[GlobalISel][IRTranslator] When translating vector geps, splat the base pointer if...

We can have geps that have a scalar base pointer, and a vector index value, which
means that the base pointer must be splatted into a vector of pointers.

This fixes crashes on arm64 GlobalISel with optimizations enabled.

diff --git a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
index e29eb15..730c633 100644 (file)
--- a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
@@ -1062,6 +1062,18 @@ bool IRTranslator::translateGetElementPtr(const User &U,
   if (auto *VT = dyn_cast<VectorType>(U.getType()))
     VectorWidth = VT->getNumElements();
 
+  // We might need to splat the base pointer into a vector if the offsets
+  // are vectors.
+  if (VectorWidth && !PtrTy.isVector()) {
+    BaseReg =
+        MIRBuilder.buildSplatVector(LLT::vector(VectorWidth, PtrTy), BaseReg)
+            .getReg(0);
+    PtrIRTy = VectorType::get(PtrIRTy, VectorWidth);
+    PtrTy = getLLTForType(*PtrIRTy, *DL);
+    OffsetIRTy = DL->getIntPtrType(PtrIRTy);
+    OffsetTy = getLLTForType(*OffsetIRTy, *DL);
+  }
+
   int64_t Offset = 0;
   for (gep_type_iterator GTI = gep_type_begin(&U), E = gep_type_end(&U);
        GTI != E; ++GTI) {

diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/translate-gep.ll b/llvm/test/CodeGen/AArch64/GlobalISel/translate-gep.ll
index b1cf6b2..1f812f9 100644 (file)
--- a/llvm/test/CodeGen/AArch64/GlobalISel/translate-gep.ll
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/translate-gep.ll
@@ -117,3 +117,24 @@ define i32* @var_then_const(%type1* %addr, i64 %idx) {
   %res = getelementptr %type1, %type1* %addr, i64 %idx, i32 2, i32 2
   ret i32* %res
 }
+
+@arr = external global [8 x i32]
+
+define <2 x i32*> @vec_gep_scalar_base(<2 x i64> %offs) {
+  ; CHECK-LABEL: name: vec_gep_scalar_base
+  ; CHECK: bb.1.entry:
+  ; CHECK:   liveins: $q0
+  ; CHECK:   [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $q0
+  ; CHECK:   [[GV:%[0-9]+]]:_(p0) = G_GLOBAL_VALUE @arr
+  ; CHECK:   [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p0>) = G_BUILD_VECTOR [[GV]](p0), [[GV]](p0)
+  ; CHECK:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+  ; CHECK:   [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[C]](s64), [[C]](s64)
+  ; CHECK:   [[MUL:%[0-9]+]]:_(<2 x s64>) = G_MUL [[COPY]], [[BUILD_VECTOR1]]
+  ; CHECK:   [[PTR_ADD:%[0-9]+]]:_(<2 x p0>) = G_PTR_ADD [[BUILD_VECTOR]], [[MUL]](<2 x s64>)
+  ; CHECK:   [[COPY1:%[0-9]+]]:_(<2 x p0>) = COPY [[PTR_ADD]](<2 x p0>)
+  ; CHECK:   $q0 = COPY [[COPY1]](<2 x p0>)
+  ; CHECK:   RET_ReallyLR implicit $q0
+entry:
+  %0 = getelementptr inbounds [8 x i32], [8 x i32]* @arr, i64 0, <2 x i64> %offs
+  ret <2 x i32*> %0
+}