implement float varargs, and some other fixes

7 files changed, 93 insertions, 53 deletions

diff --git a/amd64/isel.c b/amd64/isel.c
index 9606c87..51a58db 100644
--- a/amd64/isel.c
+++ b/amd64/isel.c
@@ -141,6 +141,12 @@ selcall(struct function *fn, struct instr *ins, struct block *blk, int *curi)
       } else {
          union ref adr = mkaddr((struct addr){mkref(RREG, RSP), .disp = abi.stk});
          *arg = mkinstr(Ostore8+ilog2(cls2siz[abi.ty.cls]), 0, adr, arg->r);
+         int iargsave = iarg;
+         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);
+         *curi += iarg - iargsave;
       }
    }
    if (call->argstksiz) {
diff --git a/amd64/sysv.c b/amd64/sysv.c
index 9d97b86..9d356a1 100644
--- a/amd64/sysv.c
+++ b/amd64/sysv.c
@@ -74,7 +74,7 @@ abiarg(short r[2], uchar cls[2], int *ni, int *nf, int *ns, union irtype typ)
    if (!typ.isagg) {
       if (kisflt(cls[0] = typ.cls) && *nf < NFLT) {
          r[0] = XMM0 + (*nf)++;
-      } else if (*ni < NINT) {
+      } else if (kisint(cls[0]) && *ni < NINT) {
          r[0] = intregs[(*ni)++];
       } else {
          r[0] = *ns;
@@ -204,7 +204,7 @@ vastart(struct function *fn, struct block *blk, int *curi)
    insertinstr(blk, i++, mkinstr(Ostore32, 0, ap, mkref(RICON, gpr0*8)));
    /* set ap->fp_offset */
    dst = insertinstr(blk, i++, mkinstr(Oadd, KPTR, ap, mkref(RICON, 4)));
-   insertinstr(blk, i++, mkinstr(Ostore32, 0, dst, mkref(RICON, 6*8 + fpr0*8)));
+   insertinstr(blk, i++, mkinstr(Ostore32, 0, dst, mkref(RICON, 6*8 + fpr0*16)));
    *curi = i;
 }
 
