preliminary pie and pic

7 files changed, 88 insertions, 25 deletions

diff --git a/amd64/emit.c b/amd64/emit.c
index df4c4cd..7c88955 100644
--- a/amd64/emit.c
+++ b/amd64/emit.c
@@ -240,14 +240,21 @@ opermatch(enum operpat pat, struct oper oper)
 #define DS(S) D(S, sizeof S - 1)
 
 static bool usebp; /* use RBP? */
+static const char *curfnsym;
+static uchar *fnstart;
 
 /* Given an instruction description table, find the first entry that matches
  * the operands (where dst, src are the operands in intel syntax order) and encode it */
 static void
 encode(uchar **pcode, const struct desc *tab, int ntab, enum irclass k, struct oper dst, struct oper src)
 {
-   const uchar *opc; int nopc, mod, rex;
-   bool sib = 0; uchar reg; struct oper mem; const struct desc *en = NULL;
+   const uchar *opc;
+   int nopc, mod, rex;
+   const char *sym;
+   uchar reg;
+   struct oper mem;
+   bool sib = 0;
+   const struct desc *en = NULL;
 
    for (int i = 0; i < ntab; ++i) {
       if ((tab[i].psiz & cls2siz[k]) && opermatch(tab[i].ptd, dst) && opermatch(tab[i].pts, src)) {
@@ -313,16 +320,17 @@ encode(uchar **pcode, const struct desc *tab, int ntab, enum irclass k, struct o
       else if (en->r8 && in_range(reg, RSP, RDI)) B(0x40);
 
       if (mem.t == OSYM) {
-         /* XXX PIC */
          D(opc, nopc);
          if (mem.cindex == NOINDEX) {
             /* %rip(var) */
             static uchar offs[NOPERENC] = { [EN_MI8] = 1, [EN_MI16] = 2, [EN_MI32] = 4 };
+            enum relockind r = (!conht[mem.con].deref && ccopt.pic) ? (rex ? REL_GOTPCRELX : REL_GOTPCRELX_REX) : REL_PCREL32;
             int off = -4 - offs[en->operenc];
             B(/*mod 0*/ (reg & 7) << 3 | RBP);
-            objreloc(xcon2sym(mem.con), REL_PCREL32, Stext, *pcode - objout.textbegin, mem.disp + off);
+            objreloc(xcon2sym(mem.con), r, Stext, *pcode - objout.textbegin, mem.disp + off);
          } else {
             /* var(,%reg,shift) */
+            assert(!ccopt.pic && !ccopt.pie && "cannot encode [RIP-rel + REG] for position independent");
             B(/*mod 0*/ (reg & 7) << 3 | RSP);
             B(mem.cshift << 6 | mem.cindex << 3 | RBP); /* SIB [index*s + disp32] */
             objreloc(xcon2sym(mem.con), REL_ABS32S, Stext, *pcode - objout.textbegin, mem.disp);
@@ -393,8 +401,15 @@ encode(uchar **pcode, const struct desc *tab, int ntab, enum irclass k, struct o
       if (rex) B(0x40 | rex);
       D(opc, nopc);
       assert(dst.t == OSYM);
-      objreloc(xcon2sym(dst.con), REL_PCREL32, Stext, *pcode - objout.textbegin, -4);
-      I32(0);
+      sym = xcon2sym(dst.con);
+      if (sym != curfnsym) {
+         enum relockind r = (ccopt.pie|ccopt.pic) ? REL_PLT32 : REL_PCREL32;
+         objreloc(sym, r, Stext, *pcode - objout.textbegin, -4);
+         I32(0);
+      } else {
+         /* self-recursive call */
+         I32(fnstart - *pcode - 4);
+      }
       break;
    }
 }
@@ -478,6 +493,7 @@ DEFINSTR2(Xxchg,
 )
 DEFINSTR2(Xlea,
    {4|8, PGPR, PMEM, "\x8D", EN_RM}, /* LEA r32/64,m32/64 */
+   {  8, PGPR, PSYM, "\x8D", EN_RM}, /* LEA rel32 */
 )
 DEFINSTR2(Xadd,
    {4|8, PGPR, PGPR, "\x03", EN_RR},   /* ADD r32/64, r32/64 */
@@ -764,12 +780,22 @@ gencopy(uchar **pcode, enum irclass cls, struct block *blk, int curi, struct ope
       }
       /* normal (not 2-address) case */
    Lea:
+      if (isaddrcon(addr->base) && ccopt.pic) {
+         assert(!addr->disp && !addr->index.bits);
+         val = addr->base;
+         goto GOTLoad;
+      }
       Xlea(pcode, cls, dst, ref2oper(val));
    } else if (val.bits == ZEROREF.bits && dst.t == OREG && !flagslivep(blk, curi)) {
       /* dst = 0 -> xor dst, dst; but only if it is ok to clobber flags */
       Xxor(pcode, kisint(cls) ? KI4 : cls, dst, dst);
    } else if (isaddrcon(val)) {
-      Xlea(pcode, cls, dst, mkoper(OSYM, .con = val.i, .cindex = NOINDEX));
+      if (ccopt.pic) GOTLoad:
+         /* for  mov reg, [rip(sym@GOTPCREL)] */ 
+         Xmov(pcode, cls, dst, mkoper(OSYM, .con = val.i, .cindex = NOINDEX));
+      else
+         /* for  lea reg, [rip(sym)] */ 
+         Xlea(pcode, cls, dst, mkoper(OSYM, .con = val.i, .cindex = NOINDEX));
    } else {
       struct oper src = mkimmdatregoper(val);
       if (memcmp(&dst, &src, sizeof dst) != 0)
@@ -1067,7 +1093,6 @@ emitbin(struct function *fn)
 {
    struct block *blk;
    uchar **pcode = &objout.code;
-   uchar *start;
    int npush = 0;
    uint epilogueaddr = 0;
    bool saverestore;
@@ -1079,7 +1104,8 @@ emitbin(struct function *fn)
    memset(blkaddr, 0, nblkaddr * sizeof *blkaddr);
 
