[SPARC] Prefer RDPC over CALL to implement GETPCX for 64-bit target #77196

koachan · 2024-01-06T12:47:12Z

On 64-bit target, prefer usng RDPC over CALL to get the value of %pc.
This is faster on modern processors (Niagara T1 and newer) and avoids polluting
the processor's predictor state.

The old behavior of using a fake CALL is still done when tuning for classic
UltraSPARC processors, since RDPC is much slower there.

A quick pgbench test on a SPARC T4 shows about 2% speedup on SELECT loads,
and about 7% speedup on INSERT/UPDATE loads.

Created using spr 1.3.4

llvmbot · 2024-01-06T12:47:38Z

@llvm/pr-subscribers-backend-sparc

Author: Koakuma (koachan)

Changes

On 64-bit target, prefer usng RDPC over CALL to get the value of %pc.
This is faster on modern processors (Niagara T1 and newer) and avoids polluting
the processor's predictor state.

The old behavior of using a fake CALL is still done when tuning for classic
UltraSPARC processors, since RDPC is much slower there.

A quick pgbench test on a SPARC T4 shows about 2% speedup on SELECT loads,
and about 7% speedup on INSERT/UPDATE loads.

Full diff: https://github.com/llvm/llvm-project/pull/77196.diff

3 Files Affected:

(modified) llvm/lib/Target/Sparc/Sparc.td (+14-4)
(modified) llvm/lib/Target/Sparc/SparcAsmPrinter.cpp (+19-2)
(added) llvm/test/CodeGen/SPARC/tune-getpcx.ll (+18)

diff --git a/llvm/lib/Target/Sparc/Sparc.td b/llvm/lib/Target/Sparc/Sparc.td
index 1a71cfed3128f0..7b103395652433 100644
--- a/llvm/lib/Target/Sparc/Sparc.td
+++ b/llvm/lib/Target/Sparc/Sparc.td
@@ -62,6 +62,13 @@ def UsePopc : SubtargetFeature<"popc", "UsePopc", "true",
 def FeatureSoftFloat : SubtargetFeature<"soft-float", "UseSoftFloat", "true",
                               "Use software emulation for floating point">;
 
+//===----------------------------------------------------------------------===//
+// SPARC Subtarget tuning features.
+//
+
+def TuneSlowRDPC : SubtargetFeature<"slow-rdpc", "HasSlowRDPC", "true",
+                                    "rd %pc, %XX is slow", [FeatureV9]>;
+
 //==== Features added predmoninantly for LEON subtarget support
 include "LeonFeatures.td"
 
@@ -89,8 +96,9 @@ def SparcAsmParserVariant : AsmParserVariant {
 // SPARC processors supported.
 //===----------------------------------------------------------------------===//
 
-class Proc<string Name, list<SubtargetFeature> Features>
- : Processor<Name, NoItineraries, Features>;
+class Proc<string Name, list<SubtargetFeature> Features,
+           list<SubtargetFeature> TuneFeatures = []>
+ : Processor<Name, NoItineraries, Features, TuneFeatures>;
 
 def : Proc<"generic",         []>;
 def : Proc<"v7",              [FeatureSoftMulDiv, FeatureNoFSMULD]>;
@@ -118,9 +126,11 @@ def : Proc<"ma2480",          [FeatureLeon, LeonCASA]>;
 def : Proc<"ma2485",          [FeatureLeon, LeonCASA]>;
 def : Proc<"ma2x8x",          [FeatureLeon, LeonCASA]>;
 def : Proc<"v9",              [FeatureV9]>;
-def : Proc<"ultrasparc",      [FeatureV9, FeatureV8Deprecated, FeatureVIS]>;
+def : Proc<"ultrasparc",      [FeatureV9, FeatureV8Deprecated, FeatureVIS],
+                              [TuneSlowRDPC]>;
 def : Proc<"ultrasparc3",     [FeatureV9, FeatureV8Deprecated, FeatureVIS,
-                               FeatureVIS2]>;
+                               FeatureVIS2],
+                              [TuneSlowRDPC]>;
 def : Proc<"niagara",         [FeatureV9, FeatureV8Deprecated, FeatureVIS,
                                FeatureVIS2]>;
 def : Proc<"niagara2",        [FeatureV9, FeatureV8Deprecated, UsePopc,
diff --git a/llvm/lib/Target/Sparc/SparcAsmPrinter.cpp b/llvm/lib/Target/Sparc/SparcAsmPrinter.cpp
index cca624e0926796..97abf10b18540d 100644
--- a/llvm/lib/Target/Sparc/SparcAsmPrinter.cpp
+++ b/llvm/lib/Target/Sparc/SparcAsmPrinter.cpp
@@ -13,6 +13,7 @@
 
 #include "MCTargetDesc/SparcInstPrinter.h"
 #include "MCTargetDesc/SparcMCExpr.h"
+#include "MCTargetDesc/SparcMCTargetDesc.h"
 #include "MCTargetDesc/SparcTargetStreamer.h"
 #include "Sparc.h"
 #include "SparcInstrInfo.h"
@@ -111,6 +112,15 @@ static void EmitCall(MCStreamer &OutStreamer,
   OutStreamer.emitInstruction(CallInst, STI);
 }
 
