For now, don't split live intervals around x87 stack register barriers. FpGET_ST0_80 must be right after a call instruction (and ADJCALLSTACKUP) so we need to find a way to prevent reload of x87 registers between them.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58230 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/include/llvm/Target/TargetInstrInfo.h b/include/llvm/Target/TargetInstrInfo.h
index 2fdcf8b..a37e9e6 100644
--- a/include/llvm/Target/TargetInstrInfo.h
+++ b/include/llvm/Target/TargetInstrInfo.h
@@ -407,6 +407,13 @@
return false;
}
+ /// IgnoreRegisterClassBarriers - Returns true if pre-register allocation
+ /// live interval splitting pass should ignore barriers of the specified
+ /// register class.
+ virtual bool IgnoreRegisterClassBarriers(const TargetRegisterClass *RC) const{
+ return true;
+ }
+
/// getPointerRegClass - Returns a TargetRegisterClass used for pointer
/// values.
virtual const TargetRegisterClass *getPointerRegClass() const {
@@ -425,7 +432,6 @@
/// GetFunctionSizeInBytes - Returns the size of the specified MachineFunction.
///
virtual unsigned GetFunctionSizeInBytes(const MachineFunction &MF) const = 0;
-
};
/// TargetInstrInfoImpl - This is the default implementation of
diff --git a/lib/CodeGen/PreAllocSplitting.cpp b/lib/CodeGen/PreAllocSplitting.cpp
index 547c528..33e3e80 100644
--- a/lib/CodeGen/PreAllocSplitting.cpp
+++ b/lib/CodeGen/PreAllocSplitting.cpp
@@ -659,6 +659,8 @@
// by the current barrier.
SmallVector<LiveInterval*, 8> Intervals;
for (const TargetRegisterClass **RC = RCs; *RC; ++RC) {
+ if (TII->IgnoreRegisterClassBarriers(*RC))
+ continue;
std::vector<unsigned> &VRs = MRI->getRegClassVirtRegs(*RC);
for (unsigned i = 0, e = VRs.size(); i != e; ++i) {
unsigned Reg = VRs[i];
diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index 04f10c0..c08fdd8 100644
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -2411,6 +2411,14 @@
return false;
}
+bool X86InstrInfo::
+IgnoreRegisterClassBarriers(const TargetRegisterClass *RC) const {
+ // FIXME: Ignore bariers of x87 stack registers for now. We can't
+ // allow any loads of these registers before FpGet_ST0_80.
+ return RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
+ RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass;
+}
+
const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
if (Subtarget->is64Bit())
diff --git a/lib/Target/X86/X86InstrInfo.h b/lib/Target/X86/X86InstrInfo.h
index 17be894..dc76f3e 100644
--- a/lib/Target/X86/X86InstrInfo.h
+++ b/lib/Target/X86/X86InstrInfo.h
@@ -405,6 +405,11 @@
virtual
bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
+ /// IgnoreRegisterClassBarriers - Returns true if pre-register allocation
+ /// live interval splitting pass should ignore barriers of the specified
+ /// register class.
+ bool IgnoreRegisterClassBarriers(const TargetRegisterClass *RC) const;
+
const TargetRegisterClass *getPointerRegClass() const;
// getBaseOpcodeFor - This function returns the "base" X86 opcode for the
diff --git a/lib/Target/X86/X86RegisterInfo.h b/lib/Target/X86/X86RegisterInfo.h
index 93e8613..25d7b07 100644
--- a/lib/Target/X86/X86RegisterInfo.h
+++ b/lib/Target/X86/X86RegisterInfo.h
@@ -94,6 +94,10 @@
/// Code Generation virtual methods...
///
+
+ /// getCrossCopyRegClass - Returns a legal register class to copy a register
+ /// in the specified class to or from. Returns NULL if it is possible to copy
+ /// between a two registers of the specified class.
const TargetRegisterClass *
getCrossCopyRegClass(const TargetRegisterClass *RC) const;
diff --git a/test/CodeGen/X86/pre-split7.ll b/test/CodeGen/X86/pre-split7.ll
new file mode 100644
index 0000000..7f7b933
--- /dev/null
+++ b/test/CodeGen/X86/pre-split7.ll
@@ -0,0 +1,34 @@
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 -pre-alloc-split
+
+@object_distance = external global double, align 8 ; <double*> [#uses=1]
+@axis_slope_angle = external global double, align 8 ; <double*> [#uses=1]
+@current_surfaces.b = external global i1 ; <i1*> [#uses=1]
+
+declare double @sin(double) nounwind readonly
+
+declare double @asin(double) nounwind readonly
+
+declare double @tan(double) nounwind readonly
+
+define fastcc void @trace_line(i32 %line) nounwind {
+entry:
+ %.b3 = load i1* @current_surfaces.b ; <i1> [#uses=1]
+ br i1 %.b3, label %bb, label %return
+
+bb: ; preds = %bb, %entry
+ %0 = tail call double @asin(double 0.000000e+00) nounwind readonly ; <double> [#uses=1]
+ %1 = add double 0.000000e+00, %0 ; <double> [#uses=2]
+ %2 = tail call double @asin(double 0.000000e+00) nounwind readonly ; <double> [#uses=1]
+ %3 = sub double %1, %2 ; <double> [#uses=2]
+ store double %3, double* @axis_slope_angle, align 8
+ %4 = fdiv double %1, 2.000000e+00 ; <double> [#uses=1]
+ %5 = tail call double @sin(double %4) nounwind readonly ; <double> [#uses=1]
+ %6 = mul double 0.000000e+00, %5 ; <double> [#uses=1]
+ %7 = tail call double @tan(double %3) nounwind readonly ; <double> [#uses=0]
+ %8 = add double 0.000000e+00, %6 ; <double> [#uses=1]
+ store double %8, double* @object_distance, align 8
+ br label %bb
+
+return: ; preds = %entry
+ ret void
+}