[ARM] f16 stack spill/reloads

[SjoerdMeijer]

Summary

This adds support for handling f16 stack spills/reloads.

[SjoerdMeijer]

Forgot the Thumb test case.

[olista01]

test/CodeGen/ARM/fp16-instructions.ll

721

This test looks more complicated than it needs to be, can it be reduced?

[SjoerdMeijer]

This IR is generated from the reduced C test case:

fn1() {
  fn2(9.765625e4F16);
  fn3(9.765625e4F16);
}

There is a bit of bitcasting going on because of the way we pass half-floats,
and a few alloca's because the easiest way to trigger this is with fast-isel
and no reg allocation, so I am not sure how I can further reduce this.

[olista01]

The test case doesn't have to be exactly the -O0 output from clang, you should reduce it to clear IR which triggers the bug. To start with, are all of those allocas and memcpy calls required?

[SjoerdMeijer]

Ok, yes, thanks. Test case now reduced.

[olista01]

LGTM, thanks

lib/Target/ARM/ARMBaseInstrInfo.cpp

storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
   MachineFrameInfo &MFI = MF.getFrameInfo();
   unsigned Align = MFI.getObjectAlignment(FI);
 
   MachineMemOperand *MMO = MF.getMachineMemOperand(
       MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore,
       MFI.getObjectSize(FI), Align);
 
   switch (TRI->getSpillSize(*RC)) {
+    case 2:
+      if (ARM::HPRRegClass.hasSubClassEq(RC)) {
+        BuildMI(MBB, I, DL, get(ARM::VSTRH))
+            .addReg(SrcReg, getKillRegState(isKill))
+            .addFrameIndex(FI)
+            .addImm(0)
+            .addMemOperand(MMO)
+            .add(predOps(ARMCC::AL));
+      } else
+        llvm_unreachable("Unknown reg class!");
+      break;
     case 4:
       if (ARM::GPRRegClass.hasSubClassEq(RC)) {
         BuildMI(MBB, I, DL, get(ARM::STRi12))
             .addReg(SrcReg, getKillRegState(isKill))
             .addFrameIndex(FI)
             .addImm(0)
             .addMemOperand(MMO)
             .add(predOps(ARMCC::AL));
loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
   MachineFunction &MF = *MBB.getParent();
   MachineFrameInfo &MFI = MF.getFrameInfo();
   unsigned Align = MFI.getObjectAlignment(FI);
   MachineMemOperand *MMO = MF.getMachineMemOperand(
       MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad,
       MFI.getObjectSize(FI), Align);
 
   switch (TRI->getSpillSize(*RC)) {
+  case 2:
+    if (ARM::HPRRegClass.hasSubClassEq(RC)) {
+      BuildMI(MBB, I, DL, get(ARM::VLDRH), DestReg)
+          .addFrameIndex(FI)
+          .addImm(0)
+          .addMemOperand(MMO)
+          .add(predOps(ARMCC::AL));
+    } else
+      llvm_unreachable("Unknown reg class!");
+    break;
   case 4:
     if (ARM::GPRRegClass.hasSubClassEq(RC)) {
       BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg)
           .addFrameIndex(FI)
           .addImm(0)
           .addMemOperand(MMO)
           .add(predOps(ARMCC::AL));
-
     } else if (ARM::SPRRegClass.hasSubClassEq(RC)) {
       BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg)
           .addFrameIndex(FI)
           .addImm(0)
           .addMemOperand(MMO)
           .add(predOps(ARMCC::AL));
     } else
       llvm_unreachable("Unknown reg class!");

test/CodeGen/ARM/fp16-instructions.ll

 ; SOFT:
 ; RUN: llc < %s -mtriple=arm-none-eabi -float-abi=soft     | FileCheck %s --check-prefixes=CHECK,CHECK-SOFT
 ; RUN: llc < %s -mtriple=thumb-none-eabi -float-abi=soft   | FileCheck %s --check-prefixes=CHECK,CHECK-SOFT
 
 ; SOFTFP:
 ; RUN: llc < %s -mtriple=arm-none-eabi -mattr=+vfp3        | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-VFP3
 ; RUN: llc < %s -mtriple=arm-none-eabi -mattr=+vfp4        | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-FP16
 ; RUN: llc < %s -mtriple=arm-none-eabi -mattr=+fullfp16    | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-FULLFP16
 
+; RUN: llc < %s -mtriple=arm-none-eabi -mattr=+fullfp16 -O0 | FileCheck %s --check-prefixes=CHECK-SPILL-RELOAD
+
 ; RUN: llc < %s -mtriple=thumbv7-none-eabi -mattr=+vfp3        | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-VFP3
 ; RUN: llc < %s -mtriple=thumbv7-none-eabi -mattr=+vfp4        | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-FP16
 ; RUN: llc < %s -mtriple=thumbv7-none-eabi -mattr=+fullfp16    | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-FULLFP16
 
 ; HARD:
 ; RUN: llc < %s -mtriple=arm-none-eabihf -mattr=+vfp3      | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-VFP3
 ; RUN: llc < %s -mtriple=arm-none-eabihf -mattr=+vfp4      | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-FP16
 ; RUN: llc < %s -mtriple=arm-none-eabihf -mattr=+fullfp16  | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-FULLFP16
entry:
   %add = fadd half %S.0.S.0., %0
   %1 = bitcast half %add to i16
   %tmp2.0.insert.ext = zext i16 %1 to i32
   ret i32 %tmp2.0.insert.ext
 
 ; CHECK-LABEL:            ThumbAddrMode5FP16
 ; CHECK-SOFTFP-FULLFP16:  vstr.16 s0, [sp, #2]
 }
+
+; Test function calls to check store/load reg to/from stack
+define i32 @fn1() {
+entry:
+  %retval = alloca i32, align 4
+  %coerce = alloca half, align 2
+  %tmp = alloca i32, align 4
+  %coerce1 = alloca half, align 2
+  %tmp2 = alloca i32, align 4
+  store half 0xH7C00, half* %coerce, align 2
+  %0 = bitcast i32* %tmp to i8*
+  %1 = bitcast half* %coerce to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %0, i8* align 2 %1, i32 2, i1 false)
+  %2 = load i32, i32* %tmp, align 4
+  %call = call i32 bitcast (i32 (...)* @fn2 to i32 (i32)*)(i32 %2)
+  store half 0xH7C00, half* %coerce1, align 2
+  %3 = bitcast i32* %tmp2 to i8*
+  %4 = bitcast half* %coerce1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %3, i8* align 2 %4, i32 2, i1 false)
+  %5 = load i32, i32* %tmp2, align 4
+  %call3 = call i32 bitcast (i32 (...)* @fn3 to i32 (i32)*)(i32 %5)
+  %6 = load i32, i32* %retval, align 4
+  ret i32 %6
+
+; CHECK-SPILL-RELOAD-LABEL: fn1:
+; CHECK-SPILL-RELOAD:       vstr.16 s0, [sp, #{{.}}]  @ 2-byte Spill
+; CHECK-SPILL-RELOAD-NEXT:  bl  fn2
+; CHECK-SPILL-RELOAD-NEXT:  vldr.16 s0, [sp, #8]      @ 2-byte Reload
+}
+
+declare dso_local i32 @fn2(...) #1
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture writeonly, i8* nocapture readonly, i32, i1) #2
+declare dso_local i32 @fn3(...) #1