ART: Special-case cb(n)z even for in-reg constant
Call out to OpCmpImmBranch in GenCompareAndBranch if the constant
is zero and we are testing == or !=, even when zero has been loaded
to a register already. This avoids a register size mismatch on 64b
architectures when basically doing a null check, and generally
squashes a cmp + branch to a cbz or cbnz on Arm and Mips. X86 is
not degraded.
Bug: 16562601
Change-Id: I1997760f43dc186a84247ad30ae91053f71d102d
diff --git a/compiler/dex/quick/gen_common.cc b/compiler/dex/quick/gen_common.cc
index adc228c..b94e816 100644
--- a/compiler/dex/quick/gen_common.cc
+++ b/compiler/dex/quick/gen_common.cc
@@ -273,13 +273,25 @@
if (rl_src2.is_const) {
// If it's already live in a register or not easily materialized, just keep going
RegLocation rl_temp = UpdateLoc(rl_src2);
+ int32_t constant_value = mir_graph_->ConstantValue(rl_src2);
if ((rl_temp.location == kLocDalvikFrame) &&
- InexpensiveConstantInt(mir_graph_->ConstantValue(rl_src2))) {
+ InexpensiveConstantInt(constant_value)) {
// OK - convert this to a compare immediate and branch
OpCmpImmBranch(cond, rl_src1.reg, mir_graph_->ConstantValue(rl_src2), taken);
return;
}
+
+ // It's also commonly more efficient to have a test against zero with Eq/Ne. This is not worse
+ // for x86, and allows a cbz/cbnz for Arm and Mips. At the same time, it works around a register
+ // mismatch for 64b systems, where a reference is compared against null, as dex bytecode uses
+ // the 32b literal 0 for null.
+ if (constant_value == 0 && (cond == kCondEq || cond == kCondNe)) {
+ // Use the OpCmpImmBranch and ignore the value in the register.
+ OpCmpImmBranch(cond, rl_src1.reg, 0, taken);
+ return;
+ }
}
+
rl_src2 = LoadValue(rl_src2);
OpCmpBranch(cond, rl_src1.reg, rl_src2.reg, taken);
}
