Generalize optimization of comparison to emtpy bitstring

bjorng · Nov 26, 2023 · 66997a2 · 66997a2
1 parent 25d1268
commit 66997a2
Showing 1 changed file with 30 additions and 5 deletions.
diff --git a/erts/emulator/beam/jit/arm/instr_common.cpp b/erts/emulator/beam/jit/arm/instr_common.cpp
@@ -1555,15 +1555,40 @@ void BeamModuleAssembler::emit_is_eq_exact(const ArgLabel &Fail,
                                            const ArgSource &Y) {
     auto x = load_source(X, ARG1);
 
-    if (exact_type<BeamTypeId::Bitstring>(X) && Y.isLiteral()) {
+    if (Y.isLiteral()) {
         Eterm literal = beamfile_get_literal(beam, Y.as<ArgLiteral>().get());
 
         if (is_bitstring(literal) && bitstring_size(literal) == 0) {
-            comment("simplified equality test with empty binary");
+            arm::Gp boxed_ptr;
 
-            arm::Gp boxed_ptr = emit_ptr_val(ARG1, x.reg);
-            a.ldur(TMP1, emit_boxed_val(boxed_ptr, sizeof(Eterm)));
-            a.cbnz(TMP1, resolve_beam_label(Fail, disp1MB));
+            comment("simplified equality test with empty binary");
+            emit_is_boxed(resolve_beam_label(Fail, dispUnknown), X, x.reg);
+            boxed_ptr = emit_ptr_val(ARG1, x.reg);
+            if (masked_types<BeamTypeId::MaybeBoxed>(X) ==
+                BeamTypeId::Bitstring) {
+                comment("eliminated check for known bitstring");
+                a.ldur(TMP1, emit_boxed_val(boxed_ptr, sizeof(Eterm)));
+                a.cbnz(TMP1, resolve_beam_label(Fail, disp1MB));
+            } else {
+                a.sub(TMP1, boxed_ptr, imm(TAG_PRIMARY_BOXED));
+                a.ldp(TMP1, TMP2, arm::Mem(TMP1));
+                /* The header mask with the binary sub tag bits
+                 * removed (0b110011) is not possible to use as an
+                 * immediate operand for 'and'. (See the note at the
+                 * beginning of the file.) Therefore, use a simpler
+                 * mask (0b110000) that will also clear the primary
+                 * tag bits. That works because we KNOW that a boxed
+                 * pointer always points to a header word and that the
+                 * primary tag for a header is 0. */
+                const auto mask = _BITSTRING_TAG_MASK & ~_TAG_PRIMARY_MASK;
+                ERTS_CT_ASSERT(TAG_PRIMARY_HEADER == 0);
+                ERTS_CT_ASSERT(_TAG_HEADER_HEAP_BITS ==
+                               (_TAG_HEADER_HEAP_BITS & mask));
+                a.and_(TMP1, TMP1, imm(mask));
+                a.cmp(TMP1, imm(_TAG_HEADER_HEAP_BITS));
+                a.ccmp(TMP2, imm(0), imm(NZCV::kNone), imm(arm::CondCode::kEQ));
+                a.b_ne(resolve_beam_label(Fail, disp1MB));
+            }
             return;
         }
     }