From 1fd36e41e408307a4e7362cc9cd83aeb77ee71dd Mon Sep 17 00:00:00 2001 From: Preston Gurd Date: Wed, 27 Mar 2013 23:16:18 +0000 Subject: [PATCH] This patch follows is a follow up to r178171, which uses the register form of call in preference to memory indirect on Atom. In this case, the patch applies the optimization to the code for reloading spilled registers. The patch also includes changes to sibcall.ll and movgs.ll, which were failing on the Atom buildbot after the first patch was applied. This patch by Sriram Murali. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178193 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/X86/X86InstrInfo.cpp | 9 + .../atom-call-reg-indirect-foldedreload32.ll | 204 +++++++++++++++++ .../atom-call-reg-indirect-foldedreload64.ll | 213 ++++++++++++++++++ test/CodeGen/X86/movgs.ll | 6 +- test/CodeGen/X86/sibcall.ll | 4 +- 5 files changed, 431 insertions(+), 5 deletions(-) create mode 100644 test/CodeGen/X86/atom-call-reg-indirect-foldedreload32.ll create mode 100644 test/CodeGen/X86/atom-call-reg-indirect-foldedreload64.ll diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp index 17714acd86c..7ba542c8752 100644 --- a/lib/Target/X86/X86InstrInfo.cpp +++ b/lib/Target/X86/X86InstrInfo.cpp @@ -3655,7 +3655,16 @@ X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, const SmallVectorImpl &MOs, unsigned Size, unsigned Align) const { const DenseMap > *OpcodeTablePtr = 0; + bool isCallRegIndirect = TM.getSubtarget().callRegIndirect(); bool isTwoAddrFold = false; + + // Atom favors register form of call. So, we do not fold loads into calls + // when X86Subtarget is Atom. + if (isCallRegIndirect && + (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) { + return NULL; + } + unsigned NumOps = MI->getDesc().getNumOperands(); bool isTwoAddr = NumOps > 1 && MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1; diff --git a/test/CodeGen/X86/atom-call-reg-indirect-foldedreload32.ll b/test/CodeGen/X86/atom-call-reg-indirect-foldedreload32.ll new file mode 100644 index 00000000000..4192f13d896 --- /dev/null +++ b/test/CodeGen/X86/atom-call-reg-indirect-foldedreload32.ll @@ -0,0 +1,204 @@ +; RUN: llc < %s -mtriple=i386-linux-gnu -mcpu=atom 2>&1 | \ +; RUN: grep "calll" | not grep "(" +; RUN: llc < %s -mtriple=i386-linux-gnu -mcpu=core2 2>&1 | \ +; RUN: grep "calll" | grep "4-byte Folded Reload" + +%struct.targettype = type {i32} +%struct.op_ptr1 = type opaque +%struct.op_ptr2 = type opaque +%union.anon = type { [8 x i32], [48 x i8] } +%struct.const1 = type { [64 x i16], i8 } +%struct.const2 = type { [17 x i8], [256 x i8], i8 } + +%struct.ref1 = type { void (%struct.ref2*)*, i32 (%struct.ref2*)*, void (%struct.ref2*)*, i32 (%struct.ref2*, i8***)*, %struct.op_ptr2** } +%struct.ref2 = type { %struct.localref13*, %struct.localref15*, %struct.localref12*, i8*, i8, i32, %struct.localref11*, i32, i32, i32, i32, i32, i32, i32, double, i8, i8, i32, i8, i8, i8, i32, i8, i32, i8, i8, i8, i32, i32, i32, i32, i32, i32, i8**, i32, i32, i32, i32, i32, [64 x i32]*, [4 x %struct.const1*], [4 x %struct.const2*], [4 x %struct.const2*], i32, %struct.ref3*, i8, i8, [16 x i8], [16 x i8], [16 x i8], i32, i8, i8, i8, i8, i16, i16, i8, i8, i8, %struct.localref10*, i32, i32, i32, i32, i8*, i32, [4 x %struct.ref3*], i32, i32, i32, [10 x i32], i32, i32, i32, i32, i32, %struct.localref8*, %struct.localref9*, %struct.ref1*, %struct.localref7*, %struct.localref6*, %struct.localref5*, %struct.localref1*, %struct.ref4*, %struct.localref2*, %struct.localref3*, %struct.localref4* } +%struct.ref3 = type { i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i8, i32, i32, i32, i32, i32, i32, %struct.const1*, i8* } +%struct.ref4 = type { void (%struct.ref2*)*, [5 x void (%struct.ref2*, %struct.ref3*, i16*, i8**, i32)*] } + +%struct.localref1 = type { void (%struct.ref2*)*, i8 (%struct.ref2*, [64 x i16]**)*, i8 } +%struct.localref2 = type { void (%struct.ref2*)*, void (%struct.ref2*, i8***, i32*, i32, i8**, i32*, i32)*, i8 } +%struct.localref3 = type { void (%struct.ref2*)*, void (%struct.ref2*, i8***, i32, i8**, i32)* } +%struct.localref4 = type { {}*, void (%struct.ref2*, i8**, i8**, i32)*, void (%struct.ref2*)*, void (%struct.ref2*)* } +%struct.localref5 = type { void (%struct.ref2*)*, i32 (%struct.ref2*)*, i8 (%struct.ref2*)*, i8, i8, i32, i32 } +%struct.localref6 = type { i32 (%struct.ref2*)*, void (%struct.ref2*)*, void (%struct.ref2*)*, void (%struct.ref2*)*, i8, i8 } +%struct.localref7 = type { void (%struct.ref2*, i32)*, void (%struct.ref2*, i8***, i32*, i32, i8**, i32*, i32)* } +%struct.localref8 = type { void (%struct.ref2*)*, void (%struct.ref2*)*, i8 } +%struct.localref9 = type { void (%struct.ref2*, i32)*, void (%struct.ref2*, i8**, i32*, i32)* } +%struct.localref10 = type { %struct.localref10*, i8, i32, i32, i8* } +%struct.localref11 = type { i8*, %struct.targettype, void (%struct.ref2*)*, i8 (%struct.ref2*)*, void (%struct.ref2*, %struct.targettype)*, i8 (%struct.ref2*, i32)*, void (%struct.ref2*)* } +%struct.localref12 = type { {}*, %struct.targettype, %struct.targettype, i32, i32 } +%struct.localref13 = type { void (%struct.localref14*)*, void (%struct.localref14*, i32)*, void (%struct.localref14*)*, void (%struct.localref14*, i8*)*, void (%struct.localref14*)*, i32, %union.anon, i32, %struct.targettype, i8**, i32, i8**, i32, i32 } +%struct.localref14 = type { %struct.localref13*, %struct.localref15*, %struct.localref12*, i8*, i8, i32 } +%struct.localref15 = type { i8* (%struct.localref14*, i32, %struct.targettype)*, i8* (%struct.localref14*, i32, %struct.targettype)*, i8** (%struct.localref14*, i32, i32, i32)*, [64 x i16]** (%struct.localref14*, i32, i32, i32)*, %struct.op_ptr1* (%struct.localref14*, i32, i8, i32, i32, i32)*, %struct.op_ptr2* (%struct.localref14*, i32, i8, i32, i32, i32)*, {}*, i8** (%struct.localref14*, %struct.op_ptr1*, i32, i32, i8)*, [64 x i16]** (%struct.localref14*, %struct.op_ptr2*, i32, i32, i8)*, void (%struct.localref14*, i32)*, {}*, %struct.targettype, %struct.targettype} + +define internal i32 @foldedreload(%struct.ref2* %cinfo, i8*** nocapture %output1) { + %1 = getelementptr inbounds %struct.ref2* %cinfo, i32 0, i32 79 + %2 = load %struct.ref1** %1, align 4 + %3 = getelementptr inbounds %struct.ref2* %cinfo, i32 0, i32 68 + %4 = load i32* %3, align 4 + %5 = add i32 %4, -1 + %6 = getelementptr inbounds %struct.ref2* %cinfo, i32 0, i32 64 + %7 = load i32* %6, align 4 + %8 = add i32 %7, -1 + %9 = getelementptr inbounds %struct.ref1* %2, i32 1, i32 1 + %10 = bitcast i32 (%struct.ref2*)** %9 to i32* + %11 = load i32* %10, align 4 + %12 = getelementptr inbounds %struct.ref1* %2, i32 1, i32 2 + %13 = bitcast void (%struct.ref2*)** %12 to i32* + %14 = load i32* %13, align 4 + %15 = icmp slt i32 %11, %14 + br i1 %15, label %.lr.ph18, label %._crit_edge19 + +.lr.ph18: + %16 = getelementptr inbounds %struct.ref1* %2, i32 1 + %17 = bitcast %struct.ref1* %16 to i32* + %18 = getelementptr inbounds %struct.ref1* %16, i32 0, i32 0 + %19 = getelementptr inbounds %struct.ref2* %cinfo, i32 0, i32 66 + %20 = getelementptr inbounds %struct.ref2* %cinfo, i32 0, i32 84 + %21 = getelementptr inbounds %struct.ref2* %cinfo, i32 0, i32 36 + %22 = getelementptr inbounds %struct.ref1* %2, i32 1, i32 3 + %23 = bitcast i32 (%struct.ref2*, i8***)** %22 to [10 x [64 x i16]*]* + %.pre = load i32* %17, align 4 + br label %24 + +;