[AArch64][GlobalISel] Legalize 128 bit divisions to libcalls

[aemerson]

Summary

Now that we have the infrastructure to support s128 types as parameters we can expand these to libcalls.

[paquette]

llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp

2959–2961

I think that you need braces around the WideTy.isVector() || WideTy == LLT::Scalar(128) here, because && has higher precedence than ||.

[aemerson]

Fix assertion expression.

[paquette]

LGTM

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

for (unsigned I = 0, E = LeftoverRegs.size(); I != E; ++I) {
     CurResultReg = NewResultReg;
     Offset += LeftoverPartSize;
   }
 }
 
 static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) {
   switch (Opcode) {
   case TargetOpcode::G_SDIV:
-    assert((Size == 32 || Size == 64) && "Unsupported size");
-    return Size == 64 ? RTLIB::SDIV_I64 : RTLIB::SDIV_I32;
+    assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
+    switch (Size) {
+    case 32:
+      return RTLIB::SDIV_I32;
+    case 64:
+      return RTLIB::SDIV_I64;
+    case 128:
+      return RTLIB::SDIV_I128;
+    default:
+      llvm_unreachable("unexpected size");
+    }
   case TargetOpcode::G_UDIV:
-    assert((Size == 32 || Size == 64) && "Unsupported size");
-    return Size == 64 ? RTLIB::UDIV_I64 : RTLIB::UDIV_I32;
+    assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
+    switch (Size) {
+    case 32:
+      return RTLIB::UDIV_I32;
+    case 64:
+      return RTLIB::UDIV_I64;
+    case 128:
+      return RTLIB::UDIV_I128;
+    default:
+      llvm_unreachable("unexpected size");
+    }
   case TargetOpcode::G_SREM:
     assert((Size == 32 || Size == 64) && "Unsupported size");
     return Size == 64 ? RTLIB::SREM_I64 : RTLIB::SREM_I32;
   case TargetOpcode::G_UREM:
     assert((Size == 32 || Size == 64) && "Unsupported size");
     return Size == 64 ? RTLIB::UREM_I64 : RTLIB::UREM_I32;
   case TargetOpcode::G_CTLZ_ZERO_UNDEF:
     assert(Size == 32 && "Unsupported size");

llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp

bool AArch64InstructionSelector::selectUnmergeValues(
 
   // The last operand is the vector source register, and every other operand is
   // a register to unpack into.
   unsigned NumElts = I.getNumOperands() - 1;
   Register SrcReg = I.getOperand(NumElts).getReg();
   const LLT NarrowTy = MRI.getType(I.getOperand(0).getReg());
   const LLT WideTy = MRI.getType(SrcReg);
   (void)WideTy;
-  assert(WideTy.isVector() && "can only unmerge from vector types!");
+  assert((WideTy.isVector() || WideTy == LLT::scalar(128)) &&
+         "can only unmerge from vector types!");
   assert(WideTy.getSizeInBits() > NarrowTy.getSizeInBits() &&
          "source register size too small!");
 
   if (!NarrowTy.isScalar())
     return selectSplitVectorUnmerge(I, MRI);
 
   MachineIRBuilder MIB(I);
 
   // Choose a lane copy opcode and subregister based off of the size of the

llvm/lib/Target/AArch64/AArch64LegalizerInfo.cpp

AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST) {
   getActionDefinitionsBuilder(G_GEP)
       .legalFor({{p0, s64}})
       .clampScalar(1, s64, s64);
 
   getActionDefinitionsBuilder(G_PTR_MASK).legalFor({p0});
 
   getActionDefinitionsBuilder({G_SDIV, G_UDIV})
       .legalFor({s32, s64})
+      .libcallFor({s128})
       .clampScalar(0, s32, s64)
       .widenScalarToNextPow2(0)
       .scalarize(0);
 
   getActionDefinitionsBuilder({G_LSHR, G_ASHR})
       .customIf([=](const LegalityQuery &Query) {
         const auto &SrcTy = Query.Types[0];
         const auto &AmtTy = Query.Types[1];

