backend: start implementing aarch64

1 files changed, 440 insertions, 0 deletions

diff --git a/aarch64/isel.c b/aarch64/isel.c
new file mode 100644
index 0000000..a61fa21
--- /dev/null
+++ b/aarch64/isel.c
@@ -0,0 +1,440 @@
+#include "all.h"
+
+/* map alloca tmp -> stack frame displacement (0 if not alloca) */
+static ushort *stkslots;
+static uint nstkslots;
+
+#define isstkslot(r) ((r).t == RTMP && (r).i < nstkslots && stkslots[(r).i])
+#define isimm32(r) (iscon(r) && concls(r) == KI32)
+
+static void
+picfixsym(union ref *r, struct block *blk, int *curi)
+{
+   if (!ccopt.pic || !isaddrcon(*r,0)) return;
+   *r = insertinstr(blk, (*curi)++, mkinstr(Ocopy, KPTR, .l = *r));
+}
+
+static inline uint
+clz(uvlong x)
+{
+#if HAS_BUILTIN(clzll)
+   return __builtin_clzll(x);
+#else
+   int i = 0;
+   for (uvlong mask = BIT(63);; ++i, mask >>= 1)
+      if (x & mask)
+         break;
+   return i;
+#endif
+}
+
+/* Encode logical immediate */
+bool
+aarch64_logimm(uint *enc, enum irclass k, uvlong x)
+{
+   /* https://github.com/v8/v8/blob/927ccc6076e25a614787c7011315468e40fe39a4/src/codegen/arm64/assembler-arm64.cc#L4409 */
+   if (k == KI32) x = (uint)x | x << 32;
+   bool neg;
+   if ((neg = x & 1)) x = ~x;
+   if (x == 0) return 0;
+   uvlong a = x & (~x + 1),
+          xa = x + a,
+          b = xa & (~xa + 1),
+          xa_b = xa - b,
+          c = xa_b & (~xa_b + 1),
+          mask;
+   uint clza = clz(a),
+        d, outn;
+   if (c != 0) {
+      d = clza - clz(c);
+      mask = BIT(d) - 1;
+      outn = 0;
+   } else {
+      assert(a != 0);
+      d = 64;
+      mask = ~0ull;
+      outn = 1;
+   }
+   if (!ispo2(d)) return 0;
+   if (((b - a) & ~mask) != 0) return 0;
+   static const uvlong M[] = {
+      0x0000000000000001, 0x0000000100000001, 0x0001000100010001,
+      0x0101010101010101, 0x1111111111111111, 0x5555555555555555,
+   };
+   int i = clz(d) - 57;
+   assert((uint)i < countof(M));
+   uvlong m = M[i];
+   uvlong y = (b - a) * m;
+   if (y != x) return 0;
+   if (enc) {
+      int clzb = b == 0 ? -1 : clz(b),
+          s = clza - clzb,
+          r;
+      if (neg) {
+         s = d - s;
+         r = (clzb + 1) & (d - 1);
+      } else {
+         r = (clza + 1) & (d - 1);
+      }
+      *enc = outn<<12 | r<<6 | (((-d * 2) | (s - 1)) & 0x3F);
+   }
+   return 1;
+
+}
+
+static void
+fixarg(union ref *r, struct instr *ins, struct block *blk, int *curi)
+{
+   enum op op = ins ? ins->op : 0;
+   if (isintcon(ins->r)) {
+      vlong x = intconval(ins->r);
+      switch (op) {
+      default:
+         if (oiscmp(op)) {
+         case Oadd: case Osub:
+            /* imm12 (lsl 12) */
+            if ((x &~ 0xFFF) == 0 || (x &~ 0xFFF000) == 0) return;
+            break;
+         case Oshl: case Osar: case Oslr:
+            if ((uvlong)x < (ins->cls == KI32 ? 32 : 64)) return;
+            break;
+         case Oand: case Oior: case Oxor:
+            if (aarch64_logimm(NULL, ins->cls, x)) return;
+            break;
+         }
+      }
+      goto Copy;
+   } else if (isstkslot(*r)) {
+      struct instr adr = mkinstr(Oadd, KPTR, mkref(RREG, FP), mkintcon(KI32, -stkslots[r->i]));
+      if (ins && ins->op == Ocopy)
+         *ins = adr;
+      else
+         *r = insertinstr(blk, (*curi)++, adr);
+   } else if (r->t != RTMP) Copy: {
+      *r = insertinstr(blk, (*curi)++, mkinstr(Ocopy, r->t == RTMP ? instrtab[r->i].cls : ins->cls ? ins->cls : KI32, *r));
+   }
+}
+
+static bool
+arithfold(struct instr *ins)
+{
+   if (isnumcon(ins->l) && (!ins->r.t || isnumcon(ins->r))) {
