GlobalISel: Handle more cases in getGCDType

[arsenm]

Summary

Try harder to find a canonical unmerge type when trying to cover the
desired target type. Handle finding a compatible unmerge type for two
vectors with different element types. This will return the largest
multiple of the source vector element that will evenly divide the
target vector type.

Also make the handling mixing scalars and vectors, and prefer the
source element type as the unmerge target type.

[arsenm]

ping

[arsenm]

Rebase

[arsenm]

ping

[arsenm]

ping

[aemerson]

llvm/lib/CodeGen/GlobalISel/Utils.cpp

546

Can we reorganize this to not have so much nesting?

Maybe duplicate greatestCommonDivisor(OrigSize, TargetSize); above this for the scalar case and early exit.

555

Invert this so the else early exits?

[arsenm]

llvm/lib/CodeGen/GlobalISel/Utils.cpp

546

I'm not sure what you mean, the fallthroughs here are significant. I don't see where duplicating greatestCommonDivisor(OrigSize, TargetSize) would help.

I also wouldn't all anything here the scalar case; care is taken to manage the behavior of vector and "not vector" to preserve pointers where appropriate. I think isScalar not being the same as !isVector is a constant source of confusion

555

These don't return though? It falls through

[arsenm]

ping

[arsenm]

ping

[arsenm]

12d5bec8c79b9841f1b25fa87ffb69b9d3cbec3b

llvm/include/llvm/CodeGen/GlobalISel/Utils.h

 }
 
 Align inferAlignFromPtrInfo(MachineFunction &MF, const MachinePointerInfo &MPO);
 
 /// Return the least common multiple type of \p Ty0 and \p Ty1, by changing
 /// the number of vector elements or scalar bitwidth. The intent is a
 /// G_MERGE_VALUES can be constructed from \p Ty0 elements, and unmerged into
 /// \p Ty1.
+LLVM_READNONE
 LLT getLCMType(LLT Ty0, LLT Ty1);
 
-/// Return a type that is greatest common divisor of \p OrigTy and \p
-/// TargetTy. This will either change the number of vector elements, or
-/// bitwidth of scalars. The intent is the result type can be used as the
-/// result of a G_UNMERGE_VALUES from \p OrigTy.
+/// Return a type where the total size is the greatest common divisor of \p
+/// OrigTy and \p TargetTy. This will try to either change the number of vector
+/// elements, or bitwidth of scalars. The intent is the result type can be used
+/// as the result of a G_UNMERGE_VALUES from \p OrigTy, and then some
+/// combination of G_MERGE_VALUES, G_BUILD_VECTOR and G_CONCAT_VECTORS (possibly
+/// with intermediate casts) can re-form \p TargetTy.
+///
+/// If these are vectors with different element types, this will try to produce
+/// a vector with a compatible total size, but the element type of \p OrigTy. If
+/// this can't be satisfied, this will produce a scalar smaller than the
+/// original vector elements.
+///
+/// In the worst case, this returns LLT::scalar(1)
+LLVM_READNONE
 LLT getGCDType(LLT OrigTy, LLT TargetTy);
 
 } // End namespace llvm.
 #endif

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

static void getUnmergeResults(SmallVectorImpl<Register> &Regs,
   for (int I = 0; I != NumResults; ++I)
     Regs[I] = MI.getOperand(I).getReg();
 }
 
 LLT LegalizerHelper::extractGCDType(SmallVectorImpl<Register> &Parts, LLT DstTy,
                                     LLT NarrowTy, Register SrcReg) {
   LLT SrcTy = MRI.getType(SrcReg);
 
-  LLT GCDTy = getGCDType(DstTy, getGCDType(SrcTy, NarrowTy));
+  LLT GCDTy = getGCDType(getGCDType(SrcTy, NarrowTy), DstTy);
   if (SrcTy == GCDTy) {
     // If the source already evenly divides the result type, we don't need to do
     // anything.
     Parts.push_back(SrcReg);
   } else {
     // Need to split into common type sized pieces.
     auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg);
     getUnmergeResults(Parts, *Unmerge);

llvm/lib/CodeGen/GlobalISel/Utils.cpp

if (Ty0.isVector() && Ty1.isVector()) {
     int Mul = Ty0.getNumElements() * Ty1.getNumElements();
     return LLT::vector(Mul / GCDElts, Ty0.getElementType());
   }
 
   llvm_unreachable("not yet handled");
 }
 
 LLT llvm::getGCDType(LLT OrigTy, LLT TargetTy) {
-  if (OrigTy.isVector() && TargetTy.isVector()) {
-    assert(OrigTy.getElementType() == TargetTy.getElementType());
+  const unsigned OrigSize = OrigTy.getSizeInBits();
+  const unsigned TargetSize = TargetTy.getSizeInBits();
+
+  if (OrigSize == TargetSize)
+    return OrigTy;
+
+  if (OrigTy.isVector()) {
+    LLT OrigElt = OrigTy.getElementType();
+    if (TargetTy.isVector()) {
+      LLT TargetElt = TargetTy.getElementType();
+      if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) {
         int GCD = greatestCommonDivisor(OrigTy.getNumElements(),
                                         TargetTy.getNumElements());
-    return LLT::scalarOrVector(GCD, OrigTy.getElementType());
+        return LLT::scalarOrVector(GCD, OrigElt);
+      }
+    } else {
+      // If the source is a vector of pointers, return a pointer element.
+      if (OrigElt.getSizeInBits() == TargetSize)
+        return OrigElt;
     }
 
