[llvm] [llvm-mca] Add command line option `-use-load-latency` (PR #94566)
Chinmay Deshpande via llvm-commits
llvm-commits at lists.llvm.org
Wed Jun 5 21:17:38 PDT 2024
https://github.com/chinmaydd created https://github.com/llvm/llvm-project/pull/94566
Latency for load operations could be modeled better by using the `LoadLatency` field in the provided scheduling model for the target architecture. This PR adds a new command line flag `use-load-latency` that enables this behavior. The flag is off by default.
This attempts a fix for the comment here - [LSUnit.h#370](https://github.com/llvm/llvm-project/blob/main/llvm/include/llvm/MCA/HardwareUnits/LSUnit.h#L370). I would be happy to know if there is a better way to approach this.
>From 553c8b2ed8c86657f296acfb4e7faa184e407bb4 Mon Sep 17 00:00:00 2001
From: Chinmay Deshpande <cddeshpa at uci.edu>
Date: Wed, 5 Jun 2024 21:05:07 -0700
Subject: [PATCH] [llvm-mca] Add command line option `-use-load-latency`
Latency for load operations could be modeled better by
using the `LoadLatency` field in the provided scheduling
model for the target architecture.
---
llvm/include/llvm/MCA/InstrBuilder.h | 4 +-
llvm/lib/MCA/InstrBuilder.cpp | 18 ++++--
.../tools/llvm-mca/X86/use-load-latency.s | 58 +++++++++++++++++++
llvm/tools/llvm-mca/llvm-mca.cpp | 8 ++-
llvm/unittests/tools/llvm-mca/MCATestBase.cpp | 3 +-
.../tools/llvm-mca/X86/TestIncrementalMCA.cpp | 6 +-
6 files changed, 87 insertions(+), 10 deletions(-)
create mode 100644 llvm/test/tools/llvm-mca/X86/use-load-latency.s
diff --git a/llvm/include/llvm/MCA/InstrBuilder.h b/llvm/include/llvm/MCA/InstrBuilder.h
index 00c7942e4fa16..6880a213d48b5 100644
--- a/llvm/include/llvm/MCA/InstrBuilder.h
+++ b/llvm/include/llvm/MCA/InstrBuilder.h
@@ -79,6 +79,7 @@ class InstrBuilder {
bool FirstCallInst;
bool FirstReturnInst;
unsigned CallLatency;
+ bool UseLoadLatency;
using InstRecycleCallback = std::function<Instruction *(const InstrDesc &)>;
InstRecycleCallback InstRecycleCB;
@@ -99,7 +100,8 @@ class InstrBuilder {
public:
InstrBuilder(const MCSubtargetInfo &STI, const MCInstrInfo &MCII,
const MCRegisterInfo &RI, const MCInstrAnalysis *IA,
- const InstrumentManager &IM, unsigned CallLatency);
+ const InstrumentManager &IM, unsigned CallLatency,
+ bool UseLoadLatency);
void clear() {
Descriptors.clear();
diff --git a/llvm/lib/MCA/InstrBuilder.cpp b/llvm/lib/MCA/InstrBuilder.cpp
index d5cbdc5de0b84..db4eda40de9d0 100644
--- a/llvm/lib/MCA/InstrBuilder.cpp
+++ b/llvm/lib/MCA/InstrBuilder.cpp
@@ -31,9 +31,10 @@ InstrBuilder::InstrBuilder(const llvm::MCSubtargetInfo &sti,
const llvm::MCInstrInfo &mcii,
const llvm::MCRegisterInfo &mri,
const llvm::MCInstrAnalysis *mcia,
- const mca::InstrumentManager &im, unsigned cl)
+ const mca::InstrumentManager &im, unsigned cl,
+ bool ull)
: STI(sti), MCII(mcii), MRI(mri), MCIA(mcia), IM(im), FirstCallInst(true),
- FirstReturnInst(true), CallLatency(cl) {
+ FirstReturnInst(true), CallLatency(cl), UseLoadLatency(ull) {
const MCSchedModel &SM = STI.getSchedModel();
ProcResourceMasks.resize(SM.getNumProcResourceKinds());
computeProcResourceMasks(STI.getSchedModel(), ProcResourceMasks);
@@ -220,8 +221,8 @@ static void initializeUsedResources(InstrDesc &ID,
static void computeMaxLatency(InstrDesc &ID, const MCInstrDesc &MCDesc,
const MCSchedClassDesc &SCDesc,
- const MCSubtargetInfo &STI,
- unsigned CallLatency) {
+ const MCSubtargetInfo &STI, unsigned CallLatency,
+ bool UseLoadLatency) {
if (MCDesc.isCall()) {
// We cannot estimate how long this call will take.