+      union ref r;
+      bool ok = ins->r.t ? foldbinop(&r, ins->op, ins->cls, ins->l, ins->r) : foldunop(&r, ins->op, ins->cls, ins->l);
+      assert(ok && "fold?");
+      *ins = mkinstr(Ocopy, insrescls(*ins), r);
+      return 1;
+   }
+   return 0;
+}
+
+static void
+selcall(struct function *fn, struct instr *ins, struct block *blk, int *curi)
+{
+   const struct call *call = &calltab.p[ins->r.i];
+   int iarg = *curi - 1;
+   enum irclass cls;
+   uint argstksiz = alignup(call->argstksiz, 16);
+
+   for (int i = call->narg - 1; i >= 0; --i) {
+      struct abiarg abi = call->abiarg[i];
+      struct instr *arg;
+      for (;; --iarg) {
+         assert(iarg >= 0 && i >= 0 && "arg?");
+         if ((arg = &instrtab[blk->ins.p[iarg]])->op == Oarg)
+            break;
+      }
+
+      if (!abi.isstk) {
+         assert(!abi.ty.isagg);
+         *arg = mkinstr(Omove, call->abiarg[i].ty.cls, mkref(RREG, abi.reg), arg->r);
+      } else {
+         union ref adr = mkaddr((struct addr){mkref(RREG, SP), .disp = abi.stk});
+         int iargsave = iarg;
+         if (!abi.ty.isagg) { /* scalar arg in stack */
+            *arg = mkinstr(Ostore8+ilog2(cls2siz[abi.ty.cls]), 0, adr, arg->r);
+            if (isaddrcon(arg->r,1) || arg->r.t == RADDR)
+               arg->r = insertinstr(blk, iarg++, mkinstr(Ocopy, abi.ty.cls, arg->r));
+            else
+               fixarg(&ins->r, ins, blk, &iarg);
+         } else { /* aggregate arg in stack, callee stack frame destination address */
+            *arg = mkinstr(Ocopy, KPTR, adr);
+         }
+         *curi += iarg - iargsave;
+      }
+   }
+   if (call->argstksiz) {
+      union ref disp = mkref(RICON, argstksiz);
+      insertinstr(blk, iarg--, (struct instr){Osub, KPTR, .keep=1, .reg = SP+1, .l=mkref(RREG,SP), disp});
+      ++*curi;
+      insertinstr(blk, *curi+1, (struct instr){Oadd, KPTR, .keep=1, .reg = SP+1, .l=mkref(RREG,SP), disp});
+   }
+   if (isimm32(ins->l))
+      ins->l = mkaddr((struct addr){.base = ins->l});
+   else if (isintcon(ins->l))
+      ins->l = insertinstr(blk, (*curi)++, mkinstr(Ocopy, KPTR, ins->l));
+
+   cls = ins->cls;
+   ins->cls = 0;
+   if (cls) {
+      /* duplicate to reuse same TMP ref */
+      insertinstr(blk, (*curi)++, *ins);
+      *ins = mkinstr(Ocopy, cls, mkref(RREG, call->abiret[0].reg));
+      for (int i = 1; i <= 2; ++i) {
+         if (*curi + i >= blk->ins.n) break;
+         if (instrtab[blk->ins.p[*curi + i]].op == Ocall2r) {
+            ins = &instrtab[blk->ins.p[*curi += i]];
+            *ins = mkinstr(Ocopy, ins->cls, mkref(RREG, call->abiret[1].reg));
+            break;
+         }
+      }
+   }
+}
+
+static bool
+aimm(struct addr *addr, int disp)
+{
+   if (addr->index.bits) return 0;
+   vlong a = addr->disp;
+   a += disp;
+   if ((int)a == a) {
+      addr->disp = a;
+      return 1;
+   }
+   return 0;
+}
+
+static bool
+ascale(struct addr *addr, union ref a, union ref b, uint siz/*1,2,4,8*/)
+{
+   if (b.t != RICON) return 0;
+   if (addr->index.bits || addr->disp) return 0;
+   if ((unsigned)b.i > 3 || 1<<b.i != siz) return 0;
+   if (a.t == RREG || a.t == RTMP) {
+      addr->index = a;
+      addr->shift = b.i;
+      return 1;
+   }
+   return 0;
+}
+
+static bool
+aadd(struct addr *addr, struct block *blk, int *curi, union ref r, uint siz/*1,2,4,8*/)
+{
+   if (r.t == RSTACK) {
+      if (addr->base.bits || addr->index.bits || !aimm(addr, -r.i)) goto Ref;
+      addr->base = mkref(RREG, FP);
+   } else if (r.t == RTMP) {
+      struct instr *ins = &instrtab[r.i];
