[GVN] Account for masked loads/stores depending on load/store instructions

[kparzysz]

Summary

This is a case where an intrinsic depends on a non-call instruction.

This fixes the testcase from https://reviews.llvm.org/D87061.

[fhahn]

LGTM thanks. Please run clang-format-diff before committing.

llvm/lib/Transforms/Scalar/GVN.cpp

417–418

nit: I think it is more common in this file to use !local_cdep instead of local_cdefp == nullptr.

llvm/test/Transforms/GVN/masked-load-store-vn-crash.ll

6

nit: I think the dso_local fastcc unnamed_addr #0 can eb removed.

llvm/lib/Transforms/Scalar/GVN.cpp

if (AA->doesNotAccessMemory(C)) {
     MemDepResult local_dep = MD->getDependency(C);
 
     if (!local_dep.isDef() && !local_dep.isNonLocal()) {
       valueNumbering[C] =  nextValueNumber;
       return nextValueNumber++;
     }
 
     if (local_dep.isDef()) {
-      CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
+      // For masked load/store intrinsics, the local_dep may actully be
+      // a normal load or store instruction.
+      CallInst *local_cdep = dyn_cast<CallInst>(local_dep.getInst());
 
-      if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
+      if (!local_cdep ||
+          local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
         valueNumbering[C] = nextValueNumber;
         return nextValueNumber++;
       }
 
       for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
         uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
         uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
         if (c_vn != cd_vn) {

llvm/test/Transforms/GVN/masked-load-store-vn-crash.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -gvn -S < %s | FileCheck %s
+@file_mask = external global [8 x i64], align 32
+
+define fastcc void @test() {
+; CHECK-LABEL: @test(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD_1_I:%.*]] = tail call <4 x i64> @llvm.masked.load.v4i64.p0v4i64(<4 x i64>* nonnull bitcast (i64* getelementptr inbounds ([8 x i64], [8 x i64]* @file_mask, i64 0, i64 7) to <4 x i64>*), i32 8, <4 x i1> <i1 true, i1 false, i1 false, i1 false>, <4 x i64> undef)
+; CHECK-NEXT:    unreachable
+;
+entry:
+  %wide.masked.load.1.i = tail call <4 x i64> @llvm.masked.load.v4i64.p0v4i64(<4 x i64>* nonnull bitcast (i64* getelementptr inbounds ([8 x i64], [8 x i64]* @file_mask, i64 0, i64 7) to <4 x i64>*), i32 8, <4 x i1> <i1 true, i1 false, i1 false, i1 false>, <4 x i64> undef) #2
+  %.pre392.i = load i64, i64* getelementptr inbounds ([8 x i64], [8 x i64]* @file_mask, i64 0, i64 7), align 8
+  %or156.4.i = or i64 %.pre392.i, undef
+  %wide.masked.load614.1.i = tail call <4 x i64> @llvm.masked.load.v4i64.p0v4i64(<4 x i64>* nonnull bitcast (i64* getelementptr inbounds ([8 x i64], [8 x i64]* @file_mask, i64 0, i64 7) to <4 x i64>*), i32 8, <4 x i1> <i1 true, i1 false, i1 false, i1 false>, <4 x i64> undef) #2
+  unreachable
+}
+
+; Function Attrs: argmemonly nounwind readonly willreturn
+declare <4 x i64> @llvm.masked.load.v4i64.p0v4i64(<4 x i64>*, i32 immarg, <4 x i1>, <4 x i64>)