+static void EmitRDPC(MCStreamer &OutStreamer, MCOperand &RD,
+                     const MCSubtargetInfo &STI) {
+  MCInst RDPCInst;
+  RDPCInst.setOpcode(SP::RDASR);
+  RDPCInst.addOperand(RD);
+  RDPCInst.addOperand(MCOperand::createReg(SP::ASR5));
+  OutStreamer.emitInstruction(RDPCInst, STI);
+}
+
 static void EmitSETHI(MCStreamer &OutStreamer,
                       MCOperand &Imm, MCOperand &RD,
                       const MCSubtargetInfo &STI)
@@ -234,8 +244,15 @@ void SparcAsmPrinter::LowerGETPCXAndEmitMCInsts(const MachineInstr *MI,
   //   add <MO>, %o7, <MO>
 
   OutStreamer->emitLabel(StartLabel);
-  MCOperand Callee =  createPCXCallOP(EndLabel, OutContext);
-  EmitCall(*OutStreamer, Callee, STI);
+  if (!STI.getTargetTriple().isSPARC64() ||
+      STI.hasFeature(Sparc::TuneSlowRDPC)) {
+    MCOperand Callee = createPCXCallOP(EndLabel, OutContext);
+    EmitCall(*OutStreamer, Callee, STI);
+  } else {
+    // TODO make it possible to store PC in other registers
+    // so that leaf function optimization becomes possible.
+    EmitRDPC(*OutStreamer, RegO7, STI);
+  }
   OutStreamer->emitLabel(SethiLabel);
   MCOperand hiImm = createPCXRelExprOp(SparcMCExpr::VK_Sparc_PC22,
                                        GOTLabel, StartLabel, SethiLabel,
diff --git a/llvm/test/CodeGen/SPARC/tune-getpcx.ll b/llvm/test/CodeGen/SPARC/tune-getpcx.ll
new file mode 100644
index 00000000000000..7454fea0e38d57
--- /dev/null
+++ b/llvm/test/CodeGen/SPARC/tune-getpcx.ll
@@ -0,0 +1,18 @@
+; RUN: llc < %s -relocation-model=pic -mtriple=sparc | FileCheck --check-prefix=CALL %s
+; RUN: llc < %s -relocation-model=pic -mtriple=sparcv9 -mcpu=ultrasparc | FileCheck --check-prefix=CALL %s
+; RUN: llc < %s -relocation-model=pic -mtriple=sparcv9 | FileCheck --check-prefix=RDPC %s
+
+;; SPARC32 and SPARC64 for classic UltraSPARCs implement GETPCX
+;; with a fake `call`.
+;; All other SPARC64 targets implement it with `rd %pc, %o7`.
+
+@value = external global i32
+
+; CALL: call
+; CALL-NOT: rd %pc
+; RDPC: rd %pc
+; RDPC-not: call
+define i32 @test() {
+  %1 = load i32, i32* @value
+  ret i32 %1
+}

llvm/test/CodeGen/SPARC/tune-getpcx.ll

llvm/lib/Target/Sparc/SparcAsmPrinter.cpp

s-barannikov · 2024-01-08T17:38:48Z

llvm/lib/Target/Sparc/Sparc.td

@@ -118,9 +126,11 @@ def : Proc<"ma2480",          [FeatureLeon, LeonCASA]>;
 def : Proc<"ma2485",          [FeatureLeon, LeonCASA]>;
 def : Proc<"ma2x8x",          [FeatureLeon, LeonCASA]>;
 def : Proc<"v9",              [FeatureV9]>;
-def : Proc<"ultrasparc",      [FeatureV9, FeatureV8Deprecated, FeatureVIS]>;
+def : Proc<"ultrasparc",      [FeatureV9, FeatureV8Deprecated, FeatureVIS],
+                              [TuneSlowRDPC]>;


Why is it in TuneFeatures? Sparc doesn't seem to support -mtune.

-mtune enablement is at PR #77195.
clang already accepts, recognises, and passes the flag on to the backend, it's just the backend haven't made any use of the provided info yet.

I see. IIUC -mtune is translated into -tune-cpu llc option. If so, the test should use this option.

On 64-bit target, prefer usng RDPC over CALL to get the value of %pc. This is faster on modern processors (Niagara T1 and newer) and avoids polluting the processor's predictor state. The old behavior of using a fake CALL is still done when tuning for classic UltraSPARC processors, since RDPC is much slower there. A quick pgbench test on a SPARC T4 shows about 2% speedup on SELECT loads, and about 7% speedup on INSERT/UPDATE loads. Pull Request: llvm#77196

Created using spr 1.3.4

llvm/test/CodeGen/SPARC/getpcx-call.ll

Created using spr 1.3.4

llvm/test/CodeGen/SPARC/getpcx-call.ll

Created using spr 1.3.4

brad0 · 2024-01-14T22:48:40Z

I intentionally restored the branch as I only noticed after the merge that the target branch was set incorrectly. It wasn't merged into the LLVM repo.

