Handle regmasks in MachineCSE.

Don't attempt to extend physreg live ranges across calls.

<rdar://problem/10942095>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@151610 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 37 insertions, 0 deletions

diff --git a/lib/CodeGen/MachineCSE.cpp b/lib/CodeGen/MachineCSE.cpp
index 491a22c..27226f1 100644
--- a/lib/CodeGen/MachineCSE.cpp
+++ b/lib/CodeGen/MachineCSE.cpp
@@ -170,6 +170,8 @@ MachineCSE::isPhysDefTriviallyDead(unsigned Reg,
     bool SeenDef = false;
     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
       const MachineOperand &MO = I->getOperand(i);
+      if (MO.isRegMask() && MO.clobbersPhysReg(Reg))
+        SeenDef = true;
       if (!MO.isReg() || !MO.getReg())
         continue;
       if (!TRI->regsOverlap(MO.getReg(), Reg))
@@ -271,6 +273,10 @@ bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
 
     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
       const MachineOperand &MO = I->getOperand(i);
+      // RegMasks go on instructions like calls that clobber lots of physregs.
+      // Don't attempt to CSE across such an instruction.
+      if (MO.isRegMask())
+        return false;
       if (!MO.isReg() || !MO.isDef())
         continue;
       unsigned MOReg = MO.getReg();
diff --git a/test/CodeGen/ARM/cse-call.ll b/test/CodeGen/ARM/cse-call.ll
new file mode 100644
index 0000000..eff5de5
--- /dev/null
+++ b/test/CodeGen/ARM/cse-call.ll
@@ -0,0 +1,31 @@
+; RUN: llc < %s -mcpu=arm1136jf-s -verify-machineinstrs | FileCheck %s
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:32:64-v128:32:128-a0:0:32-n32-S32"
+target triple = "armv6-apple-ios0.0.0"
+
+; Don't CSE a cmp across a call that clobbers CPSR.
+;
+; CHECK: cmp
+; CHECK: S_trimzeros
+; CHECK: cmp
+; CHECK: strlen
+
+@F_floatmul.man1 = external global [200 x i8], align 1
+@F_floatmul.man2 = external global [200 x i8], align 1
+
+declare i32 @strlen(i8* nocapture) nounwind readonly
+declare void @S_trimzeros(...)
+
+define i8* @F_floatmul(i8* %f1, i8* %f2) nounwind ssp {
+entry:
+  br i1 undef, label %while.end42, label %while.body37
+
+while.body37:                                     ; preds = %while.body37, %entry
+  br i1 false, label %while.end42, label %while.body37
+
+while.end42:                                      ; preds = %while.body37, %entry
+  %. = select i1 undef, i8* getelementptr inbounds ([200 x i8]* @F_floatmul.man1, i32 0, i32 0), i8* getelementptr inbounds ([200 x i8]* @F_floatmul.man2, i32 0, i32 0)
+  %.92 = select i1 undef, i8* getelementptr inbounds ([200 x i8]* @F_floatmul.man2, i32 0, i32 0), i8* getelementptr inbounds ([200 x i8]* @F_floatmul.man1, i32 0, i32 0)
+  tail call void bitcast (void (...)* @S_trimzeros to void (i8*)*)(i8* %.92) nounwind
+  %call47 = tail call i32 @strlen(i8* %.) nounwind
+  unreachable
+}