[PATCH] D105316: [GlobalISel] Translate <1 x N> getelementptrs to scalar G_PTR_ADDs

[Jessica Paquette via Phabricator via llvm-commits]

paquette created this revision.
paquette added reviewers: aemerson, sdesmalen, arsenm.
Herald added subscribers: hiraditya, rovka.
paquette requested review of this revision.
Herald added a subscriber: wdng.
Herald added a project: LLVM.

In `IRTranslator::translateGetElementPtr`, when we run into a vector gep with some scalar operands, we try to normalize those operands using `buildSplatVector`.

This is fine except for when the getelementptr has a <1 x N> type. In that case it is treated as a scalar. If we run into one of these then every call to

  // With VectorWidth = 1
  LLT::fixed_vector(VectorWidth, PtrTy)

will assert.

Here's an example (equivalent to the added testcase):

https://godbolt.org/z/hGsTnMYdW

  llc: llvm/include/llvm/Support/LowLevelTypeImpl.h:67: static llvm::LLT llvm::LLT::vector(llvm::ElementCount, llvm::LLT): Assertion `!EC.isScalar() && "invalid number of vector elements"' failed.

To get around this, this patch adds a variable, `WantSplatVector`, which is true when our vector type ought to actually be represented using a vector. When it's false, we'll translate as a scalar. This checks if `VectorWidth > 1`.

This fixes this bug: https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=35496


https://reviews.llvm.org/D105316

Files:
  llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
  llvm/test/CodeGen/AArch64/GlobalISel/irtranslator-one-by-n-vector-ptr-add.ll


Index: llvm/test/CodeGen/AArch64/GlobalISel/irtranslator-one-by-n-vector-ptr-add.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AArch64/GlobalISel/irtranslator-one-by-n-vector-ptr-add.ll
@@ -0,0 +1,42 @@
+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -O0 -global-isel -mtriple aarch64 -stop-after=irtranslator -verify-machineinstrs %s -o - | FileCheck %s
+
+; Make sure we treat <1 x N> getelementptrs like scalar getelementptrs.
+
+; We should not create a splat vector for the non-vector index on this
+; getelementptr. The entire getelementptr should be translated to a scalar
+; G_PTR_ADD.
+define <1 x i8*> @one_elt_vector_ptr_add_non_vector_idx(<1 x i8*> %vec) {
+  ; CHECK-LABEL: name: one_elt_vector_ptr_add_non_vector_idx
+  ; CHECK: bb.1 (%ir-block.0):
+  ; CHECK:   liveins: $d0
+  ; CHECK:   [[COPY:%[0-9]+]]:_(p0) = COPY $d0
+  ; CHECK:   [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+  ; CHECK:   [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
+  ; CHECK:   [[SEXT:%[0-9]+]]:_(s64) = G_SEXT [[COPY1]](s32)
+  ; CHECK:   [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[SEXT]](s64)
+  ; CHECK:   [[COPY2:%[0-9]+]]:_(p0) = COPY [[PTR_ADD]](p0)
+  ; CHECK:   $d0 = COPY [[COPY2]](p0)
+  ; CHECK:   RET_ReallyLR implicit $d0
+  %ptr_add = getelementptr i8, <1 x i8*> %vec, <1 x i32> <i32 1>
+  ret <1 x i8*> %ptr_add
+}
+
+; We should not create a splat vector for the non-vector pointer on this
+; getelementptr. The entire getelementptr should be translated to a scalar
+; G_PTR_ADD.
+define <1 x i8*> @one_elt_vector_ptr_add_non_vector_ptr(i8* %vec) {
+  ; CHECK-LABEL: name: one_elt_vector_ptr_add_non_vector_ptr
+  ; CHECK: bb.1 (%ir-block.0):
+  ; CHECK:   liveins: $x0
+  ; CHECK:   [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+  ; CHECK:   [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+  ; CHECK:   [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
+  ; CHECK:   [[SEXT:%[0-9]+]]:_(s64) = G_SEXT [[COPY1]](s32)
+  ; CHECK:   [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[SEXT]](s64)
+  ; CHECK:   [[COPY2:%[0-9]+]]:_(p0) = COPY [[PTR_ADD]](p0)
+  ; CHECK:   $d0 = COPY [[COPY2]](p0)
+  ; CHECK:   RET_ReallyLR implicit $d0
+  %ptr_add = getelementptr i8, i8* %vec, <1 x i32> <i32 1>
+  ret <1 x i8*> %ptr_add
+}
Index: llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
===================================================================
--- llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
+++ llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
@@ -1472,12 +1472,19 @@
   // Normalize Vector GEP - all scalar operands should be converted to the
   // splat vector.
   unsigned VectorWidth = 0;
-  if (auto *VT = dyn_cast<VectorType>(U.getType()))
+
+  // True if we should use a splat vector; using VectorWidth alone is not
+  // sufficient.
+  bool WantSplatVector = false;
+  if (auto *VT = dyn_cast<VectorType>(U.getType())) {
     VectorWidth = cast<FixedVectorType>(VT)->getNumElements();
+    // We don't produce 1 x N vectors; those are treated as scalars.
+    WantSplatVector = VectorWidth > 1;
+  }
 
   // We might need to splat the base pointer into a vector if the offsets
   // are vectors.
-  if (VectorWidth && !PtrTy.isVector()) {
+  if (WantSplatVector && !PtrTy.isVector()) {
     BaseReg =
         MIRBuilder
             .buildSplatVector(LLT::fixed_vector(VectorWidth, PtrTy), BaseReg)
@@ -1516,7 +1523,7 @@
       Register IdxReg = getOrCreateVReg(*Idx);
       LLT IdxTy = MRI->getType(IdxReg);
       if (IdxTy != OffsetTy) {
-        if (!IdxTy.isVector() && VectorWidth) {
+        if (!IdxTy.isVector() && WantSplatVector) {
           IdxReg = MIRBuilder.buildSplatVector(
             OffsetTy.changeElementType(IdxTy), IdxReg).getReg(0);
         }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D105316.356005.patch
Type: text/x-patch
Size: 3762 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210701/bc922064/attachment.bin>