[MergeICmps] Don't merge icmps derived from pointers with addressspaces

[vchuravy]

Summary

IIUC we can't emit memcmp between pointers in addressspaces,
doing so will trigger an assertion since the signature of the memcmp
will not match it's arguments (https://bugs.llvm.org/show_bug.cgi?id=48661).

This PR disables the attempt to merge icmps,
when the pointer is in an addressspace.

[MaskRay]

llvm/test/Transforms/MergeICmps/addressspaces.ll

2

Add a file-level header about the purpose.

8

Nit: Use named arguments and values. You could use gep-references-bb.ll as an example. I took some time to write meaningful names.

47

Nit: Replace the labels with more meaningful ones.

64

Delete unused metadata.

[MaskRay]

For such negative tests, they can easily get stale. Adding them into an existing positive test and tweaking the existing test may be a good way to prevent bit rot, but I don't know which existing test is suitable.

[vtjnash]

LGTM, modulo comments from previous reviewer

[vchuravy]

Improve and simplify test

[arichardson]

llvm/lib/Transforms/Scalar/MergeICmps.cpp

157

In the CHERI-enabled targets (for Aarch64,RISC-V and MIPS) memcmp takes addrspace(200) arguments. Would it be possible to use the memcmp signature instead of a hardcoded AS0 check?

[vchuravy]

llvm/lib/Transforms/Scalar/MergeICmps.cpp

157

Possibly, could you send me a reproducer/test case? I couldn't find any example in the test suite for memcmp with different AS.

llvm/lib/Transforms/Scalar/MergeICmps.cpp

if (LoadI->isUsedOutsideOfBlock(LoadI->getParent())) {
     return {};
   }
   // Do not optimize atomic loads to non-atomic memcmp
   if (!LoadI->isSimple()) {
     LLVM_DEBUG(dbgs() << "volatile or atomic\n");
     return {};
   }
   Value *const Addr = LoadI->getOperand(0);
+  if (Addr->getType()->getPointerAddressSpace() != 0) {
+    LLVM_DEBUG(dbgs() << "from non-zero AddressSpace\n");
+    return {};
+  }
   auto *const GEP = dyn_cast<GetElementPtrInst>(Addr);
   if (!GEP)
     return {};
   LLVM_DEBUG(dbgs() << "GEP\n");
   if (GEP->isUsedOutsideOfBlock(LoadI->getParent())) {
     LLVM_DEBUG(dbgs() << "used outside of block\n");
     return {};
   }

llvm/test/Transforms/MergeICmps/addressspaces.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -mergeicmps -S | FileCheck %s
+
+source_filename = "=="
+target datalayout = "e-m:e-i64:64-n32:64"
+target triple = "powerpc64le-unknown-linux-gnu"
+
+define void @juliaAS([2 x [5 x i64]] addrspace(11)* nocapture nonnull readonly align 8 dereferenceable(80) %0, [2 x [5 x i64]] addrspace(11)* nocapture nonnull readonly align 8 dereferenceable(80) %1) {
+; CHECK-LABEL: @juliaAS(
+; CHECK-NEXT:  top:
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* [[TMP0:%.*]], i64 0, i64 1, i64 2
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* [[TMP0]], i64 0, i64 1, i64 3
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* [[TMP0]], i64 0, i64 1, i64 4
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* [[TMP1:%.*]], i64 0, i64 1, i64 2
+; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* [[TMP1]], i64 0, i64 1, i64 3
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* [[TMP1]], i64 0, i64 1, i64 4
+; CHECK-NEXT:    [[TMP8:%.*]] = load i64, i64 addrspace(11)* [[TMP2]], align 8
+; CHECK-NEXT:    [[TMP9:%.*]] = load i64, i64 addrspace(11)* [[TMP5]], align 8
+; CHECK-NEXT:    [[DOTNOT17:%.*]] = icmp eq i64 [[TMP8]], [[TMP9]]
+; CHECK-NEXT:    br i1 [[DOTNOT17]], label [[L70:%.*]], label [[L90:%.*]]
+; CHECK:       L70:
+; CHECK-NEXT:    [[TMP10:%.*]] = load i64, i64 addrspace(11)* [[TMP3]], align 8
+; CHECK-NEXT:    [[TMP11:%.*]] = load i64, i64 addrspace(11)* [[TMP6]], align 8
+; CHECK-NEXT:    [[DOTNOT18:%.*]] = icmp eq i64 [[TMP10]], [[TMP11]]
+; CHECK-NEXT:    br i1 [[DOTNOT18]], label [[L74:%.*]], label [[L90]]
+; CHECK:       L74:
+; CHECK-NEXT:    [[TMP12:%.*]] = load i64, i64 addrspace(11)* [[TMP4]], align 8
+; CHECK-NEXT:    [[TMP13:%.*]] = load i64, i64 addrspace(11)* [[TMP7]], align 8
+; CHECK-NEXT:    [[DOTNOT19:%.*]] = icmp eq i64 [[TMP12]], [[TMP13]]
+; CHECK-NEXT:    br label [[L90]]
+; CHECK:       L90:
+; CHECK-NEXT:    [[VALUE_PHI2_OFF0:%.*]] = phi i1 [ false, [[TOP:%.*]] ], [ [[DOTNOT19]], [[L74]] ], [ false, [[L70]] ]
+; CHECK-NEXT:    ret void
+;
+top:
+  %2 = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* %0, i64 0, i64 1, i64 2
+  %3 = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* %0, i64 0, i64 1, i64 3
+  %4 = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* %0, i64 0, i64 1, i64 4
+  %5 = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* %1, i64 0, i64 1, i64 2
+  %6 = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* %1, i64 0, i64 1, i64 3
+  %7 = getelementptr inbounds [2 x [5 x i64]], [2 x [5 x i64]] addrspace(11)* %1, i64 0, i64 1, i64 4
+  %8 = load i64, i64 addrspace(11)* %2, align 8
+  %9 = load i64, i64 addrspace(11)* %5, align 8
+  %.not17 = icmp eq i64 %8, %9
+  br i1 %.not17, label %L70, label %L90
+
+L70:                                              ; preds = %top
+  %10 = load i64, i64 addrspace(11)* %3, align 8
+  %11 = load i64, i64 addrspace(11)* %6, align 8
+  %.not18 = icmp eq i64 %10, %11
+  br i1 %.not18, label %L74, label %L90
+
+L74:                                              ; preds = %L70
+  %12 = load i64, i64 addrspace(11)* %4, align 8
+  %13 = load i64, i64 addrspace(11)* %7, align 8
+  %.not19 = icmp eq i64 %12, %13
+  br label %L90
+
+L90:                                              ; preds = %L74, %L70, %top
+  %value_phi2.off0 = phi i1 [ false, %top ], [ %.not19, %L74 ], [ false, %L70 ]
+  ret void
+}
+
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 1, !"Debug Info Version", i32 3}
+