-  if (OrigTy.isVector() && !TargetTy.isVector()) {
-    assert(OrigTy.getElementType() == TargetTy);
-    return TargetTy;
+    unsigned GCD = greatestCommonDivisor(OrigSize, TargetSize);
+    if (GCD == OrigElt.getSizeInBits())
+      return OrigElt;
+
+    // If we can't produce the original element type, we have to use a smaller
+    // scalar.
+    if (GCD < OrigElt.getSizeInBits())
+      return LLT::scalar(GCD);
+    return LLT::vector(GCD / OrigElt.getSizeInBits(), OrigElt);
   }
 
-  assert(!OrigTy.isVector() && !TargetTy.isVector() &&
-         "GCD type of vector and scalar not implemented");
+  if (TargetTy.isVector()) {
+    // Try to preserve the original element type.
+    LLT TargetElt = TargetTy.getElementType();
+    if (TargetElt.getSizeInBits() == OrigSize)
+      return OrigTy;
+  }
 
-  int GCD = greatestCommonDivisor(OrigTy.getSizeInBits(),
-                                  TargetTy.getSizeInBits());
+  unsigned GCD = greatestCommonDivisor(OrigSize, TargetSize);
   return LLT::scalar(GCD);
 }

llvm/unittests/CodeGen/GlobalISel/GISelUtilsTest.cpp

 //===- GISelUtilsTest.cpp -------------------------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "llvm/CodeGen/GlobalISel/Utils.h"
 #include "gtest/gtest.h"
 
 using namespace llvm;
 
 namespace {
 constexpr LLT S1 = LLT::scalar(1);
+constexpr LLT S8 = LLT::scalar(8);
 constexpr LLT S16 = LLT::scalar(16);
 constexpr LLT S32 = LLT::scalar(32);
 constexpr LLT S64 = LLT::scalar(64);
 constexpr LLT P0 = LLT::pointer(0, 64);
 constexpr LLT P1 = LLT::pointer(1, 32);
 
+constexpr LLT V2S8 = LLT::vector(2, 8);
+constexpr LLT V4S8 = LLT::vector(4, 8);
+
 constexpr LLT V2S16 = LLT::vector(2, 16);
+constexpr LLT V3S16 = LLT::vector(3, 16);
 constexpr LLT V4S16 = LLT::vector(4, 16);
 
 constexpr LLT V2S32 = LLT::vector(2, 32);
 constexpr LLT V3S32 = LLT::vector(3, 32);
 constexpr LLT V4S32 = LLT::vector(4, 32);
 constexpr LLT V6S32 = LLT::vector(6, 32);
 
+constexpr LLT V2S64 = LLT::vector(2, 64);
+constexpr LLT V4S64 = LLT::vector(4, 64);
+
 constexpr LLT V2P0 = LLT::vector(2, P0);
 constexpr LLT V3P0 = LLT::vector(3, P0);
 constexpr LLT V4P0 = LLT::vector(4, P0);
 constexpr LLT V6P0 = LLT::vector(6, P0);
 
+constexpr LLT V2P1 = LLT::vector(2, P1);
+constexpr LLT V4P1 = LLT::vector(4, P1);
+
 TEST(GISelUtilsTest, getGCDType) {
   EXPECT_EQ(S1, getGCDType(S1, S1));
   EXPECT_EQ(S32, getGCDType(S32, S32));
   EXPECT_EQ(S1, getGCDType(S1, S32));
   EXPECT_EQ(S1, getGCDType(S32, S1));
   EXPECT_EQ(S16, getGCDType(S16, S32));
   EXPECT_EQ(S16, getGCDType(S32, S16));
 
TEST(GISelUtilsTest, getGCDType) {
 
 
   EXPECT_EQ(S16, getGCDType(P0, S16));
   EXPECT_EQ(S16, getGCDType(S16, P0));
 
   EXPECT_EQ(S32, getGCDType(P0, S32));
   EXPECT_EQ(S32, getGCDType(S32, P0));
 
-  EXPECT_EQ(S64, getGCDType(P0, S64));
+  EXPECT_EQ(P0, getGCDType(P0, S64));
   EXPECT_EQ(S64, getGCDType(S64, P0));
 
