aarch64: rem/urem, bswap*, arg bugfix

5 files changed, 67 insertions, 12 deletions

diff --git a/src/ir_op.def b/src/ir_op.def
index 4e775a1..dde4e10 100644
--- a/src/ir_op.def
+++ b/src/ir_op.def
@@ -75,5 +75,6 @@ _(phi,       1)
 _(swap,      2)
 _(vastart,   1)
 _(vaarg,     2)
-/* machine-specific instructions */
+/* machine-specific/internal instructions */
 _(xvaprologue, 1)
+_(msub,    3)
diff --git a/src/ir_regalloc.c b/src/ir_regalloc.c
index 41a4026..3b204c6 100644
--- a/src/ir_regalloc.c
+++ b/src/ir_regalloc.c
@@ -725,9 +725,9 @@ buildintervals(RegAlloc *ra)
             queue[0] = ins->r;
          } else {
             switch ((nqueue = opnarg[ins->op])) {
-            case 2: queue[2] = ins->oper[2];
-            case 1: queue[1] = ins->oper[1];
-            case 0: queue[0] = ins->oper[0];
+            case 3: queue[2] = ins->oper[2];
+            case 2: queue[1] = ins->oper[1];
+            case 1: queue[0] = ins->oper[0];
             }
          }
          if (0) {
@@ -1148,15 +1148,15 @@ devirt(RegAlloc *ra, Block *blk)
       Interval *it;
       Alloc *alloc;
       IRAddr newaddr;
-      Ref *argref[4];
+      Ref *argref[6];
       int curi0;
       int naddr = 0;
       int nargref = 0;
       int nspill = 0;
 
       /** devirtualize operands **/
-      for (int i = 0; i < 2; ++i) {
-         Ref *r = &i[&ins->l];
+      for (int i = 0; i < 3; ++i) {
+         Ref *r = &ins->oper[i];
          if (r->t == RADDR) {
             IRAddr *a = &addrtab.p[r->i];
             ++naddr;
diff --git a/src/t_aarch64_emit.c b/src/t_aarch64_emit.c
index 4679569..abf1f3e 100644
--- a/src/t_aarch64_emit.c
+++ b/src/t_aarch64_emit.c
@@ -227,6 +227,9 @@ encode(uchar **pcode, const EncDesc *tab, int ntab, enum irclass k, Oper o[3])
    case EN_ADDSUBSHFT3R: case EN_LOGSHFT3R:
       ins |= sf<<31 | o[2].shft<<22 | o[2].reg<<16 | o[2].shamt<<10 | o[1].reg<<5 | o[0].reg;
       break;
+   case EN_ARITH2R:
+      ins |= sf<<31 | o[1].reg<<5 | o[0].reg;
+      break;
    case EN_ARITH3R:
       ins |= sf<<31 | o[2].reg<<16 | o[1].reg<<5 | o[0].reg;
       break;
@@ -352,6 +355,15 @@ Xmadd(uchar **pcode, enum irclass k, Oper d, Oper n, Oper m, Oper a)
    W32(0x1B000000 | sf<<31 | m.reg<<16 | a.reg<<10 | n.reg<<5 | d.reg);
 }
 
+static void
+Xmsub(uchar **pcode, enum irclass k, Oper d, Oper n, Oper m, Oper a)
+{
+   assert(opermatch(PGPRZ, k, d) && opermatch(PGPRZ, k, n)
+       && opermatch(PGPRZ, k, a) && opermatch(PGPRZ, k, m));
+   uint sf = k > KI32;
+   W32(0x1B008000 | sf<<31 | m.reg<<16 | a.reg<<10 | n.reg<<5 | d.reg);
+}
+
 DEFINSTR3(Xsdiv, {4|8, {PGPRZ, PGPRZ, PGPRZ}, 0x1AC00C00, EN_ARITH3R})
 DEFINSTR3(Xudiv, {4|8, {PGPRZ, PGPRZ, PGPRZ}, 0x1AC00800, EN_ARITH3R})
 
@@ -515,6 +527,10 @@ DEFINSTR2(Xfcmp,
    {4|8, {PFPR, PZERO}, 0x1E602008, EN_FPCMPZ},
    {4|8, {PFPR,  PFPR}, 0x1E602000, EN_FPCMP},
 )
+DEFINSTR2(Xrev,
+   {4  , {PGPRZ, PGPRZ}, 0x5AC00800, EN_ARITH2R},
+   {  8, {PGPRZ, PGPRZ}, 0xDAC00C00, EN_ARITH2R},
+)
 
