replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[SystemZ] Improve handling of SETCC 

We previously used the default expansion to SELECT_CC, which in turn would
expand to "LHI; BRC; LHI".  In most cases it's better to use an IPM-based
sequence instead.

llvm-svn: 192784
This commit is contained in:
Richard Sandiford 2013-10-16 11:10:55 +00:00
parent 374a0e50c4
commit f722a8e30e
8 changed files with 380 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
llvm/lib/Target/SystemZ/SystemZ.h
View File

@ -68,6 +68,9 @@ namespace llvm {
const unsigned CCMASK_TM_MSB_1 = CCMASK_2 | CCMASK_3;
const unsigned CCMASK_TM = CCMASK_ANY;
// The position of the low CC bit in an IPM result.
const unsigned IPM_CC = 28;
// Mask assignments for PFD.
const unsigned PFD_READ = 1;
const unsigned PFD_WRITE = 2;

116
llvm/lib/Target/SystemZ/SystemZISelLowering.cpp
View File

@ -27,6 +27,19 @@
using namespace llvm;
namespace {
// Represents a sequence for extracting a 0/1 value from an IPM result:
// (((X ^ XORValue) + AddValue) >> Bit)
struct IPMConversion {
IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
: XORValue(xorValue), AddValue(addValue), Bit(bit) {}
int64_t XORValue;
int64_t AddValue;
unsigned Bit;
};
}
// Classify VT as either 32 or 64 bit.
static bool is32Bit(EVT VT) {
switch (VT.getSimpleVT().SimpleTy) {
@ -88,8 +101,8 @@ SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm)
++I) {
MVT VT = MVT::SimpleValueType(I);
if (isTypeLegal(VT)) {
// Expand SETCC(X, Y, COND) into SELECT_CC(X, Y, 1, 0, COND).
setOperationAction(ISD::SETCC, VT, Expand);
// Lower SET_CC into an IPM-based sequence.
setOperationAction(ISD::SETCC, VT, Custom);
// Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE).
setOperationAction(ISD::SELECT, VT, Expand);
@ -980,6 +993,73 @@ static unsigned CCMaskForCondCode(ISD::CondCode CC) {
#undef CONV
}
// Return a sequence for getting a 1 from an IPM result when CC has a
// value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
// The handling of CC values outside CCValid doesn't matter.
static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
// Deal with cases where the result can be taken directly from a bit
// of the IPM result.
if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
return IPMConversion(0, 0, SystemZ::IPM_CC);
if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
// Deal with cases where we can add a value to force the sign bit
// to contain the right value. Putting the bit in 31 means we can
// use SRL rather than RISBG(L), and also makes it easier to get a
// 0/-1 value, so it has priority over the other tests below.
//
// These sequences rely on the fact that the upper two bits of the
// IPM result are zero.
uint64_t TopBit = uint64_t(1) << 31;
if (CCMask == (CCValid & SystemZ::CCMASK_0))
return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_0
| SystemZ::CCMASK_1
| SystemZ::CCMASK_2)))
return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & SystemZ::CCMASK_3))
return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_1
| SystemZ::CCMASK_2
| SystemZ::CCMASK_3)))
return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
// Next try inverting the value and testing a bit. 0/1 could be
// handled this way too, but we dealt with that case above.
if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
return IPMConversion(-1, 0, SystemZ::IPM_CC);
// Handle cases where adding a value forces a non-sign bit to contain
// the right value.
if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
// The remaing cases are 1, 2, 0/1/3 and 0/2/3. All these are
// can be done by inverting the low CC bit and applying one of the
// sign-based extractions above.
if (CCMask == (CCValid & SystemZ::CCMASK_1))
return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & SystemZ::CCMASK_2))
return IPMConversion(1 << SystemZ::IPM_CC,
TopBit - (3 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_0
| SystemZ::CCMASK_1
| SystemZ::CCMASK_3)))
return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_0
| SystemZ::CCMASK_2
| SystemZ::CCMASK_3)))
return IPMConversion(1 << SystemZ::IPM_CC,
TopBit - (1 << SystemZ::IPM_CC), 31);
llvm_unreachable("Unexpected CC combination");
}
// If a comparison described by IsUnsigned, CCMask, CmpOp0 and CmpOp1
// can be converted to a comparison against zero, adjust the operands
// as necessary.
@ -1401,6 +1481,36 @@ static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT,
Odd = DAG.getTargetExtractSubreg(SystemZ::odd128(Is32Bit), DL, VT, Result);
}
SDValue SystemZTargetLowering::lowerSETCC(SDValue Op,
SelectionDAG &DAG) const {
SDValue CmpOp0 = Op.getOperand(0);
SDValue CmpOp1 = Op.getOperand(1);
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
SDLoc DL(Op);
unsigned CCValid, CCMask;
SDValue Glue = emitCmp(TM, DAG, DL, CmpOp0, CmpOp1, CC, CCValid, CCMask);
IPMConversion Conversion = getIPMConversion(CCValid, CCMask);
SDValue Result = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
if (Conversion.XORValue)
Result = DAG.getNode(ISD::XOR, DL, MVT::i32, Result,
DAG.getConstant(Conversion.XORValue, MVT::i32));
if (Conversion.AddValue)
Result = DAG.getNode(ISD::ADD, DL, MVT::i32, Result,
DAG.getConstant(Conversion.AddValue, MVT::i32));
// The SHR/AND sequence should get optimized to an RISBG.
Result = DAG.getNode(ISD::SRL, DL, MVT::i32, Result,
DAG.getConstant(Conversion.Bit, MVT::i32));
if (Conversion.Bit != 31)
Result = DAG.getNode(ISD::AND, DL, MVT::i32, Result,
DAG.getConstant(1, MVT::i32));
return Result;
}
SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
SDValue Chain = Op.getOperand(0);
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
@ -1997,6 +2107,8 @@ SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
return lowerBR_CC(Op, DAG);
case ISD::SELECT_CC:
return lowerSELECT_CC(Op, DAG);
case ISD::SETCC:
return lowerSETCC(Op, DAG);
case ISD::GlobalAddress:
return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
case ISD::GlobalTLSAddress:

