fixup! Optimize division by a power of two

bjorng · Jul 30, 2023 · 6eb95d5 · 6eb95d5
1 parent 64a45b3
commit 6eb95d5
Showing 1 changed file with 9 additions and 4 deletions.
diff --git a/erts/emulator/beam/jit/arm/instr_arith.cpp b/erts/emulator/beam/jit/arm/instr_arith.cpp
@@ -875,9 +875,14 @@ void BeamModuleAssembler::emit_div_rem_literal(Sint divisor,
             if (need_div) {
                 comment("optimized div by replacing with right shift");
             }
-            add(TMP1, dividend, (divisor - 1) << _TAG_IMMED1_SIZE);
-            a.cmp(dividend, imm(0));
-            a.csel(TMP3, TMP1, dividend, imm(arm::CondCode::kLT));
+            if (divisor == 2) {
+                ERTS_CT_ASSERT(_TAG_IMMED1_SMALL == _TAG_IMMED1_MASK);
+                a.add(TMP3, dividend, dividend, arm::lsr(63));
+            } else {
+                add(TMP1, dividend, (divisor - 1) << _TAG_IMMED1_SIZE);
+                a.cmp(dividend, imm(0));
+                a.csel(TMP3, TMP1, dividend, imm(arm::CondCode::kLT));
+            }
             if (need_div && need_rem && quotient == dividend) {
                 a.mov(TMP4, dividend);
                 dividend_reg = TMP4;
@@ -891,7 +896,7 @@ void BeamModuleAssembler::emit_div_rem_literal(Sint divisor,
             }
             if (need_rem) {
                 Uint mask = (Uint)-1 << (shift + _TAG_IMMED1_SIZE);
-                comment("optimized rem by replacing with masking");
+                comment("optimized rem by replacing with subtraction");
                 a.and_(TMP1, TMP3, imm(mask));
                 a.sub(remainder, dividend_reg, TMP1);
             }