GlobalISel: Handle more cases for widenScalar of G_MERGE_VALUES

Use an anyext to the requested type for the leftover operand to
produce a slightly wider type, and then truncate the final merge.

I have another implementation almost ready which handles arbitrary
widens, but I think it produces worse code in this example (which I
think is 90% due to not folding redundant copies or folding out
implicit_def users), so I wanted to add this as a baseline first.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@366366 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index 958e9b59cf443b8b8e13ddfa47e377c95e2c4db7..caa49cf2cf771d8c9290b95401c9d634d4722431 100644 (file)
--- a/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -875,10 +875,14 @@ LegalizerHelper::widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx,
 
   // Try to turn this into a merge of merges if we can use the requested type as
   // the source.
-
-  // TODO: Pad with undef if DstTy > WideTy
-  if (NumMerge > 1 && WideTy.getSizeInBits() % SrcTy.getSizeInBits() == 0) {
+  if (NumMerge > 1) {
     int PartsPerMerge = WideTy.getSizeInBits() / SrcTy.getSizeInBits();
+    if (WideTy.getSizeInBits() % SrcTy.getSizeInBits() != 0)
+      return UnableToLegalize;
+
+    int RemainderBits = DstTy.getSizeInBits() % WideTy.getSizeInBits();
+    int RemainderParts = RemainderBits / SrcTy.getSizeInBits();
+
     SmallVector<Register, 4> Parts;
     SmallVector<Register, 4> SubMerges;
 
@@ -891,7 +895,22 @@ LegalizerHelper::widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx,
       Parts.clear();
     }
 
-    MIRBuilder.buildMerge(DstReg, SubMerges);
+    if (RemainderParts == 0) {
+      MIRBuilder.buildMerge(DstReg, SubMerges);
+      MI.eraseFromParent();
+      return Legalized;
+    }
+
+    assert(RemainderParts == 1);
+
+    auto AnyExt = MIRBuilder.buildAnyExt(
+      WideTy, MI.getOperand(MI.getNumOperands() - 1).getReg());
+    SubMerges.push_back(AnyExt.getReg(0));
+
+    LLT WiderDstTy = LLT::scalar(SubMerges.size() * WideTy.getSizeInBits());
+    auto Merge = MIRBuilder.buildMerge(WiderDstTy, SubMerges);
+    MIRBuilder.buildTrunc(DstReg, Merge);
+
     MI.eraseFromParent();
     return Legalized;
   }