    nops(pcode, 16);
-   start = *pcode;
+   fnstart = *pcode;
+   curfnsym = fn->name;
 
    /** prologue **/
 
@@ -1129,7 +1155,7 @@ emitbin(struct function *fn)
       }
       if (blk->jmp.t == Jret) {
          /* epilogue */
-         uint here = *pcode - start;
+         uint here = *pcode - fnstart;
          if (epilogueaddr) {
             int disp = epilogueaddr - (here + 2);
             if ((uint)(disp + 128) < 256) {/* can use 1-byte displacement? */
@@ -1149,7 +1175,7 @@ emitbin(struct function *fn)
          }
       } else emitbranch(pcode, blk);
    } while ((blk = blk->lnext) != fn->entry);
-   objdeffunc(fn->name, fn->globl, start - objout.textbegin, *pcode - start);
+   objdeffunc(fn->name, fn->globl, fnstart - objout.textbegin, *pcode - fnstart);
 }
 
 void
diff --git a/amd64/isel.c b/amd64/isel.c
index fcd4ab3..1054502 100644
--- a/amd64/isel.c
+++ b/amd64/isel.c
@@ -44,6 +44,13 @@ static const uchar opflags[] = {
 static int iflagsrc = -1;
 
 static void
+picfixsym(union ref *r, struct block *blk, int *curi)
+{
+   if (!ccopt.pic || !isaddrcon(*r)) return;
+   *r = insertinstr(blk, (*curi)++, mkinstr(Ocopy, KPTR, .l = *r));
+}
+
+static void
 fixarg(union ref *r, struct instr *ins, struct block *blk, int *curi)
 {
    int sh;
@@ -72,6 +79,7 @@ fixarg(union ref *r, struct instr *ins, struct block *blk, int *curi)
             wr32le(data, pun.i);
          }
          *r = mkdatref(NULL, siz, /*align*/siz, data, siz, /*deref*/1);
+         picfixsym(r, blk, curi);
       } else if (in_range(op, Odiv, Ourem) && kisint(ins->cls))
          goto DivImm;
    } else if (r->t == RICON && in_range(op, Odiv, Ourem) && kisint(ins->cls)) {
@@ -195,9 +203,7 @@ aadd(struct addr *addr, union ref r)
    } else if (isnumcon(r)) {
       return acon(addr, r);
    } else if (isaddrcon(r)) {
-      /* XXX PIC */
       if (!addr->base.bits && !isaddrcon(addr->index)) addr->base = r;
-      else if (!addr->index.bits && !isaddrcon(addr->base)) addr->index = r;
       else return 0;
    } else if (r.t == RREG) {
       /* temporaries are single assignment, but register aren't, so they can't be *
@@ -212,7 +218,7 @@ aadd(struct addr *addr, union ref r)
 }
 
 static bool
-fuseaddr(union ref *r)
+fuseaddr(union ref *r, struct block *blk, int *curi)
 {
    struct addr addr = { 0 };
 