1
llvm/lib/Target/SystemZ/SystemZISelLowering.h
View File

@ -252,6 +252,7 @@ private:
const SystemZTargetMachine &TM;
// Implement LowerOperation for individual opcodes.
SDValue lowerSETCC(SDValue Op, SelectionDAG &DAG) const;
SDValue lowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
SDValue lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
SDValue lowerGlobalAddress(GlobalAddressSDNode *Node,

2
llvm/lib/Target/SystemZ/SystemZInstrInfo.cpp
View File

@ -449,7 +449,7 @@ static bool removeIPMBasedCompare(MachineInstr *Compare, unsigned SrcReg,
return false;
MachineInstr *SRL = getDef(RLL->getOperand(1).getReg(), MRI);
if (!SRL || !isShift(SRL, SystemZ::SRL, 28))
if (!SRL || !isShift(SRL, SystemZ::SRL, SystemZ::IPM_CC))
return false;
MachineInstr *IPM = getDef(SRL->getOperand(1).getReg(), MRI);

2
llvm/lib/Target/SystemZ/SystemZSelectionDAGInfo.cpp
View File

@ -183,7 +183,7 @@ static SDValue emitCLC(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
static SDValue addIPMSequence(SDLoc DL, SDValue Glue, SelectionDAG &DAG) {
SDValue IPM = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i32, IPM,
DAG.getConstant(28, MVT::i32));
DAG.getConstant(SystemZ::IPM_CC, MVT::i32));
SDValue ROTL = DAG.getNode(ISD::ROTL, DL, MVT::i32, SRL,
DAG.getConstant(31, MVT::i32));
return ROTL;

