dec,inc,sub

5 files changed, 57 insertions, 3 deletions

diff --git a/amd64/emit.c b/amd64/emit.c
index a89b021..e1d9463 100644
--- a/amd64/emit.c
+++ b/amd64/emit.c
@@ -280,11 +280,24 @@ DEFINSTR2(Xadd,
    {4, PFPR, PMEM, "\xF3\x0F\x58", EN_RM}, /* ADDSS xmm, m32 */
    {8, PFPR, PMEM, "\xF2\x0F\x58", EN_RM}, /* ADDSD xmm, m64 */
 )
+DEFINSTR2(Xsub,
+   {4|8, PGPR, PGPR, "\x2B", EN_RR},           /* SUB r32/64, r32/64 */
+   {4|8, PGPR, PI8,  "\x83", EN_RI8, .ext=5},  /* SUB r32/64, imm8 */
+   {4|8, PRAX, PI32, "\x2D", EN_I32},          /* SUB eax/rax, imm */
+   {4|8, PGPR, PI32, "\x81", EN_RI32, .ext=5}, /* SUB r32/64, imm */
+   {  8, PGPR, PMEM, "\x2B", EN_RM},           /* SUB r64, m64 */
+   {4, PFPR, PFPR, "\xF3\x0F\x5C", EN_RR}, /* SUBSS xmm, xmm */
+   {8, PFPR, PFPR, "\xF2\x0F\x5C", EN_RR}, /* SUBSD xmm, xmm */
+   {4, PFPR, PMEM, "\xF3\x0F\x5C", EN_RM}, /* SUBSS xmm, m32 */
+   {8, PFPR, PMEM, "\xF2\x0F\x5C", EN_RM}, /* SUBSD xmm, m64 */
+)
 DEFINSTR2(Xshl,
    {4|8, PGPR, P1,   "\xD1", EN_R, .ext=4},   /* SHL r32/64, 1 */
    {4|8, PGPR, PI32, "\xC1", EN_RI8, .ext=4}, /* SHL r32/64, imm */
    {4|8, PGPR, PRCX, "\xD3", EN_R, .ext=4},   /* SHL r32/64, CL */
 )
+DEFINSTR1(Xinc, {4|8, PGPR, 0, "\xFF", EN_R, .ext=0} /* INC r32/64 */)
+DEFINSTR1(Xdec, {4|8, PGPR, 0, "\xFF", EN_R, .ext=1} /* DEC r32/64 */)
 DEFINSTR1(Xidiv,
    {4|8, PGPR, 0, "\xF7", EN_R, .ext=7}, /* IDIV r32/64 */
    {4|8, PMEM, 0, "\xF7", EN_M, .ext=7}, /* IDIV m32/64 */
@@ -433,6 +446,7 @@ emitinstr(uchar **pcode, uint *stktop, struct function *fn, struct block *blk, i
    struct oper dst, src;
    uchar ksiz = cls2siz[ins->cls];
    void (*X)(uchar **, uint, struct oper, struct oper) = NULL;
+   void (*X1)(uchar **, uint, struct oper) = NULL;
 
    if (oisalloca(ins->op)) {
       uint alignlog2 = ins->op - Oalloca1;
@@ -473,12 +487,20 @@ emitinstr(uchar **pcode, uint *stktop, struct function *fn, struct block *blk, i
          Xlea(pcode, ksiz, dst, mem);
       }
       break;
+   case Osub: X = Xsub; goto ALU2;
    case Oshl: X = Xshl; goto ALU2;
    ALU2:
       dst = mkregoper(ins->l);
       assert(ins->reg-1 == dst.reg);
       X(pcode, ksiz, dst, mkimmdatregoper(ins->r));
       break;
+   case Oxinc: X1 = Xinc; goto ALU1;
+   case Oxdec: X1 = Xdec; goto ALU1;
+   ALU1:
+      dst = mkregoper(ins->l);
+      assert(ins->reg-1 == dst.reg);
+      X1(pcode, ksiz, dst);
+      break;
    case Odiv: case Orem:
       switch (ins->cls) {
       case KI8: B(0x48); /* REX.W */
diff --git a/amd64/isel.c b/amd64/isel.c
index 8ab11a7..30ad222 100644
--- a/amd64/isel.c
+++ b/amd64/isel.c
@@ -171,7 +171,11 @@ sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
       insertinstr(blk, ++(*curi), temp);
       break;
    case Osub:
-      if (iscon(ins->l)) {
+       if (ins->r.bits == mkref(RICON, 1).bits) {
+         /* sub x,1 -> dec x */
+         ins->op = Oxdec;
+         ins->r = NOREF;
+      } else if (iscon(ins->l)) {
          /* sub imm, x -> sub x, imm; neg x */
          struct instr sub = *ins;
          rswap(sub.l, sub.r);
@@ -181,6 +185,18 @@ sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
       }
       goto ALU;
    case Oadd:
+      if (ins->l.bits == mkref(RICON, 1).bits) {
+         /* add 1,x -> inc x */
+         ins->op = Oxinc;
+         ins->l = ins->r;
+         ins->r = NOREF;
+         goto ALU;
+      } else if (ins->r.bits == mkref(RICON, 1).bits) {
+         /* add x,1 -> inc x */
+         ins->op = Oxinc;
+         ins->r = NOREF;
+         goto ALU;
+      }
       if (kisint(ins->cls) && (addarg4addrp(ins->l) || addarg4addrp(ins->r))) {
          temp.op = Ocopy;
          temp.cls = ins->cls;
diff --git a/op.def b/op.def
index 8bdfbbc..42b7c3c 100644
--- a/op.def
+++ b/op.def
@@ -67,3 +67,6 @@ _(arg,      2)
 _(call,     2)
 _(intrin,   2)
 _(phi,      1)
+/* machine-specific instructions */
+_(xinc,      1)
+_(xdec,      1)
diff --git a/regalloc.c b/regalloc.c
index f03c33c..f001fee 100644
--- a/regalloc.c
+++ b/regalloc.c
@@ -219,6 +219,7 @@ regalloc(struct function *fn)
                /* an in-place operation where the destination does not
                 * match the first operand, so we need to add a move */
                insertinstr(blk, i, mkmove(ins->cls, ins->reg-1, ins->l.i));
+               ins->l.i = ins->reg-1;
             }
          }
       }
diff --git a/test/test3.c b/test/test3.c
index c39d15b..8bd9020 100644
--- a/test/test3.c
+++ b/test/test3.c
@@ -12,10 +12,22 @@ int shcl(int a, int b) {
     return a << (b+1);
 }
 
-int div(int a,int b) {
-    return a < 0 ? a / b : b / a;
+struct p { long x,y; };
+struct p divsh(int a) {
+    struct p p;
+    p.x = a << (a / 5);
+    p.y = a;
+    return p;
 }
 
+struct bitset { unsigned long long u; };
+static inline void
+bscopy(struct bitset dst[/*siz*/], const struct bitset src[/*siz*/], unsigned siz)
+{
+   while (siz--) dst++->u = src++->u;
+}
+
+
 #if 0
 long test(long x) {
     return x + (long)"abc";