[GVN] Invalidate ICF cache when clearing the instructions

[xgupta]

This fixes #48805

/home/user/llvm-project/llvm/lib/Analysis/InstructionPrecedenceTracking.cpp:93: void llvm::InstructionPrecedenceTracking::validate(const llvm::BasicBlock *) const: Assertion `It->second == &Insn && "Cached first special instruction is wrong!"' failed.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Changes

This fixes #48805

/home/user/llvm-project/llvm/lib/Analysis/InstructionPrecedenceTracking.cpp:93: void llvm::InstructionPrecedenceTracking::validate(const llvm::BasicBlock *) const: Assertion `It->second == &Insn && "Cached first special instruction is wrong!"' failed.

---

Full diff: https://github.com/llvm/llvm-project/pull/68145.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Scalar/GVN.cpp (+1)

diff --git a/llvm/lib/Transforms/Scalar/GVN.cpp b/llvm/lib/Transforms/Scalar/GVN.cpp
index bc54846ccf0ad2d..f7a905c2e13c4d4 100644
--- a/llvm/lib/Transforms/Scalar/GVN.cpp
+++ b/llvm/lib/Transforms/Scalar/GVN.cpp
@@ -2799,6 +2799,7 @@ bool GVNPass::processBlock(BasicBlock *BB) {
       salvageDebugInfo(*I);
       removeInstruction(I);
     }
+    ICF->clear();
     InstrsToErase.clear();
 
     if (AtStart)

[nikic]

Clearing ICF as a whole is not the correct way to fix this. It needs to be incrementally invalidated, as other parts of GVN already do. The first step would be to provide a test case for the assertion failure.

[w2yehia]

We hit this on AIX while building python with -flto. Here's a reduced testcase:

; ModuleID = 'reduced.bc'
target datalayout = "E-m:a-Fi64-i64:64-n32:64-S128-v256:256:256-v512:512:512"
target triple = "powerpc64-ibm-aix7.3.0.0"

%struct.PyGetSetDef = type { ptr, ptr, ptr, ptr, ptr }
%struct.anon.28 = type { ptr, ptr, ptr, ptr, ptr, ptr }

@_PyMem_Raw = external global %struct.PyGetSetDef
@__profc_PyMem_SetAllocator = external global [5 x i64]
@__profc_PyMem_RawFree = external global [1 x i64]

define i32 @_PyMem_SetDefaultAllocator() #0 {
entry:
  %call = call fastcc i32 @pymem_set_default_allocator()
  ret i32 0
}

define fastcc i32 @pymem_set_default_allocator() #0 {
entry:
  store ptr @_PyMem_RawFree, ptr getelementptr inbounds (%struct.PyGetSetDef, ptr @_PyMem_Raw, i64 0, i32 4), align 8
  %pgocount6.i31 = load i64, ptr getelementptr inbounds ([5 x i64], ptr @__profc_PyMem_SetAllocator, i64 0, i64 1), align 8
  store i64 %pgocount6.i31, ptr @__profc_PyMem_SetAllocator, align 8
  ret i32 0
}

; Function Attrs: nounwind willreturn
declare void @_PyMem_RawFree() #1

define void @PyMem_RawFree(ptr %ptr) #0 {
entry:
  %pgocount = load i64, ptr %ptr, align 8
  %0 = or i64 %pgocount, 1
  store i64 %0, ptr @__profc_PyMem_RawFree, align 8
  %1 = load ptr, ptr getelementptr inbounds (%struct.PyGetSetDef, ptr @_PyMem_Raw, i64 0, i32 4), align 8
  tail call void %1(ptr null, ptr null)
  ret void
}

declare fastcc void @pathconfig_set_from_config(ptr nocapture)

