[ARM] Consider denormal mode in ARMSubtarget

[paperchalice]

Factor out from #151275.
Add denormal mode to subtarget.

[llvmbot]

@llvm/pr-subscribers-backend-arm

Author: None (paperchalice)

Changes

Factor out from #151275.
Add denormal mode to subtarget.

---

Full diff: https://github.com/llvm/llvm-project/pull/160456.diff

3 Files Affected:

• (modified) llvm/lib/Target/ARM/ARMSubtarget.cpp (+5-7)
• (modified) llvm/lib/Target/ARM/ARMSubtarget.h (+6)
• (modified) llvm/lib/Target/ARM/ARMTargetMachine.cpp (+5-1)

diff --git a/llvm/lib/Target/ARM/ARMSubtarget.cpp b/llvm/lib/Target/ARM/ARMSubtarget.cpp
index 9f600e0c685ab..ed7a1ef21ee84 100644
--- a/llvm/lib/Target/ARM/ARMSubtarget.cpp
+++ b/llvm/lib/Target/ARM/ARMSubtarget.cpp
@@ -88,18 +88,16 @@ ARMFrameLowering *ARMSubtarget::initializeFrameLowering(StringRef CPU,
 ARMSubtarget::ARMSubtarget(const Triple &TT, const std::string &CPU,
                            const std::string &FS,
                            const ARMBaseTargetMachine &TM, bool IsLittle,
-                           bool MinSize)
+                           bool MinSize, DenormalMode DM)
     : ARMGenSubtargetInfo(TT, CPU, /*TuneCPU*/ CPU, FS),
       UseMulOps(UseFusedMulOps), CPUString(CPU), OptMinSize(MinSize),
-      IsLittle(IsLittle), TargetTriple(TT), Options(TM.Options), TM(TM),
+      IsLittle(IsLittle), DM(DM), TargetTriple(TT), Options(TM.Options), TM(TM),
       FrameLowering(initializeFrameLowering(CPU, FS)),
       // At this point initializeSubtargetDependencies has been called so
       // we can query directly.
-      InstrInfo(isThumb1Only()
-                    ? (ARMBaseInstrInfo *)new Thumb1InstrInfo(*this)
-                    : !isThumb()
-                          ? (ARMBaseInstrInfo *)new ARMInstrInfo(*this)
-                          : (ARMBaseInstrInfo *)new Thumb2InstrInfo(*this)),
+      InstrInfo(isThumb1Only() ? (ARMBaseInstrInfo *)new Thumb1InstrInfo(*this)
+                : !isThumb()   ? (ARMBaseInstrInfo *)new ARMInstrInfo(*this)
+                             : (ARMBaseInstrInfo *)new Thumb2InstrInfo(*this)),
       TLInfo(TM, *this) {
 
   CallLoweringInfo.reset(new ARMCallLowering(*getTargetLowering()));
diff --git a/llvm/lib/Target/ARM/ARMSubtarget.h b/llvm/lib/Target/ARM/ARMSubtarget.h
index 637eb4560e0f1..d848da0f1dbd5 100644
--- a/llvm/lib/Target/ARM/ARMSubtarget.h
+++ b/llvm/lib/Target/ARM/ARMSubtarget.h
@@ -183,6 +183,12 @@ class ARMSubtarget : public ARMGenSubtargetInfo {
   /// the function attribute.
   bool OptMinSize = false;
 
+  /// DM - Denormal mode
+  /// NEON and VFP RunFast mode are not IEEE 754 compliant,
+  /// use this field to determine whether to generate NEON/VFP
+  /// instructions in related function.
+  DenormalMode DM;
+
   /// IsLittle - The target is Little Endian
   bool IsLittle;
 
diff --git a/llvm/lib/Target/ARM/ARMTargetMachine.cpp b/llvm/lib/Target/ARM/ARMTargetMachine.cpp
index 131b9332e9ade..86740a92b32c5 100644
--- a/llvm/lib/Target/ARM/ARMTargetMachine.cpp
+++ b/llvm/lib/Target/ARM/ARMTargetMachine.cpp
@@ -229,6 +229,10 @@ ARMBaseTargetMachine::getSubtargetImpl(const Function &F) const {
   if (F.hasMinSize())
     Key += "+minsize";
 
+  DenormalMode DM = F.getDenormalModeRaw();
+  if (DM != DenormalMode::getIEEE())
+    Key += "denormal-fp-math=" + DM.str();
+
   auto &I = SubtargetMap[Key];
   if (!I) {
     // This needs to be done before we create a new subtarget since any
@@ -236,7 +240,7 @@ ARMBaseTargetMachine::getSubtargetImpl(const Function &F) const {
     // function that reside in TargetOptions.
     resetTargetOptions(F);
     I = std::make_unique<ARMSubtarget>(TargetTriple, CPU, FS, *this, isLittle,
-                                        F.hasMinSize());
+                                       F.hasMinSize(), DM);
 
     if (!I->isThumb() && !I->hasARMOps())
       F.getContext().emitError("Function '" + F.getName() + "' uses ARM "

[paperchalice]

Another possible method is adding a new IR module flag meta data for ARM with behavior 1 or 2.

[davemgreen]

Thanks for this. Can you pull in the change from this bit of code too? It can use this denormal feature now?

  // NEON f32 ops are non-IEEE 754 compliant. Darwin is ok with it by default.
  const FeatureBitset &Bits = getFeatureBits();
  if ((Bits[ARM::ProcA5] || Bits[ARM::ProcA8]) && // Where this matters
      (Options.UnsafeFPMath || isTargetDarwin()))

And maybe update the test, I think there are a few that this triggers on.

[davemgreen]

Thank you. LGTM

Hopefully what remains in #151275 is a lot simpler now.