ART: Fix GenSelect and GenFusedLongCmpBranch for ARM64
Depending on the result, we need to issue a wide csel. Also need
to handle constants, and src and dest being the same.
In GenFusedLongCmpBranch there is an ordering issue. If we swap
the inputs, we did not Load the second one.
Change-Id: Icb9876ca1288602d078b9fb89ea964ec2c910e0c
diff --git a/compiler/dex/quick/arm64/int_arm64.cc b/compiler/dex/quick/arm64/int_arm64.cc
index e8f5cb9..3ee3e2e 100644
--- a/compiler/dex/quick/arm64/int_arm64.cc
+++ b/compiler/dex/quick/arm64/int_arm64.cc
@@ -91,17 +91,121 @@
RegLocation rl_dest = mir_graph_->GetDest(mir);
RegisterClass src_reg_class = rl_src.ref ? kRefReg : kCoreReg;
RegisterClass result_reg_class = rl_dest.ref ? kRefReg : kCoreReg;
+
rl_src = LoadValue(rl_src, src_reg_class);
+ // rl_src may be aliased with rl_result/rl_dest, so do compare early.
+ OpRegImm(kOpCmp, rl_src.reg, 0);
+
ArmConditionCode code = ArmConditionEncoding(mir->meta.ccode);
- RegLocation rl_true = mir_graph_->reg_location_[mir->ssa_rep->uses[1]];
- RegLocation rl_false = mir_graph_->reg_location_[mir->ssa_rep->uses[2]];
- rl_true = LoadValue(rl_true, result_reg_class);
- rl_false = LoadValue(rl_false, result_reg_class);
- rl_result = EvalLoc(rl_dest, result_reg_class, true);
- OpRegImm(kOpCmp, rl_src.reg, 0);
- NewLIR4(kA64Csel4rrrc, rl_result.reg.GetReg(), rl_true.reg.GetReg(),
- rl_false.reg.GetReg(), code);
+ // The kMirOpSelect has two variants, one for constants and one for moves.
+ bool is_wide = rl_dest.ref || rl_dest.wide;
+
+ if (mir->ssa_rep->num_uses == 1) {
+ uint32_t true_val = mir->dalvikInsn.vB;
+ uint32_t false_val = mir->dalvikInsn.vC;
+
+ int opcode; // The opcode.
+ int left_op, right_op; // The operands.
+ bool rl_result_evaled = false;
+
+ // Check some simple cases.
+ // TODO: Improve this.
+ int zero_reg = (is_wide ? rs_xzr : rs_wzr).GetReg();
+
+ if ((true_val == 0 && false_val == 1) || (true_val == 1 && false_val == 0)) {
+ // CSInc cheap based on wzr.
+ if (true_val == 1) {
+ // Negate.
+ code = ArmConditionEncoding(NegateComparison(mir->meta.ccode));
+ }
+
+ left_op = right_op = zero_reg;
+ opcode = is_wide ? WIDE(kA64Csinc4rrrc) : kA64Csinc4rrrc;
+ } else if ((true_val == 0 && false_val == 0xFFFFFFFF) ||
+ (true_val == 0xFFFFFFFF && false_val == 0)) {
+ // CSneg cheap based on wzr.
+ if (true_val == 0xFFFFFFFF) {
+ // Negate.
+ code = ArmConditionEncoding(NegateComparison(mir->meta.ccode));
+ }
+
+ left_op = right_op = zero_reg;
+ opcode = is_wide ? WIDE(kA64Csneg4rrrc) : kA64Csneg4rrrc;
+ } else if (true_val == 0 || false_val == 0) {
+ // Csel half cheap based on wzr.
+ rl_result = EvalLoc(rl_dest, result_reg_class, true);
+ rl_result_evaled = true;
+ if (false_val == 0) {
+ // Negate.
+ code = ArmConditionEncoding(NegateComparison(mir->meta.ccode));
+ }
+ LoadConstantNoClobber(rl_result.reg, true_val == 0 ? false_val : true_val);
+ left_op = zero_reg;
+ right_op = rl_result.reg.GetReg();
+ opcode = is_wide ? WIDE(kA64Csel4rrrc) : kA64Csel4rrrc;
+ } else if (true_val == 1 || false_val == 1) {
+ // CSInc half cheap based on wzr.
+ rl_result = EvalLoc(rl_dest, result_reg_class, true);
+ rl_result_evaled = true;
+ if (true_val == 1) {
+ // Negate.
