x86_64: for vararg calls, write to EAX in isel

Also, in regalloc ensure fixed intervals are sorted

3 files changed, 30 insertions, 22 deletions

diff --git a/ir/regalloc.c b/ir/regalloc.c
index 794cdd6..7056dfd 100644
--- a/ir/regalloc.c
+++ b/ir/regalloc.c
@@ -560,10 +560,19 @@ usereg(struct rega *ra, int reg, struct block *blk, int pos)
 static bool
 defreg(struct rega *ra, int reg, int pos) {
    if (rstest(mctarg->rglob, reg)) return 1;
-   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->rs == 1<<reg) {
          if (fxit->range.from <= pos) {
             fxit->range.from = pos;
+            struct fixinterval **at = &ra->intervals.fixed;
+            if ((*at)->range.from > pos) {
+               /* keep sorted */
+               //DBG("moved %s\n", mctarg->rnames[reg]);
+               if (prev) prev->next = fxit->next;
+               while ((*at)->range.from < pos) at = &(*at)->next;
+               fxit->next = *at;
+               *at = fxit;
+            }
             //DBG(">>>def REG %s range %d-%d\n", mctarg->rnames[reg], fxit->range.from, fxit->range.to);
             return 1;
          }
@@ -649,14 +658,21 @@ buildintervals(struct rega *ra)
          /* gather fixed intervals */
          if (ins->op == Omove) {
             assert(ins->l.t == RREG);
-            if (ins->l.bits == ins->r.bits) /* special case `move Rx,Rx`: clobber reg, not a real use */
+            if (ins->l.bits == ins->r.bits) {/* special case `move Rx,Rx`: clobber reg, not a real use */
                usereg(ra, ins->l.i, blk, pos);
+               //DBG("@ %d clob %s\n", pos, mctarg->rnames[ins->l.i]);
+            }
             if (!defreg(ra, ins->l.i, pos)) {
-               /* dead register use. for example if
-                *   move RCX, %1
-                *   %2 = shl 1, RCX
-                * and %2 is dead, the move to RCX can be killed */
-               *ins = mkinstr(Onop,0,);
+               if (ins->keep) { /* clobber here */
+                  usereg(ra, ins->l.i, blk, pos);
+                  assert(defreg(ra, ins->l.i, pos));
+               } else {
+                  /* dead register use. for example if
+                   *   move RCX, %1
+                   *   %2 = shl 1, RCX
+                   * and %2 is dead, the move to RCX can be killed */
+                  *ins = mkinstr(Onop,0,);
+               }
             }
             if (ins->l.bits == ins->r.bits)
                continue;
@@ -917,7 +933,8 @@ linearscan(struct rega *ra)
          int end = itrange(current, current->nrange-1).to;
 
          /* exclude regs from overlapping fixed intervals */
-         for (struct fixinterval *fxit = intervals->fixed; fxit; fxit = fxit->next) {
+         for (struct fixinterval *last = NULL, *fxit = intervals->fixed; fxit; last = fxit, fxit = fxit->next) {
+            if (last) assert(last->range.from <= fxit->range.from && "unsorted fixintervals");
             if (fxit->range.to <= pos) {
                intervals->fixed = fxit->next;
                continue;
diff --git a/x86_64/emit.c b/x86_64/emit.c
index 13121b3..89585b8 100644
--- a/x86_64/emit.c
+++ b/x86_64/emit.c
@@ -1171,16 +1171,6 @@ emitinstr(uchar **pcode, struct function *fn, struct block *blk, int curi, struc
       Xpop(pcode, mkregoper(ins->l).reg);
       break;
    case Ocall:
-      if (calltab.p[ins->r.i].vararg >= 0) {
-         struct call *call = &calltab.p[ins->r.i];
-         /* variadic functions need the caller to write num of args in sse regs to %al */
-         int n = 0;
-         for (int i = 0; i < call->narg; ++i)
-            if (!call->abiarg[i].isstk && call->abiarg[i].reg >= XMM0)
-               ++n;
-         if (!n) DS("\x31\xC0"); /* XOR EAX, EAX */
-         else B(0xB0), B(n); /* MOV AL, n */
-      }
       Xcall(pcode, KPTR, ref2oper(ins->l));
       break;
    case Oxvaprologue:
diff --git a/x86_64/isel.c b/x86_64/isel.c
index 67a4358..22a1290 100644
--- a/x86_64/isel.c
+++ b/x86_64/isel.c
@@ -126,6 +126,7 @@ selcall(struct function *fn, struct instr *ins, struct block *blk, int *curi)
    int iarg = *curi - 1;
    enum irclass cls;
    uint argstksiz = alignup(call->argstksiz, 16);
+   int nsse = 0;
 
    for (int i = call->narg - 1; i >= 0; --i) {
       struct abiarg abi = call->abiarg[i];
@@ -139,6 +140,7 @@ selcall(struct function *fn, struct instr *ins, struct block *blk, int *curi)
       if (!abi.isstk) {
          assert(!abi.ty.isagg);
          *arg = mkinstr(Omove, call->abiarg[i].ty.cls, mkref(RREG, abi.reg), arg->r);
+         if (abi.reg >= XMM0) ++nsse;
       } else {
          union ref adr = mkaddr((struct addr){mkref(RREG, RSP), .disp = abi.stk});
          int iargsave = iarg;
@@ -165,10 +167,9 @@ selcall(struct function *fn, struct instr *ins, struct block *blk, int *curi)
    else if (isintcon(ins->l))
       ins->l = insertinstr(blk, (*curi)++, mkinstr(Ocopy, KPTR, ins->l));
 
-   if (call->vararg >= 0 && ins->l.t == RTMP) {
-      /* variadic calls write number of sse regs used to AL, so mark it as clobbered such that
-       * the function pointer of an indirect calls does not get allocated to RAX by regalloc */
-      insertinstr(blk, (*curi)++, mkinstr(Omove, KPTR, mkref(RREG, RAX), mkref(RREG, RAX)));
+   if (call->vararg >= 0) {
+      /* variadic calls write number of sse regs used to AL */
+      insertinstr(blk, (*curi)++, mkinstr(Omove, KI32, mkref(RREG, RAX), mkref(RICON, nsse), .keep=1));
    }
    cls = ins->cls;
    ins->cls = 0;