// Artificially set an arbitrarily high latency.
@@ -230,6 +231,13 @@ static void computeMaxLatency(InstrDesc &ID, const MCInstrDesc &MCDesc,
}
int Latency = MCSchedModel::computeInstrLatency(STI, SCDesc);
+
+ // If `UseLoadLatency` is set, we use the value in `MCSchedModel::LoadLatency`
+ // for load instructions.
+ if (MCDesc.mayLoad() && UseLoadLatency) {
+ const auto &SM = STI.getSchedModel();
+ Latency = std::max(int(SM.LoadLatency), Latency);
+ }
// If latency is unknown, then conservatively assume the MaxLatency set for
// calls.
ID.MaxLatency = Latency < 0 ? CallLatency : static_cast<unsigned>(Latency);
@@ -582,7 +590,7 @@ InstrBuilder::createInstrDescImpl(const MCInst &MCI,
}
initializeUsedResources(*ID, SCDesc, STI, ProcResourceMasks);
- computeMaxLatency(*ID, MCDesc, SCDesc, STI, CallLatency);
+ computeMaxLatency(*ID, MCDesc, SCDesc, STI, CallLatency, UseLoadLatency);
if (Error Err = verifyOperands(MCDesc, MCI))
return std::move(Err);
diff --git a/llvm/test/tools/llvm-mca/X86/use-load-latency.s b/llvm/test/tools/llvm-mca/X86/use-load-latency.s
new file mode 100644
index 0000000000000..54c37c5c8699f
--- /dev/null
+++ b/llvm/test/tools/llvm-mca/X86/use-load-latency.s
@@ -0,0 +1,58 @@
+# NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
+# RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -iterations=1 %s | FileCheck --check-prefixes=ALL,DEFAULT %s
+# RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -use-load-latency -iterations=1 %s | FileCheck --check-prefixes=ALL,CUSTOM %s
+
+movq (%rsp), %rdi
+
+# ALL: Iterations: 1
+# ALL-NEXT: Instructions: 1
+
+# CUSTOM-NEXT: Total Cycles: 8
+# DEFAULT-NEXT: Total Cycles: 6
+
+# ALL-NEXT: Total uOps: 1
+
+# ALL: Dispatch Width: 2
+
+# CUSTOM-NEXT: uOps Per Cycle: 0.13
+# CUSTOM-NEXT: IPC: 0.13
+
+# DEFAULT-NEXT: uOps Per Cycle: 0.17
+# DEFAULT-NEXT: IPC: 0.17
+
+# ALL-NEXT: Block RThroughput: 1.0
+
+# ALL: Instruction Info:
+# ALL-NEXT: [1]: #uOps
+# ALL-NEXT: [2]: Latency
+# ALL-NEXT: [3]: RThroughput
+# ALL-NEXT: [4]: MayLoad
+# ALL-NEXT: [5]: MayStore
+# ALL-NEXT: [6]: HasSideEffects (U)
+
+# ALL: [1] [2] [3] [4] [5] [6] Instructions:
+# ALL-NEXT: 1 3 1.00 * movq (%rsp), %rdi
+
+# ALL: Resources:
+# ALL-NEXT: [0] - JALU0
+# ALL-NEXT: [1] - JALU1
+# ALL-NEXT: [2] - JDiv
+# ALL-NEXT: [3] - JFPA
+# ALL-NEXT: [4] - JFPM
+# ALL-NEXT: [5] - JFPU0
+# ALL-NEXT: [6] - JFPU1
+# ALL-NEXT: [7] - JLAGU
+# ALL-NEXT: [8] - JMul
+# ALL-NEXT: [9] - JSAGU
+# ALL-NEXT: [10] - JSTC
+# ALL-NEXT: [11] - JVALU0
+# ALL-NEXT: [12] - JVALU1
+# ALL-NEXT: [13] - JVIMUL
+
+# ALL: Resource pressure per iteration:
+# ALL-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13]
+# ALL-NEXT: - - - - - - - 1.00 - - - - - -
+
+# ALL: Resource pressure by instruction:
+# ALL-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] Instructions:
+# ALL-NEXT: - - - - - - - 1.00 - - - - - - movq (%rsp), %rdi
diff --git a/llvm/tools/llvm-mca/llvm-mca.cpp b/llvm/tools/llvm-mca/llvm-mca.cpp