diff --git a/test/CodeGen/AMDGPU/GlobalISel/legalize-merge-values.mir b/test/CodeGen/AMDGPU/GlobalISel/legalize-merge-values.mir
index 7438180111db7d8645d45c2a4df4cbe830654447..20b65021b31031bebfe28e335626de67412ea6c8 100644 (file)
--- a/test/CodeGen/AMDGPU/GlobalISel/legalize-merge-values.mir
+++ b/test/CodeGen/AMDGPU/GlobalISel/legalize-merge-values.mir
@@ -535,3 +535,64 @@ body: |
     %8:_(s96) = G_MERGE_VALUES %0, %1, %2, %3, %4, %5
     $vgpr0_vgpr1_vgpr2 = COPY %8
 ...
+
+---
+name: test_merge_s56_s8_s8_s8_s8_s8_s8_s8
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: test_merge_s56_s8_s8_s8_s8_s8_s8_s8
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CHECK: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CHECK: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CHECK: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CHECK: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
+    ; CHECK: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
+    ; CHECK: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[C7]](s32)
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[C]](s32)
+    ; CHECK: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[TRUNC]]
+    ; CHECK: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CHECK: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY]], [[C9]]
+    ; CHECK: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CHECK: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C1]](s32)
+    ; CHECK: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C10]]
+    ; CHECK: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[AND1]](s32)
+    ; CHECK: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CHECK: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC2]]
+    ; CHECK: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[C2]](s32)
+    ; CHECK: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[TRUNC]]
+    ; CHECK: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C11]](s32)
+    ; CHECK: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; CHECK: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C3]](s32)
+    ; CHECK: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C10]]
+    ; CHECK: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[AND4]](s32)
+    ; CHECK: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CHECK: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND3]], [[TRUNC4]]
+    ; CHECK: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[C4]](s32)
+    ; CHECK: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[TRUNC]]
+    ; CHECK: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CHECK: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; CHECK: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C5]](s32)
+    ; CHECK: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C10]]
+    ; CHECK: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND8]], [[AND7]](s32)
+    ; CHECK: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CHECK: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC6]]
+    ; CHECK: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[C6]](s32)
+    ; CHECK: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[TRUNC7]](s16)
+    ; CHECK: [[TRUNC8:%[0-9]+]]:_(s56) = G_TRUNC [[MV]](s64)
+    ; CHECK: S_NOP 0, implicit [[TRUNC8]](s56)
+    %0:_(s8) = G_CONSTANT i8 0
+    %1:_(s8) = G_CONSTANT i8 1
+    %2:_(s8) = G_CONSTANT i8 2
+    %3:_(s8) = G_CONSTANT i8 3
+    %4:_(s8) = G_CONSTANT i8 4
+    %5:_(s8) = G_CONSTANT i8 5
+    %6:_(s8) = G_CONSTANT i8 6
+    %7:_(s56) = G_MERGE_VALUES %0, %1, %2, %3, %4, %5, %6
+    S_NOP 0, implicit %7
+...

diff --git a/unittests/CodeGen/GlobalISel/LegalizerHelperTest.cpp b/unittests/CodeGen/GlobalISel/LegalizerHelperTest.cpp
index 2ba95ab0826ef60af571660f5062b0276787f5e2..608c0ddf0847554c1f5f3e7148095d2ab7846388 100644 (file)
--- a/unittests/CodeGen/GlobalISel/LegalizerHelperTest.cpp
+++ b/unittests/CodeGen/GlobalISel/LegalizerHelperTest.cpp
@@ -9,6 +9,10 @@
 
 #include "GISelMITest.h"
 
+using namespace LegalizeActions;
+using namespace LegalizeMutations;
+using namespace LegalityPredicates;
+
 namespace {
 
 class DummyGISelObserver : public GISelChangeObserver {
@@ -900,4 +904,34 @@ TEST_F(GISelMITest, WidenScalarBuildVector) {
   EXPECT_TRUE(CheckMachineFunction(*MF, CheckStr)) << *MF;
 }
 
+TEST_F(GISelMITest, LowerMergeValues) {
+  if (!TM)
+    return;
+
+  const LLT S24 = LLT::scalar(24);
+  const LLT S9 = LLT::scalar(9);
+  const LLT S3 = LLT::scalar(3);
+
+  DefineLegalizerInfo(A, {
+    getActionDefinitionsBuilder(G_UNMERGE_VALUES)
+      .widenScalarIf(typeIs(1, LLT::scalar(3)), changeTo(1, LLT::scalar(9)));
+  });
+
+  AInfo Info(MF->getSubtarget());
+  DummyGISelObserver Observer;
+  LegalizerHelper Helper(*MF, Info, Observer, B);
+  B.setInsertPt(*EntryMBB, EntryMBB->end());
+
+  // 24 = 3 3 3   3 3 3   3 3
+  //     => 9
+  //
+  // This can do 2 merges for the first parts, but has 2 leftover operands.
+  SmallVector<Register, 7> MergeOps;
+  for (int I = 0; I != 8; ++I)
+    MergeOps.push_back(B.buildConstant(S3, I).getReg(0));
+
+  auto Merge = B.buildMerge(S24, MergeOps);
+  EXPECT_EQ(LegalizerHelper::LegalizeResult::UnableToLegalize,
+            Helper.lower(*Merge, 1, S9));
+}
 } // namespace