@@ -221,6 +227,15 @@ fuseaddr(union ref *r)
    if (r->t != RTMP) return 0;
    if (!aadd(&addr, *r)) return 0;
 
+   if (ccopt.pic || (ccopt.pie && isaddrcon(addr.base) && addr.index.bits)) {
+      /* pic needs to load from GOT */
+      /* pie cannot encode RIP-relative address with index register */
+      /*     first load symbol address into a temp register */
+      union ref temp = mkaddr((struct addr){.base = addr.base, .disp = ccopt.pic ? 0 : addr.disp});
+      addr.base = insertinstr(blk, (*curi)++, mkinstr(Ocopy, KPTR, .l = temp));
+      if (!ccopt.pic) addr.disp = 0;
+   }
+
    *r = mkaddr(addr);
    return 1;
 }
@@ -318,7 +333,7 @@ sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
          /* sub x,1 -> dec x */
          ins->op = op = Oxdec;
          ins->r = NOREF;
-      } else if (iscon(ins->l)) {
+      } else if (isintcon(ins->l)) {
          /* sub imm, x -> sub x, imm; neg x */
          struct instr sub = *ins;
          rswap(sub.l, sub.r);
@@ -333,7 +348,7 @@ sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
             temp.op = Ocopy;
             temp.cls = ins->cls;
             temp.l = mkref(RTMP, ins - instrtab);
-            if (fuseaddr(&temp.l)) {
+            if (fuseaddr(&temp.l, blk, curi)) {
                *ins = temp;
                break;
             }
@@ -362,19 +377,21 @@ sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
       if (!(in_range(op, Omul, Oumul) && kisint(ins->cls) && isimm32(ins->r))) /* for (I)MUL r,r/m,imm */
          ins->inplace = 1;
       if (iscon(ins->l))
-         ins->l = insertinstr(blk, (*curi)++, mkinstr(Ocopy, concls(ins->l), ins->l));
+         ins->l = insertinstr(blk, (*curi)++, mkinstr(Ocopy, ins->cls, ins->l));
       if (ins->r.bits)
    case Omove:
          fixarg(&ins->r, ins, blk, curi);
       break;
    case Oloads1: case Oloadu1: case Oloads2: case Oloadu2:
    case Oloads4: case Oloadu4: case Oloadi8: case Oloadf4: case Oloadf8:
-      if (!fuseaddr(&ins->l) && ins->l.t != RTMP && ins->l.t != RREG)
+      if (!fuseaddr(&ins->l, blk, curi) && ins->l.t != RTMP && ins->l.t != RREG)
          ins->l = insertinstr(blk, (*curi)++, mkinstr(Ocopy, KPTR, ins->l));
+      picfixsym(&ins->l, blk, curi);
       break;
    case Ostore1: case Ostore2: case Ostore4: case Ostore8:
-      if (!fuseaddr(&ins->l) && ins->l.t != RTMP && ins->l.t != RREG)
+      if (!fuseaddr(&ins->l, blk, curi) && ins->l.t != RTMP && ins->l.t != RREG)
          ins->l = insertinstr(blk, (*curi)++, mkinstr(Ocopy, KPTR, ins->l));
+      picfixsym(&ins->l, blk, curi);
       fixarg(&ins->r, ins, blk, curi);
       break;
    case Ocvtf4f8: case Ocvtf8f4: case Ocvtf4s: case Ocvtf8s:
diff --git a/c.c b/c.c
index 6833722..2d484c4 100644
--- a/c.c
+++ b/c.c
@@ -2447,10 +2447,11 @@ genptroff(struct function *fn, enum op op, uint siz, union ref ptr,
    assert(siz);
 