index cc5d4f5fa05de..927b2b97c4d68 100644
--- a/llvm/tools/llvm-mca/llvm-mca.cpp
+++ b/llvm/tools/llvm-mca/llvm-mca.cpp
@@ -140,6 +140,11 @@ static cl::opt<unsigned>
cl::desc("Number of cycles to assume for a call instruction"),
cl::cat(ToolOptions), cl::init(100U));
+static cl::opt<bool> UseLoadLatency(
+ "use-load-latency", cl::Hidden,
+ cl::desc("Use target specific latency for load instructions"),
+ cl::cat(ToolOptions), cl::init(false));
+
enum class SkipType { NONE, LACK_SCHED, PARSE_FAILURE, ANY_FAILURE };
static cl::opt<enum SkipType> SkipUnsupportedInstructions(
@@ -573,7 +578,8 @@ int main(int argc, char **argv) {
}
// Create an instruction builder.
- mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get(), *IM, CallLatency);
+ mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get(), *IM, CallLatency,
+ UseLoadLatency);
// Create a context to control ownership of the pipeline hardware.
mca::Context MCA(*MRI, *STI);
diff --git a/llvm/unittests/tools/llvm-mca/MCATestBase.cpp b/llvm/unittests/tools/llvm-mca/MCATestBase.cpp
index 4a39f5e663f23..515fbe5bd6b69 100644
--- a/llvm/unittests/tools/llvm-mca/MCATestBase.cpp
+++ b/llvm/unittests/tools/llvm-mca/MCATestBase.cpp
@@ -66,7 +66,8 @@ Error MCATestBase::runBaselineMCA(json::Object &Result, ArrayRef<MCInst> Insts,
// Default InstrumentManager
auto IM = std::make_unique<mca::InstrumentManager>(*STI, *MCII);
- mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get(), *IM, /*CallLatency=*/100);
+ mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get(), *IM, /*CallLatency=*/100,
+ /*UseLoadLatency*/ false);
const SmallVector<mca::Instrument *> Instruments;
SmallVector<std::unique_ptr<mca::Instruction>> LoweredInsts;
diff --git a/llvm/unittests/tools/llvm-mca/X86/TestIncrementalMCA.cpp b/llvm/unittests/tools/llvm-mca/X86/TestIncrementalMCA.cpp
index ac35dce522ae1..80a2f49ca4aa4 100644
--- a/llvm/unittests/tools/llvm-mca/X86/TestIncrementalMCA.cpp
+++ b/llvm/unittests/tools/llvm-mca/X86/TestIncrementalMCA.cpp
@@ -33,7 +33,8 @@ TEST_F(X86TestBase, TestResumablePipeline) {
P->addEventListener(SV.get());
auto IM = std::make_unique<mca::InstrumentManager>(*STI, *MCII);
- mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get(), *IM, /*CallLatency=*/100);
+ mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get(), *IM, /*CallLatency=*/100,
+ /*UseLoadLatency*/ false);
const SmallVector<mca::Instrument *> Instruments;
// Tile size = 7
@@ -124,7 +125,8 @@ TEST_F(X86TestBase, TestInstructionRecycling) {
// Default InstrumentManager
auto IM = std::make_unique<mca::InstrumentManager>(*STI, *MCII);
- mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get(), *IM, /*CallLatency=*/100);
+ mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get(), *IM, /*CallLatency=*/100,
+ /*UseLoadLatency*/ false);
IB.setInstRecycleCallback(GetRecycledInst);
const SmallVector<mca::Instrument *> Instruments;
More information about the llvm-commits
mailing list