+      if (ins->op == Oadd) {
+         if (!aadd(addr, blk, curi, ins->l, siz)) goto Ref;
+         if (!aadd(addr, blk, curi, ins->r, siz)) goto Ref;
+         ins->skip = 1;
+      } else if (ins->op == Osub) {
+         if (!aadd(addr, blk, curi, ins->l, siz)) goto Ref;
+         if (!isintcon(ins->r)) goto Ref;
+         if (!aimm(addr, -intconval(ins->r))) goto Ref;
+         ins->skip = 1;
+      } else if (ins->op == Oshl) {
+         if (!ascale(addr, ins->l, ins->r, siz)) goto Ref;
+         ins->skip = 1;
+      } else if (ins->op == Ocopy) {
+         if (!aadd(addr, blk, curi, ins->l, siz)) goto Ref;
+         ins->skip = 1;
+      } else goto Ref;
+   } else if (isnumcon(r)) {
+      assert(isintcon(r));
+      return aimm(addr, intconval(r));
+   } else if (isaddrcon(r,1)) {
+      if (!addr->base.bits && !isaddrcon(addr->index,1)) addr->base = r;
+      else return 0;
+   } else if (r.t == RREG) {
+      /* temporaries are single assignment, but register aren't, so they can't be *
+       * safely hoisted into an address value, unless they have global lifetime */
+      if (!rstest(mctarg->rglob, r.i)) return 0;
+   Ref:
+      if (r.t == RSTACK && (addr->base.bits || addr->index.bits)) {
+         r = insertinstr(blk, (*curi)++, mkinstr(Oadd, KPTR, mkref(RREG, FP), mkref(RICON, -r.i)));
+      }
+      if (!addr->base.bits) addr->base = r;
+      else if (!addr->index.bits) addr->index = r;
+      else return 0;
+   } else return 0;
+   return 1;
+}
+
+static bool
+fuseaddr(union ref *r, struct block *blk, int *curi, uint siz/*1,2,4,8*/)
+{
+   struct addr addr = {0};
+
+   if (isaddrcon(*r,1)) return 1;
+
+   if (r->t != RSTACK && r->t != RTMP) return 0;
+   if (!aadd(&addr, blk, curi, *r, siz)) return 0;
+   if (isaddrcon(addr.base,0) && (ccopt.pic || (ccopt.pie && addr.index.bits) || (conht[addr.base.i].flag & SFUNC))) {
+      /* 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;
+   }
+   if (!(addr.disp >= -256 && addr.disp < 256) /* for 9-bit signed unscaled offset */
+    && !(!(addr.disp & (siz-1)) && (uvlong)addr.disp < (1<<12)*siz)) /* 12-bit unsigned scaled offset */
+      return 0;
+   *r = mkaddr(addr);
+   return 1;
+}
+
+
+static void
+loadstoreaddr(struct block *blk, union ref *r, int *curi, uint siz)
+{
+   if (isimm32(*r)) {
+      *r = mkaddr((struct addr){.base = *r});
+   } else if (isaddrcon(*r, 0)) {
+      picfixsym(r, blk, curi);
+   } else if (r->t == RTMP || r->t == RSTACK) {
+      fuseaddr(r, blk, curi, siz);
+   } else if (r->t != RREG) {
+      *r = insertinstr(blk, (*curi)++, mkinstr(Ocopy, KPTR, *r));
+   }
+}
+
+static void
+sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
+{
+   uint siz, alignlog2;
+   int t = ins - instrtab;
+   struct instr temp = {0};
+   enum op op = ins->op;
+
+   if (oisarith(ins->op) && arithfold(ins)) {
+      fixarg(&ins->l, ins, blk, curi);
+      return;
+   }
+
+   switch (op) {
+   //default: assert(0);
+   case Onop: break;
+   case Oalloca1: case Oalloca2: case Oalloca4: case Oalloca8: case Oalloca16:
+      alignlog2 = ins->op - Oalloca1;
+      assert(ins->l.i > 0);
+      siz = ins->l.i << alignlog2;
+      fn->stksiz += siz;
+      fn->stksiz = alignup(fn->stksiz, 1 << alignlog2);
+      if (fn->stksiz > (1<<16)-1) error(NULL, "'%s' stack frame too big", fn->name);
+      stkslots[t] = fn->stksiz;
+      *ins = mkinstr(Onop,0,);
+      break;
+   case Oparam:
+      assert(ins->l.t == RICON && ins->l.i < fn->nabiarg);
+      if (!fn->abiarg[ins->l.i].isstk)