    idx = cvt(fn, targ_sizetype, tt, idx);
-   if (siz ==  1) off = idx;
-   else if (idx.t == RICON)
-      off = mkintcon(cls, idx.i * siz);
-   else if (ispo2(siz))
+   if (siz == 1) off = idx;
+   else if (idx.t == RICON) {
+      if (op == Osub) op = Oadd, idx.i = -idx.i;
+      off = mkintcon(cls, idx.i * (int)siz);
+   } else if (ispo2(siz))
       off = addinstr(fn,
             mkinstr(Oshl, cls, .l = idx, .r = mkintcon(cls, ilog2(siz))));
    else
diff --git a/common.h b/common.h
index 11c4464..549822a 100644
--- a/common.h
+++ b/common.h
@@ -122,6 +122,7 @@ struct option {
    bool pedant;
    bool trigraph;
    bool nocolor;
+   bool pie, pic;
    union {
       struct {
          bool p : 1, /* after parsing */
diff --git a/elf.c b/elf.c
index 399434b..7620961 100644
--- a/elf.c
+++ b/elf.c
@@ -118,6 +118,9 @@ static const ushort relktab[][NRELOCKIND] = {
       [REL_ABS32]   = 10, /* R_X86_64_32 */
       [REL_ABS32S]  = 11, /* R_X86_64_32S */
       [REL_PCREL32] = 2, /* R_X86_64_PC32 */
+      [REL_PLT32]   = 4, /* R_X86_64_PLT32 */
+      [REL_GOTPCRELX]     = 41, /* R_X86_64_GOTPCRELX */
+      [REL_GOTPCRELX_REX] = 42, /* R_X86_64_REX_GOTPCRELX */
    }
 };
 
@@ -132,7 +135,7 @@ elfreloc(const char *sym, enum relockind kind, enum section section, uint off, v
    case Sdata:   ++ndatarel; break;
    }
    assert(kind < NRELOCKIND);
-   vpush(&relocs, ((struct reloc) { section, relktab[mctarg->isa][kind], snam, off, addend}));
+   vpush(&relocs, ((struct reloc) { section, relktab[mctarg->isa][kind], snam, off, addend }));
 }
 
 static void
diff --git a/main.c b/main.c
index bf178d3..a51eae2 100644
--- a/main.c
+++ b/main.c
@@ -131,6 +131,16 @@ optparse(char **args)
          if (arg[1]) task.out = arg+1;
          else if (args[1]) task.out = *++args;
          else goto Bad;
+      } else if (*arg == 'f') {
+         const char *flag = arg+1;
+         bool set = 1;
+         if (!strncmp(flag, "no-", 3)) {
+            set = 0;
+            flag += 3;
+         }
+         if (!strcmp(flag, "pie") || !strcmp(flag, "PIE")) ccopt.pie = set;
+         else if (!strcmp(flag, "pic") || !strcmp(flag, "PIC")) ccopt.pic = set;
+         else goto Bad;
       } else if (!strcmp(arg, "v") || !strcmp(arg, "-verbose")) {
          task.verbose = 1;
       } else if (!strcmp(arg, "c")) {
@@ -242,8 +252,9 @@ dolink(void)
    vpush(&cmd, "/bin/gcc");
    if (task.outft == OFTdll) {
       vpush(&cmd, "-shared");
+   } else if (task.outft == OFTexe) {
+      vpush(&cmd, ccopt.pie ? "-pie" : "-no-pie");
    }
-   vpush(&cmd, "-no-pie");
    vpush(&cmd, "-o");
    vpush(&cmd, task.out);
    assert(task.ninf > 0);
@@ -335,6 +346,7 @@ main(int argc, char **argv)
    /* setup defaults */
    detectcolor();
    ccopt.cstd = STDC99;
+   ccopt.pie = 1;
 
    /* parse cli ags */
    optparse(argv);
diff --git a/obj.h b/obj.h
index 2c4753d..750ad48 100644
--- a/obj.h
+++ b/obj.h
@@ -14,6 +14,9 @@ enum relockind {
    REL_ABS32,
    REL_ABS32S,
    REL_PCREL32,
+   REL_PLT32,
+   REL_GOTPCRELX,
+   REL_GOTPCRELX_REX,
    NRELOCKIND,
 };
 enum section { Snone, Stext, Srodata, Sdata, Sbss };