improve codegen for div a little (bikeshedding)

2 files changed, 7 insertions, 1 deletions

diff --git a/amd64/isel.c b/amd64/isel.c
index 2273c99..b0cacea 100644
--- a/amd64/isel.c
+++ b/amd64/isel.c
@@ -224,6 +224,7 @@ static void
 sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
 {
    uint siz, alignlog2;
+   int t;
    struct instr temp = {0};
    enum op op = ins->op;
 
@@ -290,6 +291,10 @@ sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
       *ins = mkinstr(Ocopy, ins->cls, mkref(RREG, op < Orem ? RAX : RDX)); /* get output */
       temp = mkinstr(Ocopy, ins->cls, mkref(RREG, op < Orem ? RDX : RAX)); /* clobber other reg*/
       insertinstr(blk, ++(*curi), temp);
+      /* swap instrs so that clobber goes first */
+      t = blk->ins.p[*curi - 1];
+      blk->ins.p[*curi - 1] = blk->ins.p[*curi - 0];
+      blk->ins.p[*curi - 0] = t;
       break;
    case Osub:
        if (ins->r.bits == mkref(RICON, 1).bits) {
diff --git a/regalloc.c b/regalloc.c
index 8f31b05..a30ebe5 100644
--- a/regalloc.c
+++ b/regalloc.c
@@ -238,7 +238,8 @@ use(struct rega *ra, struct block *blk, int curi, enum op op, int hint, union re
 
       assert(ins->op != Ocall);
       if (ins->r.t == RREG && ins->inplace) excl |= 1ull<<ins->r.i;
-      if (hint == -1 && ins->op == Ocopy && ins->l.t == RREG) /* for '%x = copy Rx', hint %x to use Rx */
+      if ((hint == -1 || ra->regs[hint].t) && ins->op == Ocopy && ins->l.t == RREG)
+         /* for '%x = copy Rx', hint %x to use Rx */
          hint = ins->l.i;
       if (hint != -1 && !(excl >> hint & 1) && !ra->regs[hint].t) {
          take(ra, hint, *ref);