x86/x64: Don't fuse loads across IR_NEWREF.

Reported by Peter Cawley. #1117

1 files changed, 9 insertions, 7 deletions

diff --git a/src/lj_asm_x86.h b/src/lj_asm_x86.h
index 6b114802..ddbe9c55 100644
--- a/src/lj_asm_x86.h
+++ b/src/lj_asm_x86.h
@@ -93,7 +93,7 @@ static int asm_isk32(ASMState *as, IRRef ref, int32_t *k)
 /* Check if there's no conflicting instruction between curins and ref.
 ** Also avoid fusing loads if there are multiple references.
 */
-static int noconflict(ASMState *as, IRRef ref, IROp conflict, int noload)
+static int noconflict(ASMState *as, IRRef ref, IROp conflict, int check)
 {
   IRIns *ir = as->ir;
   IRRef i = as->curins;
@@ -102,7 +102,9 @@ static int noconflict(ASMState *as, IRRef ref, IROp conflict, int noload)
   while (--i > ref) {
     if (ir[i].o == conflict)
       return 0;  /* Conflict found. */
-    else if (!noload && (ir[i].op1 == ref || ir[i].op2 == ref))
+    else if ((check & 1) && ir[i].o == IR_NEWREF)
+      return 0;
+    else if ((check & 2) && (ir[i].op1 == ref || ir[i].op2 == ref))
       return 0;
   }
   return 1;  /* Ok, no conflict. */
@@ -118,7 +120,7 @@ static IRRef asm_fuseabase(ASMState *as, IRRef ref)
     lua_assert(irb->op2 == IRFL_TAB_ARRAY);
     /* We can avoid the FLOAD of t->array for colocated arrays. */
     if (ira->o == IR_TNEW && ira->op1 <= LJ_MAX_COLOSIZE &&
-        !neverfuse(as) && noconflict(as, irb->op1, IR_NEWREF, 1)) {
+        !neverfuse(as) && noconflict(as, irb->op1, IR_NEWREF, 0)) {
       as->mrm.ofs = (int32_t)sizeof(GCtab);  /* Ofs to colocated array. */
       return irb->op1;  /* Table obj. */
     }
@@ -337,7 +339,7 @@ static Reg asm_fuseload(ASMState *as, IRRef ref, RegSet allow)
     RegSet xallow = (allow & RSET_GPR) ? allow : RSET_GPR;
     if (ir->o == IR_SLOAD) {
       if (!(ir->op2 & (IRSLOAD_PARENT|IRSLOAD_CONVERT)) &&
-          noconflict(as, ref, IR_RETF, 0)) {
+          noconflict(as, ref, IR_RETF, 2)) {
         as->mrm.base = (uint8_t)ra_alloc1(as, REF_BASE, xallow);
         as->mrm.ofs = 8*((int32_t)ir->op1-1) + ((ir->op2&IRSLOAD_FRAME)?4:0);
         as->mrm.idx = RID_NONE;
@@ -346,12 +348,12 @@ static Reg asm_fuseload(ASMState *as, IRRef ref, RegSet allow)
     } else if (ir->o == IR_FLOAD) {
       /* Generic fusion is only ok for 32 bit operand (but see asm_comp). */
       if ((irt_isint(ir->t) || irt_isu32(ir->t) || irt_isaddr(ir->t)) &&
-          noconflict(as, ref, IR_FSTORE, 0)) {
+          noconflict(as, ref, IR_FSTORE, 2)) {
         asm_fusefref(as, ir, xallow);
         return RID_MRM;
       }
     } else if (ir->o == IR_ALOAD || ir->o == IR_HLOAD || ir->o == IR_ULOAD) {
-      if (noconflict(as, ref, ir->o + IRDELTA_L2S, 0)) {
+      if (noconflict(as, ref, ir->o + IRDELTA_L2S, 2+(ir->o != IR_ULOAD))) {
         asm_fuseahuref(as, ir->op1, xallow);
         return RID_MRM;
       }
@@ -360,7 +362,7 @@ static Reg asm_fuseload(ASMState *as, IRRef ref, RegSet allow)
       ** Fusing unaligned memory operands is ok on x86 (except for SIMD types).
       */
       if ((!irt_typerange(ir->t, IRT_I8, IRT_U16)) &&
-          noconflict(as, ref, IR_XSTORE, 0)) {
+          noconflict(as, ref, IR_XSTORE, 2)) {
         asm_fusexref(as, ir->op1, xallow);
         return RID_MRM;
       }