@@ -228,51 +228,50 @@ vaarg(struct function *fn, struct block *blk, int *curi)
    else if (ret == 1) {
       struct block *merge;
       union ref phi, phiargs[2];
-      if (ni) {
-         /* l->gp_offset < 48 - num_gp * 8 */
-         tmp = insertinstr(blk, (*curi)++, mkinstr(Oloadu32, KI32, ap));
-         tmp = insertinstr(blk, (*curi)++, mkinstr(Oulte, KI32, tmp, mkref(RICON, 48 - ni*8)));
-         merge = blksplitafter(fn, blk, *curi);
-         blk->jmp.t = 0;
-         useblk(fn, blk);
-         putcondbranch(fn, tmp, newblk(fn), newblk(fn));
-         useblk(fn, blk->s1);
-         {
-            /* phi0: &l->reg_save_area[l->gp_offset] */
-            union ref sav = addinstr(fn, mkinstr(Oloadi64, KPTR, irbinop(fn, Oadd, KPTR, ap, mkref(RICON, 16))));
-            union ref gpoff = addinstr(fn, mkinstr(Oloadu32, KI32, ap));
-            phiargs[0] = irbinop(fn, Oadd, KPTR, sav, gpoff);
-            /* l->gp_offset += num_gp * 8 */
-            gpoff = irbinop(fn, Oadd, KI32, gpoff, mkref(RICON, ni * 8));
-            addinstr(fn, mkinstr(Ostore32, 0, ap, gpoff));
-            assert(merge->npred == 1);
-            blkpred(merge, 0) = blk->s1;
-            blk->s1->jmp.t = Jb;
-            blk->s1->s1 = merge;
-         }
-         useblk(fn, blk->s2);
-         {
-            /* phi1: l->overflow_arg_area */
-            union ref adr = irbinop(fn, Oadd, KPTR, ap, mkref(RICON, 8));
-            union ref ovf = addinstr(fn, mkinstr(Oloadi64, KPTR, adr));
-            /* align no-op */
-            
-            phiargs[1] = ovf;
-            /* update l->overflow_arg_area += num_gp*8 */
-            int siz = 8;
-            addinstr(fn, mkinstr(Ostore64, 0, adr, irbinop(fn, Oadd, KPTR, ovf, mkref(RICON, siz))));
-            putbranch(fn, merge);
-         }
-         assert(merge->npred == 2);
-         vpush(&merge->ins, 0);
-         memmove(merge->ins.p+1, merge->ins.p, (merge->ins.n-1)*sizeof *merge->ins.p);
-         merge->ins.p[0] = var;
-         phi = insertphi(merge, KPTR);
-         memcpy(phitab.p[instrtab[phi.i].l.i], phiargs, sizeof phiargs);
-         instrtab[var] = mkinstr(cls[0] == KI32 ? Oloads32 : Oloadi64, cls[0], phi);
-      } else {
-         assert(0&&nf);
+      /* int: l->gp_offset < 48 - num_gp * 8 */
+      /* sse: l->fp_offset < 304 - num_gp * 16 (why 304? ... 176) */
+      tmp = ni ? ap : insertinstr(blk, (*curi)++, mkinstr(Oadd, KPTR, ap, mkref(RICON, 4)));
+      tmp = insertinstr(blk, (*curi)++, mkinstr(Oloadu32, KI32, tmp));
+      tmp = insertinstr(blk, (*curi)++, mkinstr(Oulte, KI32, tmp, mkref(RICON, ni ? 48 - ni*8 : 176 - nf*16)));
+      merge = blksplitafter(fn, blk, *curi);
+      blk->jmp.t = 0;
+      useblk(fn, blk);
+      putcondbranch(fn, tmp, newblk(fn), newblk(fn));
+      useblk(fn, blk->s1);
+      {
+         /* phi0: &l->reg_save_area[l->gp/fp_offset] */
+         union ref sav = addinstr(fn, mkinstr(Oloadi64, KPTR, irbinop(fn, Oadd, KPTR, ap, mkref(RICON, 16))));
+         union ref roff = addinstr(fn, mkinstr(Oloadu32, KI32, irbinop(fn, Oadd, KPTR, ap, mkref(RICON, ni ? 0 : 4))));
+         phiargs[0] = irbinop(fn, Oadd, KPTR, sav, roff);
+         /* l->gp/fp_offset += num_gp/fp * 8(16) */
+         roff = irbinop(fn, Oadd, KI32, roff, mkref(RICON, ni ? ni * 8 : nf * 16));
+         addinstr(fn, mkinstr(Ostore32, 0, irbinop(fn, Oadd, KPTR, ap, mkref(RICON, ni ? 0 : 4)), roff));
+         assert(merge->npred == 1);
+         blkpred(merge, 0) = blk->s1;
+         blk->s1->jmp.t = Jb;
+         blk->s1->s1 = merge;
+      }
+      useblk(fn, blk->s2);
+      {
+         /* phi1: l->overflow_arg_area */
+         union ref adr = irbinop(fn, Oadd, KPTR, ap, mkref(RICON, 8));
+         union ref ovf = addinstr(fn, mkinstr(Oloadi64, KPTR, adr));
+         /* align no-op */
+         
+         phiargs[1] = ovf;
+         /* update l->overflow_arg_area += size */
+         int siz = 8;
+         addinstr(fn, mkinstr(Ostore64, 0, adr, irbinop(fn, Oadd, KPTR, ovf, mkref(RICON, siz))));
+         putbranch(fn, merge);
       }
+      assert(merge->npred == 2);
+      vpush(&merge->ins, 0);
+      memmove(merge->ins.p+1, merge->ins.p, (merge->ins.n-1)*sizeof *merge->ins.p);
+      merge->ins.p[0] = var;
+      phi = insertphi(merge, KPTR);
+      memcpy(phitab.p[instrtab[phi.i].l.i], phiargs, sizeof phiargs);
+      instrtab[var] = mkinstr(cls[0] == KI32 ? Oloads32 : Oloadi64, cls[0], phi);
+      fn->prop &= ~FNUSE;
    } else {
       assert(!"nyi");
    }
diff --git a/c/c.c b/c/c.c
index 05308e1..2b2688d 100644
--- a/c/c.c
+++ b/c/c.c
@@ -2677,6 +2677,9 @@ expraddr(struct function *fn, const struct expr *ex)
    case ECALL:
       assert(isagg(ex->ty));
       return compilecall(fn, ex);
+   case EVAARG:
+      assert(isagg(ex->ty));
+      return builtin_va_arg_comp(fn, ex, 0);
    case EINIT:
       if (fn) {
          /* compound literal, allocate temp */
diff --git a/ir/abi0.c b/ir/abi0.c
index fe1a7e3..4462be9 100644
--- a/ir/abi0.c
+++ b/ir/abi0.c
@@ -437,7 +437,9 @@ abi0(struct function *fn)
       blk->id = id++;
    } while ((blk = blk->lnext) != fn->entry);
 