   EXPECT_EQ(S32, getGCDType(P0, P1));
   EXPECT_EQ(S32, getGCDType(P1, P0));
 
   EXPECT_EQ(P0, getGCDType(V3P0, V2P0));
   EXPECT_EQ(P0, getGCDType(V2P0, V3P0));
+
+  EXPECT_EQ(P0, getGCDType(P0, V2P0));
+  EXPECT_EQ(P0, getGCDType(V2P0, P0));
+
+
+  EXPECT_EQ(V2P0, getGCDType(V2P0, V2P0));
+  EXPECT_EQ(P0, getGCDType(V3P0, V2P0));
+  EXPECT_EQ(P0, getGCDType(V2P0, V3P0));
+  EXPECT_EQ(V2P0, getGCDType(V4P0, V2P0));
+
+  EXPECT_EQ(V2P0, getGCDType(V2P0, V4P1));
+  EXPECT_EQ(V4P1, getGCDType(V4P1, V2P0));
+
+  EXPECT_EQ(V2P0, getGCDType(V4P0, V4P1));
+  EXPECT_EQ(V4P1, getGCDType(V4P1, V4P0));
+
+  // Elements have same size, but have different pointeriness, so prefer the
+  // original element type.
+  EXPECT_EQ(V2P0, getGCDType(V2P0, V4S64));
+  EXPECT_EQ(V2S64, getGCDType(V4S64, V2P0));
+
+  EXPECT_EQ(V2S16, getGCDType(V2S16, V4P1));
+  EXPECT_EQ(P1, getGCDType(V4P1, V2S16));
+  EXPECT_EQ(V2P1, getGCDType(V4P1, V4S16));
+  EXPECT_EQ(V4S16, getGCDType(V4S16, V2P1));
+
+  EXPECT_EQ(P0, getGCDType(P0, V2S64));
+  EXPECT_EQ(S64, getGCDType(V2S64, P0));
+
+  EXPECT_EQ(S16, getGCDType(V2S16, V3S16));
+  EXPECT_EQ(S16, getGCDType(V3S16, V2S16));
+  EXPECT_EQ(S16, getGCDType(V3S16, S16));
+  EXPECT_EQ(S16, getGCDType(S16, V3S16));
+
+  EXPECT_EQ(V2S16, getGCDType(V2S16, V2S32));
+  EXPECT_EQ(S32, getGCDType(V2S32, V2S16));
+
+  EXPECT_EQ(V4S8, getGCDType(V4S8, V2S32));
+  EXPECT_EQ(S32, getGCDType(V2S32, V4S8));
+
+  // Test cases where neither element type nicely divides.
+  EXPECT_EQ(LLT::scalar(3), getGCDType(LLT::vector(3, 5), LLT::vector(2, 6)));
+  EXPECT_EQ(LLT::scalar(3), getGCDType(LLT::vector(2, 6), LLT::vector(3, 5)));
+
+  // Have to go smaller than a pointer element.
+  EXPECT_EQ(LLT::scalar(3), getGCDType(LLT::vector(2, LLT::pointer(3, 6)),
+                                       LLT::vector(3, 5)));
+  EXPECT_EQ(LLT::scalar(3), getGCDType(LLT::vector(3, 5),
+                                       LLT::vector(2, LLT::pointer(3, 6))));
+
+  EXPECT_EQ(V4S8, getGCDType(V4S8, S32));
+  EXPECT_EQ(S32, getGCDType(S32, V4S8));
+  EXPECT_EQ(V4S8, getGCDType(V4S8, P1));
+  EXPECT_EQ(P1, getGCDType(P1, V4S8));
+
+  EXPECT_EQ(V2S8, getGCDType(V2S8, V4S16));
+  EXPECT_EQ(S16, getGCDType(V4S16, V2S8));
+
+  EXPECT_EQ(S8, getGCDType(V2S8, LLT::vector(4, 2)));
+  EXPECT_EQ(LLT::vector(4, 2), getGCDType(LLT::vector(4, 2), S8));
+
+
+  EXPECT_EQ(LLT::pointer(4, 8), getGCDType(LLT::vector(2, LLT::pointer(4, 8)),
+                                           LLT::vector(4, 2)));
+
+  EXPECT_EQ(LLT::vector(4, 2), getGCDType(LLT::vector(4, 2),
+                                          LLT::vector(2, LLT::pointer(4, 8))));
+
+  EXPECT_EQ(LLT::scalar(4), getGCDType(LLT::vector(3, 4), S8));
+  EXPECT_EQ(LLT::scalar(4), getGCDType(S8, LLT::vector(3, 4)));
 }
 
 TEST(GISelUtilsTest, getLCMType) {
   EXPECT_EQ(S1, getLCMType(S1, S1));
   EXPECT_EQ(S32, getLCMType(S32, S1));
   EXPECT_EQ(S32, getLCMType(S1, S32));
   EXPECT_EQ(S32, getLCMType(S32, S32));