30
llvm/test/CodeGen/SystemZ/branch-07.ll
View File

@ -97,10 +97,9 @@ exit:
; Test a vector of 0/-1 results for i32 EQ.
define i64 @f7(i64 %a, i64 %b) {
; CHECK-LABEL: f7:
; CHECK: lhi [[REG:%r[0-5]]], -1
; CHECK: crje {{%r[0-5]}}
; CHECK: lhi [[REG]], 0
; CHECK-NOT: sra
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK: afi [[REG]], -268435456
; CHECK: sra [[REG]], 31
; CHECK: br %r14
%avec = bitcast i64 %a to <2 x i32>
%bvec = bitcast i64 %b to <2 x i32>
@ -113,10 +112,9 @@ define i64 @f7(i64 %a, i64 %b) {
; Test a vector of 0/-1 results for i32 NE.
define i64 @f8(i64 %a, i64 %b) {
; CHECK-LABEL: f8:
; CHECK: lhi [[REG:%r[0-5]]], -1
; CHECK: crjlh {{%r[0-5]}}
; CHECK: lhi [[REG]], 0
; CHECK-NOT: sra
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK: afi [[REG]], 1879048192
; CHECK: sra [[REG]], 31
; CHECK: br %r14
%avec = bitcast i64 %a to <2 x i32>
%bvec = bitcast i64 %b to <2 x i32>
@ -129,10 +127,10 @@ define i64 @f8(i64 %a, i64 %b) {
; Test a vector of 0/-1 results for i64 EQ.
define void @f9(i64 %a, i64 %b, <2 x i64> *%dest) {
; CHECK-LABEL: f9:
; CHECK: lghi [[REG:%r[0-5]]], -1
; CHECK: crje {{%r[0-5]}}
; CHECK: lghi [[REG]], 0
; CHECK-NOT: sra
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK: afi [[REG]], -268435456
; CHECK: sllg [[REG2:%r[0-5]]], [[REG]], 32
; CHECK: srag {{%r[0-5]}}, [[REG2]], 63
; CHECK: br %r14
%avec = bitcast i64 %a to <2 x i32>
%bvec = bitcast i64 %b to <2 x i32>
@ -145,10 +143,10 @@ define void @f9(i64 %a, i64 %b, <2 x i64> *%dest) {
; Test a vector of 0/-1 results for i64 NE.
define void @f10(i64 %a, i64 %b, <2 x i64> *%dest) {
; CHECK-LABEL: f10:
; CHECK: lghi [[REG:%r[0-5]]], -1
; CHECK: crjlh {{%r[0-5]}}
; CHECK: lghi [[REG]], 0
; CHECK-NOT: sra
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK: afi [[REG]], 1879048192
; CHECK: sllg [[REG2:%r[0-5]]], [[REG]], 32
; CHECK: srag {{%r[0-5]}}, [[REG2]], 63
; CHECK: br %r14
%avec = bitcast i64 %a to <2 x i32>
%bvec = bitcast i64 %b to <2 x i32>

73
llvm/test/CodeGen/SystemZ/setcc-01.ll Normal file
View File

@ -0,0 +1,73 @@
; Test SETCC for every integer condition.
;
; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
; Test CC in { 0 }, with 3 don't care.
define i32 @f1(i32 %a, i32 %b) {
; CHECK-LABEL: f1:
; CHECK: ipm %r2
; CHECK-NEXT: afi %r2, -268435456
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = icmp eq i32 %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 1 }, with 3 don't care.
define i32 @f2(i32 %a, i32 %b) {
; CHECK-LABEL: f2:
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK-NEXT: risbg %r2, [[REG]], 63, 191, 36
; CHECK: br %r14
%cond = icmp slt i32 %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 0, 1 }, with 3 don't care.
define i32 @f3(i32 %a, i32 %b) {
; CHECK-LABEL: f3:
; CHECK: ipm %r2
; CHECK-NEXT: afi %r2, -536870912
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = icmp sle i32 %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 2 }, with 3 don't care.
define i32 @f4(i32 %a, i32 %b) {
; CHECK-LABEL: f4:
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK-NEXT: risbg %r2, [[REG]], 63, 191, 35
; CHECK: br %r14
%cond = icmp sgt i32 %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 0, 2 }, with 3 don't care.
define i32 @f5(i32 %a, i32 %b) {
; CHECK-LABEL: f5:
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK-NEXT: xilf [[REG]], 4294967295
; CHECK-NEXT: risbg %r2, [[REG]], 63, 191, 36
; CHECK: br %r14
%cond = icmp sge i32 %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 1, 2 }, with 3 don't care.
define i32 @f6(i32 %a, i32 %b) {
; CHECK-LABEL: f6:
; CHECK: ipm %r2
; CHECK-NEXT: afi %r2, 1879048192
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = icmp ne i32 %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}