+ code = ArmConditionEncoding(NegateComparison(mir->meta.ccode));
+ }
+ LoadConstantNoClobber(rl_result.reg, true_val == 1 ? false_val : true_val);
+ left_op = rl_result.reg.GetReg();
+ right_op = zero_reg;
+ opcode = is_wide ? WIDE(kA64Csinc4rrrc) : kA64Csinc4rrrc;
+ } else if (true_val == 0xFFFFFFFF || false_val == 0xFFFFFFFF) {
+ // CSneg half cheap based on wzr.
+ rl_result = EvalLoc(rl_dest, result_reg_class, true);
+ rl_result_evaled = true;
+ if (true_val == 0xFFFFFFFF) {
+ // Negate.
+ code = ArmConditionEncoding(NegateComparison(mir->meta.ccode));
+ }
+ LoadConstantNoClobber(rl_result.reg, true_val == 0xFFFFFFFF ? false_val : true_val);
+ left_op = rl_result.reg.GetReg();
+ right_op = zero_reg;
+ opcode = is_wide ? WIDE(kA64Csneg4rrrc) : kA64Csneg4rrrc;
+ } else {
+ // Csel. The rest. Use rl_result and a temp.
+ // TODO: To minimize the constants being loaded, check whether one can be inexpensively
+ // loaded as n - 1 or ~n.
+ rl_result = EvalLoc(rl_dest, result_reg_class, true);
+ rl_result_evaled = true;
+ LoadConstantNoClobber(rl_result.reg, true_val);
+ RegStorage t_reg2 = AllocTypedTemp(false, result_reg_class);
+ if (rl_dest.wide) {
+ if (t_reg2.Is32Bit()) {
+ t_reg2 = As64BitReg(t_reg2);
+ }
+ }
+ LoadConstantNoClobber(t_reg2, false_val);
+
+ // Use csel.
+ left_op = rl_result.reg.GetReg();
+ right_op = t_reg2.GetReg();
+ opcode = is_wide ? WIDE(kA64Csel4rrrc) : kA64Csel4rrrc;
+ }
+
+ if (!rl_result_evaled) {
+ rl_result = EvalLoc(rl_dest, result_reg_class, true);
+ }
+
+ NewLIR4(opcode, rl_result.reg.GetReg(), left_op, right_op, code);
+ } else {
+ RegLocation rl_true = mir_graph_->reg_location_[mir->ssa_rep->uses[1]];
+ RegLocation rl_false = mir_graph_->reg_location_[mir->ssa_rep->uses[2]];
+
+ rl_true = LoadValue(rl_true, result_reg_class);
+ rl_false = LoadValue(rl_false, result_reg_class);
+ rl_result = EvalLoc(rl_dest, result_reg_class, true);
+
+ int opcode = is_wide ? WIDE(kA64Csel4rrrc) : kA64Csel4rrrc;
+ NewLIR4(opcode, rl_result.reg.GetReg(),
+ rl_true.reg.GetReg(), rl_false.reg.GetReg(), code);
+ }
StoreValue(rl_dest, rl_result);
}
@@ -110,7 +214,6 @@
RegLocation rl_src2 = mir_graph_->GetSrcWide(mir, 2);
LIR* taken = &block_label_list_[bb->taken];
LIR* not_taken = &block_label_list_[bb->fall_through];
- rl_src1 = LoadValueWide(rl_src1, kCoreReg);
// Normalize such that if either operand is constant, src2 will be constant.
ConditionCode ccode = mir->meta.ccode;
if (rl_src1.is_const) {
@@ -118,16 +221,22 @@
ccode = FlipComparisonOrder(ccode);
}
+ rl_src1 = LoadValueWide(rl_src1, kCoreReg);
+
if (rl_src2.is_const) {
- rl_src2 = UpdateLocWide(rl_src2);
+ // TODO: Optimize for rl_src1.is_const? (Does happen in the boot image at the moment.)
+
int64_t val = mir_graph_->ConstantValueWide(rl_src2);
// Special handling using cbz & cbnz.
if (val == 0 && (ccode == kCondEq || ccode == kCondNe)) {
OpCmpImmBranch(ccode, rl_src1.reg, 0, taken);
OpCmpImmBranch(NegateComparison(ccode), rl_src1.reg, 0, not_taken);
return;
+ }
+
// Only handle Imm if src2 is not already in a register.
- } else if (rl_src2.location != kLocPhysReg) {
+ rl_src2 = UpdateLocWide(rl_src2);
+ if (rl_src2.location != kLocPhysReg) {
OpRegImm64(kOpCmp, rl_src1.reg, val);
OpCondBranch(ccode, taken);
OpCondBranch(NegateComparison(ccode), not_taken);