[PATCH] D53485: [ScheduleDAGRRList] Do not preschedule the node has ADJCALLSTACKDOWN parent
Shiva Chen via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Wed Nov 21 23:56:42 PST 2018
shiva0217 updated this revision to Diff 175015.
shiva0217 added a comment.
Add the test case provided by Tim.
Repository:
rL LLVM
https://reviews.llvm.org/D53485
Files:
lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
test/CodeGen/AVR/pre-schedule.ll
Index: test/CodeGen/AVR/pre-schedule.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AVR/pre-schedule.ll
@@ -0,0 +1,32 @@
+; RUN: llc < %s -march=avr | FileCheck %s
+target triple = "avr-unknown-unknown"
+
+; The case illustrate DAG schedular may pre-schedule the node has
+; ADJCALLSTACKDOWN parent, so ADJCALLSTACKUP may hold CallResource too long
+; and make other calls can't be scheduled. If there's no other available node
+; to schedule, the scheduler will try to rename the register by creating the
+; copy to avoid the conflict which will fail because CallResource is not a real
+; physical register.
+;
+; The issue is found by Tim on https://github.com/avr-rust/rust/issues/111 and
+; discuss in http://lists.llvm.org/pipermail/llvm-dev/2018-October/127083.html.
+
+define void @"main"() addrspace(1) {
+start:
+ %0 = or i64 undef, undef
+ br i1 undef, label %mul_and_call, label %fail
+
+ mul_and_call:
+ %1 = mul i64 %0, %0
+ call addrspace(1) void @"three_ints"(i64 undef, i64 %1, i64 %0)
+; The CHECK line only want to make sure the following assertion message
+; won't trigger due to create copy of artificial CallResource register.
+; llc: llvm/lib/CodeGen/TargetRegisterInfo.cpp:203: const llvm::TargetRegisterClass* llvm::TargetRegisterInfo::getMinimalPhysRegClass(unsigned int, llvm::MVT) const: Assertion `BestRC && "Couldn't find the register class"' failed.
+; CHECK: call __muldi3
+ ret void
+
+ fail:
+ ret void
+}
+
+declare void @"three_ints"(i64, i64, i64) addrspace(1)
Index: lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
+++ lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
@@ -2939,6 +2939,29 @@
(cast<RegisterSDNode>(N->getOperand(1))->getReg()))
continue;
+ SDNode *PredFrameSetup = nullptr;
+ for (const SDep &Pred : SU.Preds)
+ if (Pred.isCtrl() && Pred.getSUnit()) {
+ // Find the predecessor which is not data dependence.
+ SDNode *PredND = Pred.getSUnit()->getNode();
+
+ // If PredND is FrameSetup, we should not pre-scheduled the node,
+ // or else, when bottom up scheduling, ADJCALLSTACKDOWN and
+ // ADJCALLSTACKUP may hold CallResource too long and make other
+ // calls can't be scheduled. If there's no other available node
+ // to schedule, the schedular will try to rename the register by
+ // creating copy to avoid the conflict which will fail because
+ // CallResource is not a real physical register.
+ if (PredND && PredND->isMachineOpcode() &&
+ (PredND->getMachineOpcode() == TII->getCallFrameSetupOpcode())) {
+ PredFrameSetup = PredND;
+ break;
+ }
+ }
+ // Skip the node has FrameSetup parent.
+ if (PredFrameSetup != nullptr)
+ continue;
+
// Locate the single data predecessor.
SUnit *PredSU = nullptr;
for (const SDep &Pred : SU.Preds)
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D53485.175015.patch
Type: text/x-patch
Size: 3056 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20181122/143ca858/attachment.bin>
More information about the llvm-commits
mailing list