173
llvm/test/CodeGen/SystemZ/setcc-02.ll Normal file
View File

@ -0,0 +1,173 @@
; Test SETCC for every floating-point condition.
;
; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
; Test CC in { 0 }
define i32 @f1(float %a, float %b) {
; CHECK-LABEL: f1:
; CHECK: ipm %r2
; CHECK-NEXT: afi %r2, -268435456
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp oeq float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 1 }
define i32 @f2(float %a, float %b) {
; CHECK-LABEL: f2:
; CHECK: ipm %r2
; CHECK-NEXT: xilf %r2, 268435456
; CHECK-NEXT: afi %r2, -268435456
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp olt float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 0, 1 }
define i32 @f3(float %a, float %b) {
; CHECK-LABEL: f3:
; CHECK: ipm %r2
; CHECK-NEXT: afi %r2, -536870912
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp ole float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 2 }
define i32 @f4(float %a, float %b) {
; CHECK-LABEL: f4:
; CHECK: ipm %r2
; CHECK-NEXT: xilf %r2, 268435456
; CHECK-NEXT: afi %r2, 1342177280
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp ogt float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 0, 2 }
define i32 @f5(float %a, float %b) {
; CHECK-LABEL: f5:
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK-NEXT: xilf [[REG]], 4294967295
; CHECK-NEXT: risbg %r2, [[REG]], 63, 191, 36
; CHECK: br %r14
%cond = fcmp oge float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 1, 2 }
define i32 @f6(float %a, float %b) {
; CHECK-LABEL: f6:
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK-NEXT: afi [[REG]], 268435456
; CHECK-NEXT: risbg %r2, [[REG]], 63, 191, 35
; CHECK: br %r14
%cond = fcmp one float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 0, 1, 2 }
define i32 @f7(float %a, float %b) {
; CHECK-LABEL: f7:
; CHECK: ipm %r2
; CHECK-NEXT: afi %r2, -805306368
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp ord float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 3 }
define i32 @f8(float %a, float %b) {
; CHECK-LABEL: f8:
; CHECK: ipm %r2
; CHECK-NEXT: afi %r2, 1342177280
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp uno float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 0, 3 }
define i32 @f9(float %a, float %b) {
; CHECK-LABEL: f9:
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK-NEXT: afi [[REG]], -268435456
; CHECK-NEXT: risbg %r2, [[REG]], 63, 191, 35
; CHECK: br %r14
%cond = fcmp ueq float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 1, 3 }
define i32 @f10(float %a, float %b) {
; CHECK-LABEL: f10:
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK-NEXT: risbg %r2, [[REG]], 63, 191, 36
; CHECK: br %r14
%cond = fcmp ult float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 0, 1, 3 }
define i32 @f11(float %a, float %b) {
; CHECK-LABEL: f11:
; CHECK: ipm %r2
; CHECK-NEXT: xilf %r2, 268435456
; CHECK-NEXT: afi %r2, -805306368
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp ule float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 2, 3 }
define i32 @f12(float %a, float %b) {
; CHECK-LABEL: f12:
; CHECK: ipm [[REG:%r[0-5]]]
; CHECK-NEXT: risbg %r2, [[REG]], 63, 191, 35
; CHECK: br %r14
%cond = fcmp ugt float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 0, 2, 3 }
define i32 @f13(float %a, float %b) {
; CHECK-LABEL: f13:
; CHECK: ipm %r2
; CHECK-NEXT: xilf %r2, 268435456
; CHECK-NEXT: afi %r2, 1879048192
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp uge float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}
; Test CC in { 1, 2, 3 }
define i32 @f14(float %a, float %b) {
; CHECK-LABEL: f14:
; CHECK: ipm %r2
; CHECK-NEXT: afi %r2, 1879048192
; CHECK-NEXT: srl %r2, 31
; CHECK: br %r14
%cond = fcmp une float %a, %b
%res = zext i1 %cond to i32
ret i32 %res
}