[AggressiveInstCombine] Fix invalid TypeSize conversion when combining loads.

Much of foldLoadsRecursive relies on knowing the size of loaded
data, which is not possible for scalable vector types.  However,
the logic of combining two small loads into one bigger load does
not apply for vector types so rather than converting the algorithm
to use TypeSize I've simply added an early exit for vectors.

Fixes #59510

Differential Revision: https://reviews.llvm.org/D140106

diff --git a/llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp b/llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp
index 5da7dbb..473b412 100644 (file)
--- a/llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp
+++ b/llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp
@@ -769,6 +769,10 @@ static bool foldLoadsRecursive(Value *V, LoadOps &LOps, const DataLayout &DL,
 // of the loads is to form a wider load.
 static bool foldConsecutiveLoads(Instruction &I, const DataLayout &DL,
                                  TargetTransformInfo &TTI, AliasAnalysis &AA) {
+  // Only consider load chains of scalar values.
+  if (isa<VectorType>(I.getType()))
+    return false;
+
   LoadOps LOps;
   if (!foldLoadsRecursive(&I, LOps, DL, AA) || !LOps.FoundRoot)
     return false;

diff --git a/llvm/test/Transforms/AggressiveInstCombine/vector-or-load.ll b/llvm/test/Transforms/AggressiveInstCombine/vector-or-load.ll
new file mode 100644 (file)
index 0000000..f7d48ce
--- /dev/null
+++ b/llvm/test/Transforms/AggressiveInstCombine/vector-or-load.ll
@@ -0,0 +1,50 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=aggressive-instcombine -S | FileCheck %s
+
+; Use the same idoim as X86/or-load.ll @loadCombine_2consecutive to represent a
+; sequence of operations that can be replaced by a single double-width load when
+; using scalar types but whose logic does not apply to fixed length vectors.
+define <8 x i16> @or-load-fixed-length-vector(ptr %p1) {
+; CHECK-LABEL: @or-load-fixed-length-vector(
+; CHECK-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P1:%.*]], i32 1
+; CHECK-NEXT:    [[L1:%.*]] = load <8 x i8>, ptr [[P1]], align 1
+; CHECK-NEXT:    [[L2:%.*]] = load <8 x i8>, ptr [[P2]], align 1
+; CHECK-NEXT:    [[E1:%.*]] = zext <8 x i8> [[L1]] to <8 x i16>
+; CHECK-NEXT:    [[E2:%.*]] = zext <8 x i8> [[L2]] to <8 x i16>
+; CHECK-NEXT:    [[S2:%.*]] = shl <8 x i16> [[E2]], <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
+; CHECK-NEXT:    [[OR:%.*]] = or <8 x i16> [[E1]], [[S2]]
+; CHECK-NEXT:    ret <8 x i16> [[OR]]
+;
+  %p2 = getelementptr i8, ptr %p1, i32 1
+  %l1 = load <8 x i8>, ptr %p1, align 1
+  %l2 = load <8 x i8>, ptr %p2, align 1
+  %e1 = zext <8 x i8> %l1 to <8 x i16>
+  %e2 = zext <8 x i8> %l2 to <8 x i16>
+  %s2 = shl <8 x i16> %e2, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
+  %or = or <8 x i16> %e1, %s2
+  ret <8 x i16> %or
+}
+
+; Use the same idoim as X86/or-load.ll @loadCombine_2consecutive to represent a
+; sequence of operations that can be replaced by a single double-width load when
+; using scalar types but whose logic does not apply to scalable length vectors.
+define <vscale x 8 x i16> @or-load-scalable-vector(ptr %p1) {
+; CHECK-LABEL: @or-load-scalable-vector(
+; CHECK-NEXT:    [[P2:%.*]] = getelementptr i8, ptr [[P1:%.*]], i32 1
+; CHECK-NEXT:    [[L1:%.*]] = load <vscale x 8 x i8>, ptr [[P1]], align 1
+; CHECK-NEXT:    [[L2:%.*]] = load <vscale x 8 x i8>, ptr [[P2]], align 1
+; CHECK-NEXT:    [[E1:%.*]] = zext <vscale x 8 x i8> [[L1]] to <vscale x 8 x i16>
+; CHECK-NEXT:    [[E2:%.*]] = zext <vscale x 8 x i8> [[L2]] to <vscale x 8 x i16>
+; CHECK-NEXT:    [[S2:%.*]] = shl <vscale x 8 x i16> [[E2]], shufflevector (<vscale x 8 x i16> insertelement (<vscale x 8 x i16> poison, i16 8, i32 0), <vscale x 8 x i16> poison, <vscale x 8 x i32> zeroinitializer)
+; CHECK-NEXT:    [[OR:%.*]] = or <vscale x 8 x i16> [[E1]], [[S2]]
+; CHECK-NEXT:    ret <vscale x 8 x i16> [[OR]]
+;
+  %p2 = getelementptr i8, ptr %p1, i32 1
+  %l1 = load <vscale x 8 x i8>, ptr %p1, align 1
+  %l2 = load <vscale x 8 x i8>, ptr %p2, align 1
+  %e1 = zext <vscale x 8 x i8> %l1 to <vscale x 8 x i16>
+  %e2 = zext <vscale x 8 x i8> %l2 to <vscale x 8 x i16>
+  %s2 = shl <vscale x 8 x i16> %e2, shufflevector (<vscale x 8 x i16> insertelement (<vscale x 8 x i16> poison, i16 8, i32 0), <vscale x 8 x i16> poison, <vscale x 8 x i32> zeroinitializer)
+  %or = or <vscale x 8 x i16> %e1, %s2
+  ret <vscale x 8 x i16> %or
+}