Add a hack to the scheduler to disable pseudo-two-address dependencies in

basic blocks containing calls. This works around a problem in which
these artificial dependencies can get tied up in calling seqeunce
scheduling in a way that makes the graph unschedulable with the current
approach of using artificial physical register dependencies for calling
sequences. This fixes PR11314.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@144124 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 30 insertions, 10 deletions

diff --git a/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp b/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
index cab303dd5c..f965a5e8ab 100644
--- a/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
+++ b/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp
@@ -1666,7 +1666,7 @@ public:
 
 protected:
   bool canClobber(const SUnit *SU, const SUnit *Op);
-  void AddPseudoTwoAddrDeps();
+  void AddPseudoTwoAddrDeps(const TargetInstrInfo *TII);
   void PrescheduleNodesWithMultipleUses();
   void CalculateSethiUllmanNumbers();
 };
@@ -2628,7 +2628,7 @@ bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
 void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
   SUnits = &sunits;
   // Add pseudo dependency edges for two-address nodes.
-  AddPseudoTwoAddrDeps();
+  AddPseudoTwoAddrDeps(TII);
   // Reroute edges to nodes with multiple uses.
   if (!TracksRegPressure)
     PrescheduleNodesWithMultipleUses();
@@ -2855,7 +2855,17 @@ void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
 /// one that has a CopyToReg use (more likely to be a loop induction update).
 /// If both are two-address, but one is commutable while the other is not
 /// commutable, favor the one that's not commutable.
-void RegReductionPQBase::AddPseudoTwoAddrDeps() {
+void RegReductionPQBase::AddPseudoTwoAddrDeps(const TargetInstrInfo *TII) {
+  // If the graph contains any calls, disable this optimization.
+  // FIXME: This is a kludge to work around the fact that the artificial edges
+  // can combine with the way call sequences use physical register dependencies
+  // to model their resource usage to create unschedulable graphs.
+  for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
+    for (SDNode *Node = (*SUnits)[i].getNode(); Node; Node = Node->getGluedNode())
+      if (Node->isMachineOpcode() &&
+          Node->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode())
+        return;
+
   for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
     SUnit *SU = &(*SUnits)[i];
     if (!SU->isTwoAddress)
diff --git a/test/CodeGen/X86/fold-pcmpeqd-0.ll b/test/CodeGen/X86/fold-pcmpeqd-0.ll
index 647bbdb7f0..6095a9cd20 100644
--- a/test/CodeGen/X86/fold-pcmpeqd-0.ll
+++ b/test/CodeGen/X86/fold-pcmpeqd-0.ll
@@ -3,15 +3,9 @@
 
 ; This testcase shouldn't need to spill the -1 value,
 ; so it should just use pcmpeqd to materialize an all-ones vector.
-; For i386, cp load of -1 are folded.
 
-; With -regalloc=greedy, the live range is split before spilling, so the first
-; pcmpeq doesn't get folded as a constant pool load.
-
-; I386-NOT: pcmpeqd
-; I386: orps LCPI0_2, %xmm
+; I386: pcmpeqd
 ; I386-NOT: pcmpeqd
-; I386: orps LCPI0_2, %xmm
 
 ; X86-64: pcmpeqd
 ; X86-64-NOT: pcmpeqd
diff --git a/test/CodeGen/X86/multiple-libcalls-and-twoaddr-deps-scheduling.ll b/test/CodeGen/X86/multiple-libcalls-and-twoaddr-deps-scheduling.ll
new file mode 100644
index 0000000000..b114a57966
--- /dev/null
+++ b/test/CodeGen/X86/multiple-libcalls-and-twoaddr-deps-scheduling.ll
@@ -0,0 +1,16 @@
+; RUN: llc -march=x86 -mcpu=pentium4 -mtriple=i686-none-linux < %s
+; PR11314
+
+; Make sure the scheduler's hack to insert artificial dependencies to optimize
+; two-address instruction scheduling doesn't interfere with the scheduler's
+; hack to model call sequences as artificial physical registers.
+
+define inreg { i64, i64 } @sscanf(i32 inreg %base.1.i) nounwind {
+entry:
+  %conv38.i92.i = sext i32 %base.1.i to i64
+  %rem.i93.i = urem i64 10, %conv38.i92.i
+  %div.i94.i = udiv i64 10, %conv38.i92.i
+  %a = insertvalue { i64, i64 } undef, i64 %rem.i93.i, 0
+  %b = insertvalue { i64, i64 } %a, i64 %div.i94.i, 1
+  ret { i64, i64 } %b
+}