if (NOp1.isReg() && I.getOperand(0).getReg() == NOp1.getReg())
secondRegMatch = true;
- for (auto T : CurrentPacketMIs) {
- SUnit *PacketSU = MIToSUnit.find(T)->second;
- MachineInstr &PI = *PacketSU->getInstr();
+ for (MachineInstr *PI : CurrentPacketMIs) {
// NVJ can not be part of the dual jump - Arch Spec: section 7.8.
- if (PI.isCall()) {
+ if (PI->isCall()) {
Dependence = true;
break;
}
// 3. If the second operand of the nvj is newified, (which means
// first operand is also a reg), first reg is not defined in
// the same packet.
- if (PI.getOpcode() == Hexagon::S2_allocframe || PI.mayStore() ||
- HII->isLoopN(PI)) {
+ if (PI->getOpcode() == Hexagon::S2_allocframe || PI->mayStore() ||
+ HII->isLoopN(*PI)) {
Dependence = true;
break;
}
// Check #2/#3.
const MachineOperand &OpR = secondRegMatch ? NOp0 : NOp1;
- if (OpR.isReg() && PI.modifiesRegister(OpR.getReg(), HRI)) {
+ if (OpR.isReg() && PI->modifiesRegister(OpR.getReg(), HRI)) {
Dependence = true;
break;
}
}
+ GlueToNewValueJump = true;
if (Dependence)
return false;
- GlueToNewValueJump = true;
}
// There no dependency between a prolog instruction and its successor.
if (ChangedOffset != INT64_MAX)
undoChangedOffset(I);
- else if (updateOffset(SUI, SUJ)) {
+
+ if (GlueToNewValueJump) {
+ // Putting I and J together would prevent the new-value jump from being
+ // packetized with the producer. In that case I and J must be separated.
+ GlueToNewValueJump = false;
+ return false;
+ }
+
+ if (ChangedOffset == INT64_MAX && updateOffset(SUI, SUJ)) {
FoundSequentialDependence = false;
Dependence = false;
return true;
--- /dev/null
+# RUN: llc -march=hexagon -run-pass hexagon-packetizer %s -o - | FileCheck %s
+
+# Make sure that the new-value jump is packetized with the producer. In this
+# case, the loads cold be packetized together (with updating the offset in
+# the second load), but then the new-value jump would not be possible to
+# put in the same packet.
+
+# CHECK-LABEL: name: fred
+# CHECK: BUNDLE
+# CHECK-NEXT: %r3 = L2_loadri_io %r1, 0
+# CHECK-NEXT: J4_cmpgtu_f_jumpnv_t internal killed %r3
+
+
+--- |
+ define void @fred() { ret void }
+ @array = external global [256 x i32], align 8
+...
+
+---
+name: fred
+tracksRegLiveness: true
+body: |
+ bb.0:
+ successors: %bb.1
+ %r1 = A2_tfrsi @array
+ %r2, %r1 = L2_loadri_pi %r1, 4
+ %r3 = L2_loadri_io %r1, 0
+ J4_cmpgtu_f_jumpnv_t killed %r3, killed %r2, %bb.1, implicit-def %pc
+
+ bb.1:
+...