summaryrefslogtreecommitdiffstats
path: root/contrib/llvm/lib/Target/AArch64/AArch64InstrInfo.td
diff options
context:
space:
mode:
Diffstat (limited to 'contrib/llvm/lib/Target/AArch64/AArch64InstrInfo.td')
-rw-r--r--contrib/llvm/lib/Target/AArch64/AArch64InstrInfo.td123
1 files changed, 73 insertions, 50 deletions
diff --git a/contrib/llvm/lib/Target/AArch64/AArch64InstrInfo.td b/contrib/llvm/lib/Target/AArch64/AArch64InstrInfo.td
index 37be5e4..d2cfc7d 100644
--- a/contrib/llvm/lib/Target/AArch64/AArch64InstrInfo.td
+++ b/contrib/llvm/lib/Target/AArch64/AArch64InstrInfo.td
@@ -70,12 +70,20 @@ def A64cmn : PatFrag<(ops node:$lhs, node:$rhs),
// made for a variable/address at ISelLowering.
// + The output of ISelLowering should be selectable (hence the Wrapper,
// rather than a bare target opcode)
-def SDTAArch64Wrapper : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
- SDTCisSameAs<1, 2>,
- SDTCisVT<3, i32>,
- SDTCisPtrTy<0>]>;
+def SDTAArch64WrapperLarge : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>,
+ SDTCisSameAs<0, 2>,
+ SDTCisSameAs<0, 3>,
+ SDTCisSameAs<0, 4>,
+ SDTCisPtrTy<0>]>;
-def A64WrapperSmall : SDNode<"AArch64ISD::WrapperSmall", SDTAArch64Wrapper>;
+def A64WrapperLarge :SDNode<"AArch64ISD::WrapperLarge", SDTAArch64WrapperLarge>;
+
+def SDTAArch64WrapperSmall : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
+ SDTCisSameAs<1, 2>,
+ SDTCisVT<3, i32>,
+ SDTCisPtrTy<0>]>;
+
+def A64WrapperSmall :SDNode<"AArch64ISD::WrapperSmall", SDTAArch64WrapperSmall>;
def SDTAArch64GOTLoad : SDTypeProfile<1, 1, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
@@ -159,49 +167,55 @@ let Defs = [XSP], Uses = [XSP] in {
// Atomic operation pseudo-instructions
//===----------------------------------------------------------------------===//
-let usesCustomInserter = 1 in {
-multiclass AtomicSizes<string opname> {
- def _I8 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
- [(set i32:$dst, (!cast<SDNode>(opname # "_8") i64:$ptr, i32:$incr))]>;
- def _I16 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
- [(set i32:$dst, (!cast<SDNode>(opname # "_16") i64:$ptr, i32:$incr))]>;
- def _I32 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
- [(set i32:$dst, (!cast<SDNode>(opname # "_32") i64:$ptr, i32:$incr))]>;
- def _I64 : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$incr),
- [(set i64:$dst, (!cast<SDNode>(opname # "_64") i64:$ptr, i64:$incr))]>;
-}
-}
-
-defm ATOMIC_LOAD_ADD : AtomicSizes<"atomic_load_add">;
-defm ATOMIC_LOAD_SUB : AtomicSizes<"atomic_load_sub">;
-defm ATOMIC_LOAD_AND : AtomicSizes<"atomic_load_and">;
-defm ATOMIC_LOAD_OR : AtomicSizes<"atomic_load_or">;
-defm ATOMIC_LOAD_XOR : AtomicSizes<"atomic_load_xor">;
-defm ATOMIC_LOAD_NAND : AtomicSizes<"atomic_load_nand">;
-defm ATOMIC_SWAP : AtomicSizes<"atomic_swap">;
+// These get selected from C++ code as a pretty much direct translation from the
+// generic DAG nodes. The one exception is the AtomicOrdering is added as an
+// operand so that the eventual lowering can make use of it and choose
+// acquire/release operations when required.
