Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

ARM64: Use stlur for volatile stores #91553

Merged
merged 8 commits into from
Sep 8, 2023
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions src/coreclr/jit/codegen.h
Original file line number Diff line number Diff line change
Expand Up @@ -1235,6 +1235,10 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
#else
instruction genGetInsForOper(genTreeOps oper, var_types type);
#endif
instruction genGetVolatileLdStIns(instruction currentIns,
regNumber targetReg,
GenTreeIndir* indir,
bool* needsBarrier);
bool genEmitOptimizedGCWriteBarrier(GCInfo::WriteBarrierForm writeBarrierForm, GenTree* addr, GenTree* data);
GenTree* getCallTarget(const GenTreeCall* call, CORINFO_METHOD_HANDLE* methHnd);
regNumber getCallIndirectionCellReg(GenTreeCall* call);
Expand Down
18 changes: 3 additions & 15 deletions src/coreclr/jit/codegenarm64.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -4207,22 +4207,10 @@ void CodeGen::genCodeForStoreInd(GenTreeStoreInd* tree)

if (tree->IsVolatile())
{
bool addrIsInReg = addr->isUsedFromReg();
bool addrIsAligned = ((tree->gtFlags & GTF_IND_UNALIGNED) == 0);
bool needsBarrier = true;
ins = genGetVolatileLdStIns(ins, dataReg, tree, &needsBarrier);

if ((ins == INS_strb) && addrIsInReg)
{
ins = INS_stlrb;
}
else if ((ins == INS_strh) && addrIsInReg && addrIsAligned)
{
ins = INS_stlrh;
}
else if ((ins == INS_str) && genIsValidIntReg(dataReg) && addrIsInReg && addrIsAligned)
{
ins = INS_stlr;
}
else
if (needsBarrier)
{
// issue a full memory barrier before a volatile StInd
// Note: We cannot issue store barrier ishst because it is a weaker barrier.
Expand Down
181 changes: 130 additions & 51 deletions src/coreclr/jit/codegenarmarch.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -1728,6 +1728,134 @@ void CodeGen::genCodeForIndexAddr(GenTreeIndexAddr* node)
genProduceReg(node);
}

//------------------------------------------------------------------------
// genGetVolatileLdStIns: Determine the most efficient instruction to perform a
// volatile load or store and whether an explicit barrier is required or not.
//
// Arguments:
// currentIns - the current instruction to perform load/store
// targetReg - the target register
// indir - the indirection node representing the volatile load/store
// needsBarrier - OUT parameter. Set to true if an explicit memory barrier is required.
//
// Return Value:
// instruction to perform the volatile load/store with.
//
instruction CodeGen::genGetVolatileLdStIns(instruction currentIns,
regNumber targetReg,
GenTreeIndir* indir,
bool* needsBarrier)
{
assert(indir->IsVolatile());

if (!genIsValidIntReg(targetReg))
{
// We don't have special instructions to perform volatile loads/stores for non-integer registers.
*needsBarrier = true;
return currentIns;
}

assert(!varTypeIsFloating(indir));
assert(!varTypeIsSIMD(indir));

#ifdef TARGET_ARM64
*needsBarrier = false;

if (indir->IsUnaligned() && (currentIns != INS_ldrb) && (currentIns != INS_strb))
{
// We have to use a normal load/store + explicit memory barrier
// to avoid unaligned access exceptions.
*needsBarrier = true;
return currentIns;
}

const bool addrIsInReg = indir->Addr()->isUsedFromReg();

// With RCPC2 (arm64 v8.4+) we can work with simple addressing modes like [reg + simm9]
const bool shouldUseRcpc2 = compiler->compOpportunisticallyDependsOn(InstructionSet_Rcpc2) && !addrIsInReg &&
indir->Addr()->OperIs(GT_LEA) && !indir->HasIndex() && (indir->Scale() == 1) &&
emitter::emitIns_valid_imm_for_unscaled_ldst_offset(indir->Offset());

if (shouldUseRcpc2)
{
assert(!addrIsInReg);
switch (currentIns)
{
// Loads

case INS_ldrb:
return INS_ldapurb;

case INS_ldrh:
return INS_ldapurh;

case INS_ldr:
return INS_ldapur;