define void @_PyConfig_InitPathConfig(i1 %cmp.not.i) #0 {
entry:
  %pathconfig.i11 = alloca [0 x [0 x %struct.anon.28]], i32 0, align 8
  %prefix.i.i.i.i = getelementptr %struct.anon.28, ptr %pathconfig.i11, i64 0, i32 1
  call fastcc void @pathconfig_set_from_config(ptr %pathconfig.i11)
  br i1 %cmp.not.i, label %if.end.i, label %config_calculate_pathconfig.exit

if.end.i:                                         ; preds = %entry
  %0 = load ptr, ptr %pathconfig.i11, align 8
  %cmp.not.i.i = icmp eq ptr %0, null
  %1 = load ptr, ptr %prefix.i.i.i.i, align 8
  %cmp.not.i66.i = icmp eq ptr %1, null
  %or.cond = select i1 %cmp.not.i.i, i1 %cmp.not.i66.i, i1 false
  br i1 %or.cond, label %config_calculate_pathconfig.exit, label %common.ret

common.ret:                                       ; preds = %config_calculate_pathconfig.exit, %if.end.i
  ret void

config_calculate_pathconfig.exit:                 ; preds = %if.end.i, %entry
  %call.i89.i = call i32 @_PyMem_SetDefaultAllocator()
  %2 = load ptr, ptr %pathconfig.i11, align 8
  call void @PyMem_RawFree(ptr %2)
  call void @PyMem_RawFree(ptr %prefix.i.i.i.i)
  br label %common.ret
}

attributes #0 = { "target-cpu"="pwr7" "target-features"="+altivec,+bpermd,+extdiv,+isa-v206-instructions,+vsx,-aix-small-local-exec-tls,-crbits,-crypto,-direct-move,-htm,-isa-v207-instructions,-isa-v30-instructions,-power8-vector,-power9-vector,-privileged,-quadword-atomics,-rop-protect,-spe" }
attributes #1 = { nounwind willreturn }

[xgupta]

Thanks, @w2yehia, Can you also share how you got this test case for future purposes, I mean how do you run llvm-reduce for lto issue?

[nikic]

Somewhat reduced test case for -passes=gvn:

declare void @_PyMem_RawFree() nounwind willreturn

define i64 @test(ptr %p) {
entry:
  %a = alloca [2 x ptr], align 8
  %a2 = getelementptr ptr, ptr %a, i64 1
  call void null(ptr %a)
  br i1 false, label %if, label %exit

if:
  %p0 = load ptr, ptr %a, align 8
  %p1 = load ptr, ptr %a2, align 8
  br label %exit

exit:
  store ptr @_PyMem_RawFree, ptr %p
  %p2 = load ptr, ptr %a, align 8
  %pgocount.i = load i64, ptr %p2, align 8
  %fn = load ptr, ptr %p
  tail call void %fn()
  %res = load i64, ptr %a2, align 8
  ret i64 %res
}

[nikic]

Fixed in 46aac94. We need to remove users from ICF before performing RAUW.

[xgupta]

Thanks, @nikic for the fix. How do you get the bc/ll file for llvm-reduce i.e. which flag is used when building python?

[w2yehia]

Thanks, @w2yehia, Can you also share how you got this test case for future purposes, I mean how do you run llvm-reduce for lto issue?

I used -save-temps to dump intermediate IR, picked the *.0.2.internalize.bc file, and tried it with opt "-passes=lto<O3>" and was able to reproduce the failure. After that I ran a bugpoint, which took few hours but was slowly making progress, so took the reduced IR, and fed into llvm-reduce which finished in few minutes.

[xgupta]

Thanks for the reply, I did set the variable for -save-temps

export CC=/home/shivam/llvm/new/llvm-project/build/bin/clang CXX=/home/shivam/llvm/new/llvm-project/build/bin/clang++ CFLAGS="-save-temps" CXXFLAGS="-save-temps" LDFLAGS="-save-temps"

And then run
./configure --with-lto --with-ensurepip=upgrade --enable-optimizations

It crashes and I checked the generated .bc file
find . -type f -name '*.bc'

