NewMIs.push_back(MIB);
}
+/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
+/// instruction with two undef reads of the register being defined. This is
+/// used for mapping:
+/// %xmm4 = V_SET0
+/// to:
+/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
+///
+static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) {
+ assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
+ unsigned Reg = MI->getOperand(0).getReg();
+ MI->setDesc(Desc);
+
+ // MachineInstr::addOperand() will insert explicit operands before any
+ // implicit operands.
+ MachineInstrBuilder(MI).addReg(Reg, RegState::Undef)
+ .addReg(Reg, RegState::Undef);
+ // But we don't trust that.
+ assert(MI->getOperand(1).getReg() == Reg &&
+ MI->getOperand(2).getReg() == Reg && "Misplaced operand");
+ return true;
+}
+
+bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
+ bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
+ switch (MI->getOpcode()) {
+ case X86::V_SET0:
+ return Expand2AddrUndef(MI, get(HasAVX ? X86::VPXORrr : X86::PXORrr));
+ }
+ return false;
+}
+
MachineInstr*
X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
int FrameIx, uint64_t Offset,
case X86::AVX_SET0PDY:
Alignment = 32;
break;
- case X86::V_SET0PS:
- case X86::V_SET0PD:
- case X86::V_SET0PI:
+ case X86::V_SET0:
case X86::V_SETALLONES:
- case X86::AVX_SET0PS:
- case X86::AVX_SET0PD:
- case X86::AVX_SET0PI:
case X86::AVX_SETALLONES:
Alignment = 16;
break;
SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
switch (LoadMI->getOpcode()) {
- case X86::V_SET0PS:
- case X86::V_SET0PD:
- case X86::V_SET0PI:
+ case X86::V_SET0:
case X86::V_SETALLONES:
- case X86::AVX_SET0PS:
- case X86::AVX_SET0PD:
- case X86::AVX_SET0PI:
case X86::AVX_SET0PSY:
case X86::AVX_SET0PDY:
case X86::AVX_SETALLONES:
case X86::FsFLD0SS:
case X86::VFsFLD0SD:
case X86::VFsFLD0SS: {
- // Folding a V_SET0P? or V_SETALLONES as a load, to ease register pressure.
+ // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
// Create a constant-pool entry and operands to load from it.
// Medium and large mode can't fold loads this way.
{ X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
{ X86::ORPSrm, X86::ORPDrm, X86::PORrm },
{ X86::ORPSrr, X86::ORPDrr, X86::PORrr },
- { X86::V_SET0PS, X86::V_SET0PD, X86::V_SET0PI },
{ X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
{ X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
// AVX 128-bit support
{ X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
{ X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
{ X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
- { X86::AVX_SET0PS, X86::AVX_SET0PD, X86::AVX_SET0PI },
{ X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
{ X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
// AVX 256-bit support
// AVX & SSE - Zero/One Vectors
//===----------------------------------------------------------------------===//
-// Alias instructions that map zero vector to pxor / xorp* for sse.
+// Alias instruction that maps zero vector to pxor / xorp* for sse.
+// This is expanded by ExpandPostRAPseudos to an xorps / vxorps, and then
+// swizzled by ExecutionDepsFix to pxor.
// We set canFoldAsLoad because this can be converted to a constant-pool
// load of an all-zeros value if folding it would be beneficial.
-// FIXME: Change encoding to pseudo! This is blocked right now by the x86
-// JIT implementation, it does not expand the instructions below like
-// X86MCInstLower does.
let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
- isCodeGenOnly = 1 in {
-def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
- [(set VR128:$dst, (v4f32 immAllZerosV))]>;
-def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
- [(set VR128:$dst, (v2f64 immAllZerosV))]>;
-let ExeDomain = SSEPackedInt in
-def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
- [(set VR128:$dst, (v4i32 immAllZerosV))]>;
+ isPseudo = 1 in {
+def V_SET0 : I<0, Pseudo, (outs VR128:$dst), (ins), "", []>;
}
-// The same as done above but for AVX. The 128-bit versions are the
-// same, but re-encoded. The 256-bit does not support PI version, and
-// doesn't need it because on sandy bridge the register is set to zero
+def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
+def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
+def : Pat<(v4i32 immAllZerosV), (V_SET0)>;
+def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
+def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
+def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
+
+
+// The same as done above but for AVX. The 256-bit ISA does not support PI,
+// and doesn't need it because on sandy bridge the register is set to zero
// at the rename stage without using any execution unit, so SET0PSY
// and SET0PDY can be used for vector int instructions without penalty
// FIXME: Change encoding to pseudo! This is blocked right now by the x86
// X86MCInstLower does.
