9);
+            o(0x5a0f66); /* cvtpd2ps */
+            o(0xc0 + REG_VALUE(vtop->r)*9);
         } else if (tbt == VT_LDOUBLE) {
+            save_reg(RC_ST0);
             /* movsd %xmm0,-0x10(%rsp) */
-            o(0x44110ff2);
+            o(0x110ff2);
+            o(0x44 + REG_VALUE(vtop->r)*8);
             o(0xf024);
             o(0xf02444dd); /* fldl -0x10(%rsp) */
             vtop->r = TREG_ST0;
         }
     } else {
+        int r;
         gv(RC_ST0);
+        r = get_reg(RC_FLOAT);
         if (tbt == VT_DOUBLE) {
             o(0xf0245cdd); /* fstpl -0x10(%rsp) */
             /* movsd -0x10(%rsp),%xmm0 */
-            o(0x44100ff2);
+            o(0x100ff2);
+            o(0x44 + REG_VALUE(r)*8);
             o(0xf024);
-            vtop->r = TREG_XMM0;
+            vtop->r = r;
         } else if (tbt == VT_FLOAT) {
             o(0xf0245cd9); /* fstps -0x10(%rsp) */
             /* movss -0x10(%rsp),%xmm0 */
-            o(0x44100ff3);
+            o(0x100ff3);
+            o(0x44 + REG_VALUE(r)*8);
             o(0xf024);
-            vtop->r = TREG_XMM0;
+            vtop->r = r;
         }
     }
 }
@@ -1400,11 +2179,8 @@ void gen_cvt_ftoi(int t)
     } else {
         assert(0);
     }
-    if (size == 8) {
-        o(0x48 + REX_BASE(r));
-    }
-    o(0x2c0f); /* cvttss2si or cvttsd2si */
-    o(0xc0 + (REG_VALUE(r) << 3));
+    orex(size == 8, r, 0, 0x2c0f); /* cvttss2si or cvttsd2si */
+    o(0xc0 + REG_VALUE(vtop->r) + REG_VALUE(r)*8);
     vtop->r = r;
 }
 
@@ -1415,5 +2191,39 @@ void ggoto(void)
     vtop--;
 }
 
+/* Save the stack pointer onto the stack and return the location of its address */
+ST_FUNC void gen_vla_sp_save(int addr) {
+    /* mov %rsp,addr(%rbp)*/
+    gen_modrm64(0x89, TREG_RSP, VT_LOCAL, NULL, addr);
+}
+
+/* Restore the SP from a location on the stack */
+ST_FUNC void gen_vla_sp_restore(int addr) {
+    gen_modrm64(0x8b, TREG_RSP, VT_LOCAL, NULL, addr);
+}
+
+/* Subtract from the stack pointer, and push the resulting value onto the stack */
+ST_FUNC void gen_vla_alloc(CType *type, int align) {
+#ifdef TCC_TARGET_PE
+    /* alloca does more than just adjust %rsp on Windows */
+    vpush_global_sym(&func_old_type, TOK_alloca);
+    vswap(); /* Move alloca ref past allocation size */
+    gfunc_call(1);
+#else
+    int r;
+    r = gv(RC_INT); /* allocation size */
+    /* sub r,%rsp */
+    o(0x2b48);
+    o(0xe0 | REG_VALUE(r));
+    /* We align to 16 bytes rather than align */
+    /* and ~15, %rsp */
+    o(0xf0e48348);
+    vpop();
+#endif
+}
+
+
 /* end of x86-64 code generator */
 /*************************************************************/
+#endif /* ! TARGET_DEFS_ONLY */
+/******************************************************/