[Scalarizer] Test *_with_overflow crash with min-bits=32 and min-bits=16

[farzonl]

first reported here: #127520
introduced by: #126815

; RUN: opt %s -passes='function(scalarizer<load-store;min-bits=16>,dce)' -S | FileCheck %s --check-prefixes=CHECK,MIN16
; RUN: opt %s -passes='function(scalarizer<load-store;min-bits=32>,dce)' -S | FileCheck %s --check-prefixes=CHECK,MIN32

define <3 x i32> @call_v3i32(<3 x i32> %a, <3 x i32> %b) {
; CHECK-LABEL: @call_v3i32(
; CHECK-NEXT:    [[T:%.*]] = call { <3 x i32>, <3 x i1> } @llvm.uadd.with.overflow.v3i32(<3 x i32> [[A:%.*]], <3 x i32> [[B:%.*]])
; CHECK-NEXT:    [[R:%.*]] = extractvalue { <3 x i32>, <3 x i1> } [[T]], 0
; CHECK-NEXT:    ret <3 x i32> [[R]]
;
  %t = call { <3 x i32>, <3 x i1> } @llvm.uadd.with.overflow.v3i32(<3 x i32> %a, <3 x i32> %b)
  %r = extractvalue { <3 x i32>, <3 x i1> } %t, 0
  ret <3 x i32> %r
}

opt: include/llvm/Support/Casting.h:578: decltype(auto) llvm::cast(From *) [To = llvm::VectorType, From = llvm::Type]: Assertion `isa<To>(Val) && "cast<Ty>() argument of incompatible type!"' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.	Program arguments: opt -passes=function(scalarizer<load-store;min-bits=16>,dce) -S
1.	Running pass "function(scalarizer,dce)" on module "<stdin>"
2.	Running pass "scalarizer" on function "call_v3i32"
 #0 0x0000000004028908 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (opt+0x4028908)
 #1 0x000000000402643e llvm::sys::RunSignalHandlers() (opt+0x402643e)
 #2 0x0000000004029131 SignalHandler(int, siginfo_t*, void*) Signals.cpp:0:0
 #3 0x000075c457a45330 (/lib/x86_64-linux-gnu/libc.so.6+0x45330)
 #4 0x000075c457a9eb2c __pthread_kill_implementation ./nptl/pthread_kill.c:44:76
 #5 0x000075c457a9eb2c __pthread_kill_internal ./nptl/pthread_kill.c:78:10
 #6 0x000075c457a9eb2c pthread_kill ./nptl/pthread_kill.c:89:10
 #7 0x000075c457a4527e raise ./signal/../sysdeps/posix/raise.c:27:6
 #8 0x000075c457a288ff abort ./stdlib/abort.c:81:7
 #9 0x000075c457a2881b _nl_load_domain ./intl/loadmsgcat.c:1177:9
#10 0x000075c457a3b517 (/lib/x86_64-linux-gnu/libc.so.6+0x3b517)
#11 0x0000000004143bea (opt+0x4143bea)
#12 0x00000000049addf4 (anonymous namespace)::Scatterer::operator[](unsigned int) Scalarizer.cpp:0:0
#13 0x00000000049aa142 (anonymous namespace)::ScalarizerVisitor::visit(llvm::Function&) Scalarizer.cpp:0:0
#14 0x00000000049a5419 llvm::ScalarizerPass::run(llvm::Function&, llvm::AnalysisManager<llvm::Function>&) (opt+0x49a5419)

No line numbers, Looks like we have some missing debug symbols I suspect the assert is coming from ScalarizerVisitor::visitExtractValueInst

There is a use of Scatterer::operator[] as part of creating the ExtractValue IR.

for (unsigned OpIdx = 0; OpIdx < Op0.size(); ++OpIdx) {
    Value *ResElem = Builder.CreateExtractValue(
        Op0[OpIdx], Index, EVI.getName() + ".elem" + Twine(Index));
    Res.push_back(ResElem);
  }

We likely need to check VS.NumPacked before calling Scatterer::operator[]

Because that operator is using Frag * VS.NumPacked to create the exrtractElement instruction.

 CV[Frag] = Builder.CreateExtractElement(V, Frag * VS.NumPacked,
                                            V->getName() + ".i" + Twine(Frag))

[farzonl]

@Icohedron can you look into this issue. I'm also noticing a second potential issue:

call { <3 x i32>, <3 x i1> }

The struct return elements are not the same type. one is vec3 of i32 one is and one is vec3 of i1. you need to add the *_with_overflow intrinsics to isVectorIntrinsicWithStructReturnOverloadAtField if these types are not the same.

[Icohedron]

I'm on it

[jayfoad]

As mentioned in #127996, note that you can also provoke this bug by adding min-bits=64 to test/Transforms/Scalarizer/frexp.ll.

[Icohedron]

@Icohedron can you look into this issue. I'm also noticing a second potential issue:

call { <3 x i32>, <3 x i1> }

The struct return elements are not the same type. one is vec3 of i32 one is and one is vec3 of i1. you need to add the *_with_overflow intrinsics to isVectorIntrinsicWithStructReturnOverloadAtField if these types are not the same.

isVectorIntrinsicWithStructReturnOverloadAtField appears to only be for intrinsic name-mangling when there is more than one possible type for the the 2nd element in the struct return type (see splitCall and getOrInsertDeclaration). For the *_with_overflow intrinsics, the 2nd element of the return struct is always i1, so it does not need to be added to isVectorIntrinsicWithStructReturnOverloadAtField.

Adding the *_with_overflow intrinsic to isVectorIntrinsicWithStructReturnOverloadAtField results in a broken LLVM module.

Intrinsic name not mangled correctly for type arguments! Should be: llvm.uadd.with.overflow.i32
ptr @llvm.uadd.with.overflow.i32.i1
LLVM ERROR: Broken module found, compilation aborted!