// Stores

case INS_strb:
return INS_stlurb;

case INS_strh:
return INS_stlurh;

case INS_str:
return INS_stlur;

default:
*needsBarrier = true;
return currentIns;
}
}

// Only RCPC2 (arm64 v8.4+) allows us to deal with contained addresses.
// In other cases we'll have to emit a normal load/store + explicit memory barrier.
// It means that lower should generally avoid generating contained addresses for volatile loads/stores
// so we can use cheaper instructions.
if (!addrIsInReg)
{
*needsBarrier = true;
return currentIns;
}

// RCPC1 (arm64 v8.3+) offers a bit more relaxed memory ordering than ldar. Which is sufficient for
// .NET memory model's requirements, see https://github.com/dotnet/runtime/issues/67374
const bool hasRcpc1 = compiler->compOpportunisticallyDependsOn(InstructionSet_Rcpc);
switch (currentIns)
{
// Loads

case INS_ldrb:
return hasRcpc1 ? INS_ldaprb : INS_ldarb;

case INS_ldrh:
return hasRcpc1 ? INS_ldaprh : INS_ldarh;

case INS_ldr:
return hasRcpc1 ? INS_ldapr : INS_ldar;

// Stores

case INS_strb:
return INS_stlrb;

case INS_strh:
return INS_stlrh;

case INS_str:
return INS_stlr;

default:
*needsBarrier = true;
return currentIns;
}
#else // TARGET_ARM64
*needsBarrier = true;
return currentIns;
#endif // !TARGET_ARM64
}

//------------------------------------------------------------------------
// genCodeForIndir: Produce code for a GT_IND node.
//
Expand Down Expand Up @@ -1755,58 +1883,9 @@ void CodeGen::genCodeForIndir(GenTreeIndir* tree)

bool emitBarrier = false;

if ((tree->gtFlags & GTF_IND_VOLATILE) != 0)
if (tree->IsVolatile())
{
#ifdef TARGET_ARM64
bool addrIsInReg = tree->Addr()->isUsedFromReg();
bool addrIsAligned = ((tree->gtFlags & GTF_IND_UNALIGNED) == 0);

// On arm64-v8.3+ we can use ldap* instructions with acquire/release semantics to avoid
// full memory barriers if mixed with STLR
bool hasRcpc = compiler->compOpportunisticallyDependsOn(InstructionSet_Rcpc);

// On arm64-v8.4+ we can use ldapur* instructions with acquire/release semantics to
// avoid full memory barriers if address is contained and unscaled
bool hasRcpc2 = compiler->compOpportunisticallyDependsOn(InstructionSet_Rcpc2);

bool handledWithLdapur = false;
if (hasRcpc2 && !addrIsInReg && tree->Addr()->OperIs(GT_LEA) && !tree->HasIndex() && (tree->Scale() == 1) &&
emitter::emitIns_valid_imm_for_unscaled_ldst_offset(tree->Offset()))
{
if (ins == INS_ldrb)
{
ins = INS_ldapurb;
handledWithLdapur = true;
}
else if ((ins == INS_ldrh) && addrIsAligned)
{
ins = INS_ldapurh;
handledWithLdapur = true;
}
else if ((ins == INS_ldr) && addrIsAligned && genIsValidIntReg(targetReg))
{
ins = INS_ldapur;
handledWithLdapur = true;
}
}

if ((ins == INS_ldrb) && addrIsInReg)
{
ins = hasRcpc ? INS_ldaprb : INS_ldarb;
}
else if ((ins == INS_ldrh) && addrIsInReg && addrIsAligned)
{
ins = hasRcpc ? INS_ldaprh : INS_ldarh;
}
else if ((ins == INS_ldr) && addrIsInReg && addrIsAligned && genIsValidIntReg(targetReg))
{
ins = hasRcpc ? INS_ldapr : INS_ldar;
}
else if (!handledWithLdapur)
#endif // TARGET_ARM64
{
emitBarrier = true;
}
ins = genGetVolatileLdStIns(ins, targetReg, tree, &emitBarrier);
}

GetEmitter()->emitInsLoadStoreOp(ins, emitActualTypeSize(type), targetReg, tree);
Expand Down