-   /* vaargs might break */
+
+   /* vaargs might break these */
+   if (!(fn->prop & FNUSE)) filluses(fn);
    fn->prop &= ~(FNBLKID | FNRPO);
 
    if (ccopt.dbg.a) {
diff --git a/ir/ir.c b/ir/ir.c
index 3b37405..412d0f9 100644
--- a/ir/ir.c
+++ b/ir/ir.c
@@ -326,7 +326,7 @@ blksplitafter(struct function *fn, struct block *blk, int idx)
    memset(blk->jmp.arg, 0, sizeof blk->jmp.arg);
    for (int i = 0; i < 2; ++i) {
       struct block *s = (&blk->s1)[i];
-      for (int i = 0; i < s->npred; ++i) {
+      if (s) for (int i = 0; i < s->npred; ++i) {
          if (blkpred(s, i) == blk)
             blkpred(s, i) = new;
       }
@@ -334,6 +334,7 @@ blksplitafter(struct function *fn, struct block *blk, int idx)
    new->s1 = blk->s1, new->s2 = blk->s2;
    blk->s1 = new, blk->s2 = NULL;
    addpred(new, blk);
+   fn->prop &= ~FNUSE;
    return new;
 }
 
diff --git a/ir/regalloc.c b/ir/regalloc.c
index d3c1ee0..e060ebf 100644
--- a/ir/regalloc.c
+++ b/ir/regalloc.c
@@ -517,11 +517,18 @@ usereg(struct rega *ra, int reg, struct block *blk, int pos)
 {
    struct fixinterval *fxit;
    if (rstest(mctarg->rglob, reg)) return; /* regalloc never allocates globally live regs, so don't need intervals for those */
-   for (struct fixinterval *fxit = ra->intervals.fixed; fxit; fxit = fxit->next) {
+   for (struct fixinterval *prev = NULL, *fxit = ra->intervals.fixed; fxit; prev = fxit, fxit = fxit->next) {
       if (fxit->range.from > pos) break;
       if (fxit->rs == 1<<reg && fxit->range.from <= pos && pos < fxit->range.to) {
          /* contained by existing interval */
          fxit->range.from = blk->inumstart;
+         /* insert at head */
+         //DBG(">>>extend REG %s range %d-%d\n", mctarg->rnames[reg], fxit->range.from, fxit->range.to);
+         if (prev) {
+            prev->next = fxit->next;
+            fxit->next = ra->intervals.fixed;
+            ra->intervals.fixed = fxit;
+         }
          return;
       }
    }
@@ -539,7 +546,7 @@ defreg(struct rega *ra, int reg, int pos) {
       if (fxit->rs == 1<<reg) {
          assert(fxit->range.from <= pos);
          fxit->range.from = pos;
-         // DBG(">>>REG %s range @%d: %d-%d\n", mctarg->rnames[reg], fxit->range.from.blk, fxit->range.from.ins, fxit->range.to.ins);
+         //DBG(">>>def REG %s range %d-%d\n", mctarg->rnames[reg], fxit->range.from, fxit->range.to);
          return;
       }
    }
@@ -643,9 +650,9 @@ buildintervals(struct rega *ra)
                ra->intervals.fixed = fxit;
             }
             for (int j = call->narg - 1; j >= 0; --j) {
-               int reg = call->abiarg[j].reg;
-               if (reg >= 0) {
-                  usereg(ra, reg, blk, pos);
+               struct abiarg abi = call->abiarg[j];
+               if (!abi.isstk) {
+                  usereg(ra, abi.reg, blk, pos);
                }
             }
          }
diff --git a/test/10-varargs.c b/test/10-varargs.c
index c15b887..a3715ab 100644
--- a/test/10-varargs.c
+++ b/test/10-varargs.c
@@ -1,5 +1,7 @@
 /* EXPECT:
 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8 = 36
+<1.5>
+1.1 + 2.1 + 3.1 + 5.1 + 1.5 + -1.5 + 1.5 + -1.5 + 1.5 + -1 = 11.9
 */
 
 
@@ -18,6 +20,26 @@ int sum(int x, ...) {
    return x;
 }
 
+double stkarg(double a, double b, double c, double d, double e, double f, double g, double h, double x) {
+   printf("");
+   return x;
+}
+
+double sumf(double x, ...) {
+   va_list ap;
+   va_start(ap, x);
+   printf("%g", x);
+   for (double y; (y = va_arg(ap, double));) {
+      printf(" + %g",y);
+      x += y;
+   }
+   va_end(ap);
+   return x;
+}
+
+
 int main() {
    printf(" = %d\n", sum(1,2,3,4,5,6,7,8,0,0));
+   printf("<%g>\n", stkarg(0,0,0,0,0,0,0,0,1.5));
+   printf(" = %g\n", sumf(1.1, 2.1, 3.1, 5.1, 1.5, -1.5, 1.5, -1.5, 1.5, -1.0, 0.0));
 }