Is there a way of re-targeting the PR or make a new PR?

koachan · 2024-01-15T01:52:13Z

Lemme see if I can do it
I used spr to make the PRs, it should still be mergeable from there, I think?

koachan · 2024-01-15T13:55:28Z

Uh oh, seems like I couldn't merge it with the tool either...
Guess I'll open a new PR to merge this, would that be okay? @brad0 @s-barannikov

s-barannikov · 2024-01-15T14:41:37Z

Yes, sure

brad0 · 2024-01-15T21:38:32Z

Uh oh, seems like I couldn't merge it with the tool either... Guess I'll open a new PR to merge this, would that be okay? @brad0 @s-barannikov

Ya, sure. Whatever path is easier and quickest for you.

koachan · 2024-01-16T14:03:18Z

Okay, new PR is at #78280.

[𝘀𝗽𝗿] initial version

fe29e4f

Created using spr 1.3.4

llvmbot added the backend:Sparc label Jan 6, 2024

koachan requested review from jrtc27, brad0, rorth and s-barannikov January 7, 2024 00:45

s-barannikov reviewed Jan 8, 2024

View reviewed changes

Implement suggestions & clarify comment on RDPC emission

0c85312

Created using spr 1.3.4

s-barannikov reviewed Jan 13, 2024

View reviewed changes

llvm/test/CodeGen/SPARC/getpcx-call.ll Outdated Show resolved Hide resolved

Add nounwind to test cases

f483082

Created using spr 1.3.4

s-barannikov reviewed Jan 14, 2024

View reviewed changes

llvm/test/CodeGen/SPARC/getpcx-call.ll Outdated Show resolved Hide resolved

s-barannikov approved these changes Jan 14, 2024

View reviewed changes

Fixup load pointer type in testcase

811e1de

Created using spr 1.3.4

brad0 merged commit 63f9829 into users/koachan/main.sparc-prefer-rdpc-over-call-to-implement-getpcx-for-64-bit-target Jan 14, 2024

brad0 deleted the users/koachan/sparc-prefer-rdpc-over-call-to-implement-getpcx-for-64-bit-target branch January 14, 2024 21:28

brad0 restored the users/koachan/sparc-prefer-rdpc-over-call-to-implement-getpcx-for-64-bit-target branch January 14, 2024 21:33

s-barannikov deleted the users/koachan/sparc-prefer-rdpc-over-call-to-implement-getpcx-for-64-bit-target branch January 14, 2024 21:42

brad0 restored the users/koachan/sparc-prefer-rdpc-over-call-to-implement-getpcx-for-64-bit-target branch January 14, 2024 22:48

koachan mentioned this pull request Jan 16, 2024

[SPARC] Prefer RDPC over CALL to implement GETPCX for 64-bit target #78280

Merged

brad0 deleted the users/koachan/sparc-prefer-rdpc-over-call-to-implement-getpcx-for-64-bit-target branch January 28, 2025 20:08

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[SPARC] Prefer RDPC over CALL to implement GETPCX for 64-bit target #77196

[SPARC] Prefer RDPC over CALL to implement GETPCX for 64-bit target #77196

Uh oh!

koachan commented Jan 6, 2024

Uh oh!

llvmbot commented Jan 6, 2024

Uh oh!

Uh oh!

Uh oh!

Uh oh!

s-barannikov Jan 8, 2024

Uh oh!

koachan Jan 8, 2024

Uh oh!

s-barannikov Jan 8, 2024 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

brad0 commented Jan 14, 2024 •

edited

Loading

Uh oh!

koachan commented Jan 15, 2024

Uh oh!

koachan commented Jan 15, 2024

Uh oh!

s-barannikov commented Jan 15, 2024

Uh oh!

brad0 commented Jan 15, 2024

Uh oh!

koachan commented Jan 16, 2024

Uh oh!

Uh oh!

[SPARC] Prefer RDPC over CALL to implement GETPCX for 64-bit target #77196

[SPARC] Prefer RDPC over CALL to implement GETPCX for 64-bit target #77196

Uh oh!

Conversation

koachan commented Jan 6, 2024

Uh oh!

llvmbot commented Jan 6, 2024

Uh oh!

Uh oh!

Uh oh!

Uh oh!

s-barannikov Jan 8, 2024

Choose a reason for hiding this comment

Uh oh!

koachan Jan 8, 2024

Choose a reason for hiding this comment

Uh oh!

s-barannikov Jan 8, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

brad0 commented Jan 14, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

koachan commented Jan 15, 2024

Uh oh!

koachan commented Jan 15, 2024

Uh oh!

s-barannikov commented Jan 15, 2024

Uh oh!

brad0 commented Jan 15, 2024

Uh oh!

koachan commented Jan 16, 2024

Uh oh!

Uh oh!

s-barannikov Jan 8, 2024 •

edited

Loading

brad0 commented Jan 14, 2024 •

edited

Loading