There is no *.0.2.internalize.bc, only these

_collectionsmodule.bc, getplatform.bc, _operator.bc, typeobject.bc, ceval.bc, string_parser.bc, ast.bc, namespaceobject.bc, gcmodule.bc, longobject.bc, pylifecycle.bc, bootstrap_hash.bc, signalmodule.bc, ast_opt.bc, myreadline.bc, hamt.bc, rangeobject.bc, errnomodule.bc, genobject.bc, fileobject.bc, pystrcmp.bc, errors.bc, pyfpe.bc, token.bc, pytime.bc, unicodeobject.bc, methodobject.bc, getpath.bc, setobject.bc, symtablemodule.bc, bltinmodule.bc, picklebufobject.bc, odictobject.bc, _math.bc, pystrhex.bc, bytearrayobject.bc, capsule.bc, asdl.bc, structmember.bc, obmalloc.bc, pyarena.bc, moduleobject.bc, _weakref.bc, getcopyright.bc, pyctype.bc, stringio.bc, bytesobject.bc, mystrtoul.bc, _abc.bc, preconfig.bc, _threadmodule.bc, pystate.bc, context.bc, memoryobject.bc, modsupport.bc, _functoolsmodule.bc, dtoa.bc, descrobject.bc, _iomodule.bc, Python-ast.bc, tupleobject.bc, importdl.bc, atexitmodule.bc, _warnings.bc, main.bc, pymath.bc, unicodectype.bc, faulthandler.bc, symtable.bc, buff eredio.bc, accu.bc, iterobject.bc, classobject.bc, _localemodule.bc, mysnprintf.bc, bytesio.bc, frozenmain.bc, textio.bc, getbuildinfo.bc, traceback.bc, unionobject.bc, _tracemalloc.bc, config.bc, sliceobject.bc, getversion.bc, codeobject.bc, _testembed.bc, exceptions.bc, timemodule.bc, formatter_unicode.bc, thread.bc, ast_unparse.bc, python.bc, peg_api.bc, complexobject.bc, marshal.bc, dynamic_annotations.bc, floatobject.bc, cellobject.bc, pythonrun.bc, interpreteridobject.bc, enumobject.bc, compile.bc, bytes_methods.bc, dynload_shlib.bc, dictobject.bc, getopt.bc, fileutils.bc, xxsubtype.bc, codecs.bc, _sre.bc, future.bc, pegen.bc, sysmodule.bc, object.bc, boolobject.bc, structseq.bc, parser.bc, weakrefobject.bc, genericaliasobject.bc, abstract.bc, getargs.bc, initconfig.bc, _stat.bc, import.bc, pyhash.bc, itertoolsmodule.bc, posixmodule.bc, listobject.bc, suggestions.bc, getcompiler.bc, iobase.bc, hashtable.bc, _codecsmodule.bc, frozen.bc, call.bc, fileio.bc, pwdmodule.bc, frameobject.bc, pathconfig.bc, pystrtod.bc, tokenizer.bc, funcobject.bc

And none of them failed with -passes=lto running with opt.

Is this not right?

[w2yehia]

on AIX, we use libLTO.so (LTOCodegenerator) and it has the -save-temps option.
This option emits IR at intermediate steps of link-time LTO.
If you're using lld, you can pass -save-temps directly (i.e. without -plugin-opt=)
When I try it on ppclinux, I get these files for a simple hello world program:

a.out.0.0.preopt.bc  a.out.0.2.internalize.bc  a.out.0.4.opt.bc  a.out.0.5.precodegen.bc

Try a hello world testcase first, before applying it to the python link.

[w2yehia]

@nikic thanks for fixing; I couldn't find out if this was reviewed. If it wasn't, maybe we should?

[xgupta]

on AIX, we use libLTO.so (LTOCodegenerator) and it has the -save-temps option. This option emits IR at intermediate steps of link-time LTO. If you're using lld, you can pass -save-temps directly (i.e. without -plugin-opt=)

Thanks. I get this working by using -Wl,-plugin-opt=save-temps so thought to add some documentation -
#68389.