let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
isCodeGenOnly = 1, Predicates = [HasAVX] in {
-def AVX_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
- [(set VR128:$dst, (v4f32 immAllZerosV))]>, VEX_4V;
-def AVX_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
- [(set VR128:$dst, (v2f64 immAllZerosV))]>, VEX_4V;
def AVX_SET0PSY : PSI<0x57, MRMInitReg, (outs VR256:$dst), (ins), "",
[(set VR256:$dst, (v8f32 immAllZerosV))]>, VEX_4V;
def AVX_SET0PDY : PDI<0x57, MRMInitReg, (outs VR256:$dst), (ins), "",
[(set VR256:$dst, (v4f64 immAllZerosV))]>, VEX_4V;
-let ExeDomain = SSEPackedInt in
-def AVX_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
- [(set VR128:$dst, (v4i32 immAllZerosV))]>;
}
-def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
-def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
-def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
// AVX has no support for 256-bit integer instructions, but since the 128-bit
// VPXOR instruction writes zero to its upper part, it's safe build zeros.
-def : Pat<(v8i32 immAllZerosV), (SUBREG_TO_REG (i32 0), (AVX_SET0PI), sub_xmm)>;
+def : Pat<(v8i32 immAllZerosV), (SUBREG_TO_REG (i32 0), (V_SET0), sub_xmm)>;
def : Pat<(bc_v8i32 (v8f32 immAllZerosV)),
- (SUBREG_TO_REG (i32 0), (AVX_SET0PI), sub_xmm)>;
+ (SUBREG_TO_REG (i32 0), (V_SET0), sub_xmm)>;
-def : Pat<(v4i64 immAllZerosV), (SUBREG_TO_REG (i64 0), (AVX_SET0PI), sub_xmm)>;
+def : Pat<(v4i64 immAllZerosV), (SUBREG_TO_REG (i64 0), (V_SET0), sub_xmm)>;
def : Pat<(bc_v4i64 (v8f32 immAllZerosV)),
- (SUBREG_TO_REG (i64 0), (AVX_SET0PI), sub_xmm)>;
+ (SUBREG_TO_REG (i64 0), (V_SET0), sub_xmm)>;
// We set canFoldAsLoad because this can be converted to a constant-pool
// load of an all-ones value if folding it would be beneficial.
// Move scalar to XMM zero-extended, zeroing a VR128 then do a
// MOVSS to the lower bits.
def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
- (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
+ (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
- (MOVSSrr (v4f32 (V_SET0PS)),
+ (MOVSSrr (v4f32 (V_SET0)),
(f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
- (MOVSSrr (v4i32 (V_SET0PI)),
+ (MOVSSrr (v4i32 (V_SET0)),
(EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
}
// Move scalar to XMM zero-extended, zeroing a VR128 then do a
// MOVSD to the lower bits.
def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
- (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
+ (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
}
let AddedComplexity = 20 in {
// Move scalar to XMM zero-extended, zeroing a VR128 then do a
// MOVS{S,D} to the lower bits.
def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
- (VMOVSSrr (v4f32 (AVX_SET0PS)), FR32:$src)>;
+ (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
- (VMOVSSrr (v4f32 (AVX_SET0PS)),
+ (VMOVSSrr (v4f32 (V_SET0)),
(f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
- (VMOVSSrr (v4i32 (AVX_SET0PI)),
+ (VMOVSSrr (v4i32 (V_SET0)),
(EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
- (VMOVSDrr (v2f64 (AVX_SET0PS)), FR64:$src)>;
+ (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
}
let AddedComplexity = 20 in {
def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
(v4f32 (scalar_to_vector FR32:$src)), (i32 0)))),
(SUBREG_TO_REG (i32 0),
- (v4f32 (VMOVSSrr (v4f32 (AVX_SET0PS)), FR32:$src)),
+ (v4f32 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)),
sub_xmm)>;
def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
(v2f64 (scalar_to_vector FR64:$src)), (i32 0)))),
(SUBREG_TO_REG (i64 0),
- (v2f64 (VMOVSDrr (v2f64 (AVX_SET0PS)), FR64:$src)),
+ (v2f64 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)),
sub_xmm)>;
// Extract and store.
projects / oota-llvm.git / commitdiff