 static void
 gencopy(uchar **pcode, enum irclass cls, Block *blk, int curi, Oper dst, Ref val)
@@ -744,12 +760,18 @@ emitinstr(uchar **pcode, Function *fn, Block *blk, int curi, Instr *ins)
    Cvt:
       X2(pcode, cls, reg2oper(ins->reg-1), ref2oper(ins->l));
       break;
+   case Obswap32: case Obswap64:
+      Xrev(pcode, cls, reg2oper(ins->reg-1), ref2oper(ins->l));
+      break;
    case Oadd: X3 = kisint(cls) ? Xadd : Xfadd; goto ALU3;
    case Osub: X3 = kisint(cls) ? Xsub : Xfsub; goto ALU3;
    case Omul: if (kisflt(cls)) { X3 = Xfmul; goto ALU3; }
       /* MUL Rd,Rn,Rm ==> MADD Rd,Rn,Rm,zr */
       Xmadd(pcode, cls, reg2oper(ins->reg-1), ref2oper(ins->l), ref2oper(ins->r), REGZR);
       break;
+   case Omsub:
+      Xmsub(pcode, cls, reg2oper(ins->reg-1), ref2oper(ins->l), ref2oper(ins->r), ref2oper(ins->oper[2]));
+      break;
    case Odiv: X3 = kisint(cls) ? Xsdiv : Xfdiv; goto ALU3;
    case Oudiv: X3 = Xudiv; goto ALU3;
    case Oand: X3 = Xand; goto ALU3;
diff --git a/src/t_aarch64_isel.c b/src/t_aarch64_isel.c
index 381a57d..29056bc 100644
--- a/src/t_aarch64_isel.c
+++ b/src/t_aarch64_isel.c
@@ -137,7 +137,15 @@ fixarg(Ref *r, Instr *ins, Block *blk, int *curi)
       else
          *r = insertinstr(blk, (*curi)++, adr);
    } else if (r->t != RTMP) Reg: {
-      regarg(r, r->t == RTMP ? instrtab[r->i].cls : ins->cls ? ins->cls : KI32, blk, curi);
+      enum irclass k;
+      if (r->t == RTMP) k = insrescls(instrtab[r->i]);
+      else if (ins->op == Oarg) {
+         IRType ty = ref2type(ins->l);
+         k = ty.isagg ? KPTR : ty.cls;
+      } else {
+         k = ins->cls;
+      }
+      regarg(r, k, blk, curi);
    }
 }
 
@@ -351,8 +359,9 @@ loadstoreaddr(Block *blk, Ref *r, int *curi, enum op op)
 static void
 sel(Function *fn, Instr *ins, Block *blk, int *curi)
 {
-   enum op op = ins->op;
+   Ref tmp;
    enum irclass cls;
+   enum op op = ins->op;
 
    if (oisarith(ins->op) && arithfold(ins)) {
       fixarg(&ins->l, ins, blk, curi);
@@ -390,6 +399,18 @@ sel(Function *fn, Instr *ins, Block *blk, int *curi)
       regarg(&ins->l, ins->cls, blk, curi);
       ins->op = Ocopy;
       break;
+   case Obswap32: case Obswap64:
+      regarg(&ins->l, ins->cls, blk, curi);
+      break;
+   case Obswap16:
+      /* %tmp = rev %x
+       * %res = lsr %tmp, 16 */
+      regarg(&ins->l, ins->cls, blk, curi);
+      tmp = insertinstr(blk, (*curi)++, mkinstr1(Obswap32, KI32, ins->l));
+      ins->op = Oslr;
+      ins->l = tmp;
+      ins->r = mkref(RICON, 16);
+      break;
    case Oadd:
       if (isnumcon(ins->l)) {
          /* swap to have const in rhs */
@@ -409,7 +430,7 @@ sel(Function *fn, Instr *ins, Block *blk, int *curi)
    case Oand: case Oior: case Oxor:
       if (isnumcon(ins->l)) {
          /* swap to have const in rhs */
-         Ref tmp = ins->l;
+         tmp = ins->l;
          ins->l = ins->r;
          ins->r = tmp;
       }
@@ -421,9 +442,20 @@ sel(Function *fn, Instr *ins, Block *blk, int *curi)
       regarg(&ins->l, ins->cls, blk, curi);
       fixarg(&ins->r, ins, blk, curi);
       break;
-   case Omul: case Odiv: case Oudiv: case Ourem:
+   case Omul: case Odiv: case Oudiv:
+      regarg(&ins->l, ins->cls, blk, curi);
+      regarg(&ins->r, ins->cls, blk, curi);
+      break;
+   case Orem: case Ourem:
       regarg(&ins->l, ins->cls, blk, curi);
       regarg(&ins->r, ins->cls, blk, curi);
+      /* %tmp = div %l, %r
+       * %res = msub %tmp, %r, %l */
+      tmp = insertinstr(blk, (*curi)++,
+            mkinstr2(op == Orem ? Odiv : Oudiv, ins->cls, ins->l, ins->r));
+      ins->op = Omsub;
+      ins->oper[2] = ins->l;
+      ins->oper[0] = tmp;
       break;
    case Oarg:
       fixarg(&ins->r, ins, blk, curi);
diff --git a/test/run.sh b/test/run.sh
index ec40161..07f967e 100755
--- a/test/run.sh
+++ b/test/run.sh
@@ -26,7 +26,7 @@ run() {
             echo '-------'
         else
             actual=build/"$(echo "$f" | sed 's/\.c$/.actual/')"
-            x "$exe" $args > "$actual"
+            x $QEMU "$exe" $args > "$actual"
             if ! cmp "$actual"  "$expected" > /dev/null; then
                 echo --- !TEST ERROR "$f"
                 diff --unified=0 "$expected" "$actual"