llvm/test/CodeGen/AArch64/GlobalISel/legalize-s128-div.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple aarch64-apple-ios -run-pass=legalizer -verify-machineinstrs %s -o - | FileCheck %s
+--- |
+  target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+  target triple = "aarch64-apple-ios"
+
+  define void @udiv_test(i128* %v1ptr, i128* %v2ptr) { ret void }
+
+  define void @sdiv_test(i128* %v1ptr, i128* %v2ptr) { ret void }
+
+...
+---
+name:            udiv_test
+alignment:       2
+tracksRegLiveness: true
+liveins:
+  - { reg: '$x0' }
+  - { reg: '$x1' }
+machineFunctionInfo: {}
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: udiv_test
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p0) :: (load 16 from %ir.v1ptr)
+    ; CHECK: [[LOAD1:%[0-9]+]]:_(s128) = G_LOAD [[COPY1]](p0) :: (load 16 from %ir.v2ptr)
+    ; CHECK: ADJCALLSTACKDOWN 0, 0, implicit-def $sp, implicit $sp
+    ; CHECK: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[LOAD]](s128)
+    ; CHECK: [[UV2:%[0-9]+]]:_(s64), [[UV3:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[LOAD1]](s128)
+    ; CHECK: $x0 = COPY [[UV]](s64)
+    ; CHECK: $x1 = COPY [[UV1]](s64)
+    ; CHECK: $x2 = COPY [[UV2]](s64)
+    ; CHECK: $x3 = COPY [[UV3]](s64)
+    ; CHECK: BL &__udivti3, csr_aarch64_aapcs, implicit-def $lr, implicit $sp, implicit $x0, implicit $x1, implicit $x2, implicit $x3, implicit-def $x0, implicit-def $x1
+    ; CHECK: [[COPY2:%[0-9]+]]:_(s64) = COPY $x0
+    ; CHECK: [[COPY3:%[0-9]+]]:_(s64) = COPY $x1
+    ; CHECK: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[COPY2]](s64), [[COPY3]](s64)
+    ; CHECK: ADJCALLSTACKUP 0, 0, implicit-def $sp, implicit $sp
+    ; CHECK: G_STORE [[MV]](s128), [[COPY]](p0) :: (store 16 into %ir.v1ptr)
+    ; CHECK: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %1:_(p0) = COPY $x1
+    %2:_(s128) = G_LOAD %0(p0) :: (load 16 from %ir.v1ptr)
+    %3:_(s128) = G_LOAD %1(p0) :: (load 16 from %ir.v2ptr)
+    %4:_(s128) = G_UDIV %2, %3
+    G_STORE %4(s128), %0(p0) :: (store 16 into %ir.v1ptr)
+    RET_ReallyLR
+
+...
+---
+name:            sdiv_test
+alignment:       2
+tracksRegLiveness: true
+liveins:
+  - { reg: '$x0' }
+  - { reg: '$x1' }
+machineFunctionInfo: {}
+body:             |
+  bb.1 (%ir-block.0):
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: sdiv_test
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p0) :: (load 16 from %ir.v1ptr)
+    ; CHECK: [[LOAD1:%[0-9]+]]:_(s128) = G_LOAD [[COPY1]](p0) :: (load 16 from %ir.v2ptr)
+    ; CHECK: ADJCALLSTACKDOWN 0, 0, implicit-def $sp, implicit $sp
+    ; CHECK: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[LOAD]](s128)
+    ; CHECK: [[UV2:%[0-9]+]]:_(s64), [[UV3:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[LOAD1]](s128)
+    ; CHECK: $x0 = COPY [[UV]](s64)
+    ; CHECK: $x1 = COPY [[UV1]](s64)
+    ; CHECK: $x2 = COPY [[UV2]](s64)
+    ; CHECK: $x3 = COPY [[UV3]](s64)
+    ; CHECK: BL &__divti3, csr_aarch64_aapcs, implicit-def $lr, implicit $sp, implicit $x0, implicit $x1, implicit $x2, implicit $x3, implicit-def $x0, implicit-def $x1
+    ; CHECK: [[COPY2:%[0-9]+]]:_(s64) = COPY $x0
+    ; CHECK: [[COPY3:%[0-9]+]]:_(s64) = COPY $x1
+    ; CHECK: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[COPY2]](s64), [[COPY3]](s64)
+    ; CHECK: ADJCALLSTACKUP 0, 0, implicit-def $sp, implicit $sp
+    ; CHECK: G_STORE [[MV]](s128), [[COPY]](p0) :: (store 16 into %ir.v1ptr)
+    ; CHECK: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %1:_(p0) = COPY $x1
+    %2:_(s128) = G_LOAD %0(p0) :: (load 16 from %ir.v1ptr)
+    %3:_(s128) = G_LOAD %1(p0) :: (load 16 from %ir.v2ptr)
+    %4:_(s128) = G_SDIV %2, %3
+    G_STORE %4(s128), %0(p0) :: (store 16 into %ir.v1ptr)
+    RET_ReallyLR
+
+...