+
+let usesCustomInserter = 1, hasCtrlDep = 1, mayLoad = 1, mayStore = 1 in {
+multiclass AtomicSizes {
+ def _I8 : PseudoInst<(outs GPR32:$dst),
+ (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
+ def _I16 : PseudoInst<(outs GPR32:$dst),
+ (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
+ def _I32 : PseudoInst<(outs GPR32:$dst),
+ (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
+ def _I64 : PseudoInst<(outs GPR64:$dst),
+ (ins GPR64xsp:$ptr, GPR64:$incr, i32imm:$ordering), []>;
+}
+}
+
+defm ATOMIC_LOAD_ADD : AtomicSizes;
+defm ATOMIC_LOAD_SUB : AtomicSizes;
+defm ATOMIC_LOAD_AND : AtomicSizes;
+defm ATOMIC_LOAD_OR : AtomicSizes;
+defm ATOMIC_LOAD_XOR : AtomicSizes;
+defm ATOMIC_LOAD_NAND : AtomicSizes;
+defm ATOMIC_SWAP : AtomicSizes;
let Defs = [NZCV] in {
// These operations need a CMP to calculate the correct value
- defm ATOMIC_LOAD_MIN : AtomicSizes<"atomic_load_min">;
- defm ATOMIC_LOAD_MAX : AtomicSizes<"atomic_load_max">;
- defm ATOMIC_LOAD_UMIN : AtomicSizes<"atomic_load_umin">;
- defm ATOMIC_LOAD_UMAX : AtomicSizes<"atomic_load_umax">;
-}
-
-let usesCustomInserter = 1, Defs = [NZCV] in {
-def ATOMIC_CMP_SWAP_I8
- : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
- [(set i32:$dst, (atomic_cmp_swap_8 i64:$ptr, i32:$old, i32:$new))]>;
-def ATOMIC_CMP_SWAP_I16
- : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
- [(set i32:$dst, (atomic_cmp_swap_16 i64:$ptr, i32:$old, i32:$new))]>;
-def ATOMIC_CMP_SWAP_I32
- : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
- [(set i32:$dst, (atomic_cmp_swap_32 i64:$ptr, i32:$old, i32:$new))]>;
-def ATOMIC_CMP_SWAP_I64
- : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$old, GPR64:$new),
- [(set i64:$dst, (atomic_cmp_swap_64 i64:$ptr, i64:$old, i64:$new))]>;
+ defm ATOMIC_LOAD_MIN : AtomicSizes;
+ defm ATOMIC_LOAD_MAX : AtomicSizes;
+ defm ATOMIC_LOAD_UMIN : AtomicSizes;
+ defm ATOMIC_LOAD_UMAX : AtomicSizes;
+}
+
+class AtomicCmpSwap<RegisterClass GPRData>
+ : PseudoInst<(outs GPRData:$dst),
+ (ins GPR64xsp:$ptr, GPRData:$old, GPRData:$new,
+ i32imm:$ordering), []> {
+ let usesCustomInserter = 1;
+ let hasCtrlDep = 1;
+ let mayLoad = 1;
+ let mayStore = 1;
+ let Defs = [NZCV];
}
+def ATOMIC_CMP_SWAP_I8 : AtomicCmpSwap<GPR32>;
+def ATOMIC_CMP_SWAP_I16 : AtomicCmpSwap<GPR32>;
+def ATOMIC_CMP_SWAP_I32 : AtomicCmpSwap<GPR32>;
+def ATOMIC_CMP_SWAP_I64 : AtomicCmpSwap<GPR64>;
+
//===----------------------------------------------------------------------===//
// Add-subtract (extended register) instructions
//===----------------------------------------------------------------------===//
@@ -2579,7 +2593,8 @@ defm LDAR : A64I_LRex<"ldar", 0b101>;
class acquiring_load<PatFrag base>
: PatFrag<(ops node:$ptr), (base node:$ptr), [{
- return cast<AtomicSDNode>(N)->getOrdering() == Acquire;
+ AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
+ return Ordering == Acquire || Ordering == SequentiallyConsistent;
}]>;
def atomic_load_acquire_8 : acquiring_load<atomic_load_8>;
@@ -2610,7 +2625,8 @@ class A64I_SLexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
class releasing_store<PatFrag base>
: PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
- return cast<AtomicSDNode>(N)->getOrdering() == Release;
+ AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
+ return Ordering == Release || Ordering == SequentiallyConsistent;
}]>;
def atomic_store_release_8 : releasing_store<atomic_store_8>;
@@ -3863,7 +3879,7 @@ multiclass movw_operands<string prefix, string instname, int width> {
let DiagnosticType = "MOVWUImm16";
}
- def _imm : Operand<i32> {
+ def _imm : Operand<i64> {
let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_imm_asmoperand");
let PrintMethod = "printMoveWideImmOperand";
let EncoderMethod = "getMoveWideImmOpValue";
@@ -3934,7 +3950,7 @@ multiclass movalias_operand<string prefix, string basename,
# "A64Imms::" # immpredicate # ">";
}
- def _movimm : Operand<i32> {
+ def _movimm : Operand<i64> {
let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
@@ -3958,6 +3974,15 @@ def : movalias<MOVZxii, GPR64, movz64_movimm>;
def : movalias<MOVNwii, GPR32, movn32_movimm>;
def : movalias<MOVNxii, GPR64, movn64_movimm>;
+def movw_addressref : ComplexPattern<i64, 2, "SelectMOVWAddressRef">;
+
+def : Pat<(A64WrapperLarge movw_addressref:$G3, movw_addressref:$G2,
+ movw_addressref:$G1, movw_addressref:$G0),
+ (MOVKxii (MOVKxii (MOVKxii (MOVZxii movw_addressref:$G3),
+ movw_addressref:$G2),
+ movw_addressref:$G1),
+ movw_addressref:$G0)>;
+
//===----------------------------------------------------------------------===//
// PC-relative addressing instructions
//===----------------------------------------------------------------------===//
@@ -4454,8 +4479,6 @@ def : ADRP_ADD<A64WrapperSmall, tjumptable>;
// GOT access patterns
//===----------------------------------------------------------------------===//
-// FIXME: Wibble
-
class GOTLoadSmall<SDNode addrfrag>
: Pat<(A64GOTLoad (A64WrapperSmall addrfrag:$Hi, addrfrag:$Lo12, 8)),
(LS64_LDR (ADRPxi addrfrag:$Hi), addrfrag:$Lo12)>;
OpenPOWER on IntegriCloud