+         *ins = mkinstr(Ocopy, ins->cls, mkref(RREG, fn->abiarg[ins->l.i].reg));
+      else /* stack */
+         *ins = mkinstr(Oadd, KPTR, mkref(RREG, FP), mkref(RICON, 16+fn->abiarg[ins->l.i].stk));
+      break;
+   case Oadd: case Osub:
+      if (ins->r.t == RICON && ins->r.i < 0) {
+         op = ins->op ^= 1;
+         ins->r.i = -ins->r.i;
+      }
+      fixarg(&ins->l, ins, blk, curi);
+      fixarg(&ins->r, ins, blk, curi);
+      break;
+   case Oand: case Oior: case Oxor:
+   case Oshl: case Osar: case Oslr:
+      fixarg(&ins->r, ins, blk, curi);
+      break;
+   case Oarg:
+      fixarg(&ins->r, ins, blk, curi);
+      break;
+   case Ocall:
+      selcall(fn, ins, blk, curi);
+      break;
+   case Oloads8: case Oloadu8: case Oloads16: case Oloadu16:
+   case Oloads32: case Oloadu32: case Oloadi64:
+      loadstoreaddr(blk, &ins->l, curi, 1<<((op - Oloads8)/2));
+      break;
+   case Ostore8: case Ostore16: case Ostore32: case Ostore64:
+      loadstoreaddr(blk, &ins->l, curi, 1<<(op - Ostore8));
+      fixarg(&ins->r, ins, blk, curi);
+      break;
+   }
+}
+
+static void
+seljmp(struct function *fn, struct block *blk)
+{
+   if (blk->jmp.t == Jb && blk->jmp.arg[0].bits) {
+      int curi = blk->ins.n;
+      fixarg(&blk->jmp.arg[0], NULL, blk, &curi);
+      union ref c = blk->jmp.arg[0];
+      if (c.t != RTMP) {
+         enum irclass cls = c.t == RICON ? KI32 : c.t == RXCON && conht[c.i].cls ? conht[c.i].cls : KPTR;
+         int curi = blk->ins.n;
+
+         c = insertinstr(blk, blk->ins.n, mkinstr(Ocopy, cls, c));
+         sel(fn, &instrtab[c.i], blk, &curi);
+      }
+      if (!oiscmp(instrtab[c.i].op)) {
+         struct instr *ins;
+         int curi = blk->ins.n;
+         blk->jmp.arg[0] = insertinstr(blk, blk->ins.n, mkinstr(Oneq, insrescls(instrtab[c.i]), c, ZEROREF));
+         ins = &instrtab[blk->jmp.arg[0].i];
+         if (kisflt(ins->cls)) {
+            ins->r = insertinstr(blk, curi, mkinstr(Ocopy, ins->cls, ZEROREF));
+         }
+         ins->keep = 1;
+      }
+   } else if (blk->jmp.t == Jret) {
+      if (blk->jmp.arg[0].bits) {
+         int curi;
+         union ref r = mkref(RREG, fn->abiret[0].reg);
+         struct instr *ins = &instrtab[insertinstr(blk, blk->ins.n, mkinstr(Omove, fn->abiret[0].ty.cls, r, blk->jmp.arg[0])).i];
+         curi = blk->ins.n-1;
+         fixarg(&ins->r, ins, blk, &curi);
+         blk->jmp.arg[0] = r;
+         if (blk->jmp.arg[1].bits) {
+            r = mkref(RREG, fn->abiret[1].reg);
+            ins = &instrtab[insertinstr(blk, blk->ins.n, mkinstr(Omove, fn->abiret[1].ty.cls, r, blk->jmp.arg[1])).i];
+         }
+      }
+   }
+}
+
+void
+aarch64_isel(struct function *fn)
+{
+   extern int ninstr;
+   struct block *blk = fn->entry;
+
+   fn->stksiz = 0;
+   stkslots = allocz(fn->passarena, (nstkslots = ninstr) * sizeof *stkslots, 0);
+   do {
+      int i;
+      for (i = 0; i < blk->phi.n; ++i) {
+         struct instr *ins = &instrtab[blk->phi.p[i]];
+         union ref *phi = phitab.p[ins->l.i];
+         for (int i = 0; i < blk->npred; ++i) {
+            int curi = blkpred(blk, i)->ins.n;
+            fixarg(&phi[i], ins, blkpred(blk, i), &curi);
+         }
+      }
+      for (i = 0; i < blk->ins.n; ++i) {
+         struct instr *ins = &instrtab[blk->ins.p[i]];
+         sel(fn, ins, blk, &i);
+      }
+      seljmp(fn, blk);
+   } while ((blk = blk->lnext) != fn->entry);
+
+   if (ccopt.dbg.i) {
+      bfmt(ccopt.dbgout, "<< After isel >>\n");
+      irdump(fn);
+   }
+
+   fn->prop = 0;
+}
+
+/* vim:set ts=3 sw=3 expandtab: */