https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91758
--- Comment #3 from Martin Liška <marxin at gcc dot gnu.org> ---
So that's what I have:
1) reduced LLVM test-case:
$ cat /tmp/llvm-project/clang/test/CodeGen/lanai-regparm.c
void f2(int a) __attribute((regparm(0)));
void f0() {
f2(1);
}
2) I applied the following local patch to make the change smaller:
diff --git a/clang/include/clang/CodeGen/CGFunctionInfo.h
b/clang/include/clang/CodeGen/CGFunctionInfo.h
index 1f81072e23d..ef7f7410c9b 100644
--- a/clang/include/clang/CodeGen/CGFunctionInfo.h
+++ b/clang/include/clang/CodeGen/CGFunctionInfo.h
@@ -87,14 +87,14 @@ private:
unsigned AllocaFieldIndex; // isInAlloca()
};
Kind TheKind;
- bool PaddingInReg : 1;
- bool InAllocaSRet : 1; // isInAlloca()
- bool IndirectByVal : 1; // isIndirect()
- bool IndirectRealign : 1; // isIndirect()
- bool SRetAfterThis : 1; // isIndirect()
+ bool PaddingInReg;
+ bool InAllocaSRet; // isInAlloca()
+ bool IndirectByVal; // isIndirect()
+ bool CanBeFlattened; // isDirect()
+ bool SignExt ; // isExtend()
+ bool SRetAfterThis ; // isIndirect()
+ bool IndirectRealign: 1; // isIndirect()
bool InReg : 1; // isDirect() || isExtend() || isIndirect()
- bool CanBeFlattened: 1; // isDirect()
- bool SignExt : 1; // isExtend()
bool canHavePaddingType() const {
return isDirect() || isExtend() || isIndirect() || isExpand();
diff --git a/clang/lib/CodeGen/TargetInfo.cpp
b/clang/lib/CodeGen/TargetInfo.cpp
index f2696a33cfb..b115d9da145 100644
--- a/clang/lib/CodeGen/TargetInfo.cpp
+++ b/clang/lib/CodeGen/TargetInfo.cpp
@@ -7567,7 +7567,10 @@ public:
if (!getCXXABI().classifyReturnType(FI))
FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
for (auto &I : FI.arguments())
+ {
I.info = classifyArgumentType(I.type, State);
+ __builtin_printf ("after: I.info: %d\n", I.info.getInReg ());
+ }
}
ABIArgInfo getIndirectResult(QualType Ty, bool ByVal, CCState &State) const;
@@ -7658,7 +7661,9 @@ ABIArgInfo LanaiABIInfo::classifyArgumentType(QualType
Ty,
}
if (InReg)
return ABIArgInfo::getDirectInReg();
- return ABIArgInfo::getDirect();
+ ABIArgInfo ret = ABIArgInfo::getDirect();
+ __builtin_printf ("before: %d\n", ret.getInReg ());
+ return ret;
}
namespace {
3) I see the problematic file is:
/tmp/llvm-project/clang/lib/CodeGen/TargetInfo.cpp
4) I took the patch from r261089 and applied it to r255894 and it still fails
5) apparently one needs -O3 to expose the issue
6) If I add following dbg_cnt:
diff --git a/gcc/dbgcnt.def b/gcc/dbgcnt.def
index 0421fae7bc0..3830666dc6c 100644
--- a/gcc/dbgcnt.def
+++ b/gcc/dbgcnt.def
@@ -195,3 +195,4 @@ DEBUG_COUNTER (tree_sra)
DEBUG_COUNTER (vect_loop)
DEBUG_COUNTER (vect_slp)
DEBUG_COUNTER (dom_unreachable_edges)
+DEBUG_COUNTER (store_merging)
diff --git a/gcc/gimple-ssa-store-merging.c b/gcc/gimple-ssa-store-merging.c
index 3c63e75fcf6..2369fd4bf5d 100644
--- a/gcc/gimple-ssa-store-merging.c
+++ b/gcc/gimple-ssa-store-merging.c
@@ -166,6 +166,7 @@
#include "rtl.h"
#include "expr.h" /* For get_bit_range. */
#include "optabs-tree.h"
+#include "dbgcnt.h"
#include "selftest.h"
/* The maximum size (in bits) of the stores this pass should generate. */
@@ -3898,7 +3899,8 @@ imm_store_chain_info::output_merged_stores ()
bool ret = false;
FOR_EACH_VEC_ELT (m_merged_store_groups, i, merged_store)
{
- if (output_merged_store (merged_store))
+ if (dbg_cnt (store_merging)
+ && output_merged_store (merged_store))
{
unsigned int j;
store_immediate_info *store;
I can bisect that to one store merging transformation:
before:
MEM[(struct SmallVectorBase *)&Elements].Size = 0;
MEM[(struct SmallVectorBase *)&Elements].Capacity = 3;
after:
MEM[(unsigned int *)&Elements + 8B] = 12884901888;
if (SizeInRegs_144 > 3)
The transformation looks fine to me and it must be an issue in RTL, because
this is the only difference I see in tree optimized dump file.
The change happens in {anonymous}::ARCABIInfo::computeInfo (const struct
ARCABIInfo * const this, struct CGFunctionInfo & FI)
function. And the valgrind also points to the same function:
valgrind /tmp/llvm-project/build/bin/clang -cc1 -triple lanai-unknown-unknown
-mregparm 4 /tmp/llvm-project/clang/test/CodeGen/lanai-regparm.c -emit-llvm
==766== Memcheck, a memory error detector
==766== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et al.
==766== Using Valgrind-3.15.0 and LibVEX; rerun with -h for copyright info
==766== Command: /tmp/llvm-project/build/bin/clang -cc1 -triple
lanai-unknown-unknown -mregparm 4
/tmp/llvm-project/clang/test/CodeGen/lanai-regparm.c -emit-llvm
==766==
before: 0
==766== Conditional jump or move depends on uninitialised value(s)
==766== at 0x7295287: __vfprintf_internal (vfprintf-internal.c:1644)
==766== by 0x7280C7A: printf (printf.c:33)
==766== by 0x2776A28: (anonymous
namespace)::LanaiABIInfo::computeInfo(clang::CodeGen::CGFunctionInfo&) const
(in /tmp/llvm-project/build/bin/clang-10)
==766== by 0x27F7533:
clang::CodeGen::CodeGenTypes::arrangeLLVMFunctionInfo(clang::CanQual<clang::Type>,
bool, bool, llvm::ArrayRef<clang::CanQual<clang::Type> >,
clang::FunctionType::ExtInfo,
llvm::ArrayRef<clang::FunctionType::ExtParameterInfo>,
clang::CodeGen::RequiredArgs) (in /tmp/llvm-project/build/bin/clang-10)
==766== by 0x27F9D94:
clang::CodeGen::CodeGenTypes::arrangeFreeFunctionType(clang::CanQual<clang::FunctionProtoType>)
(in /tmp/llvm-project/build/bin/clang-10)
==766== by 0x272AAC9:
clang::CodeGen::CodeGenTypes::ConvertFunctionTypeInternal(clang::QualType) (in
/tmp/llvm-project/build/bin/clang-10)
==766== by 0x272BD45:
clang::CodeGen::CodeGenTypes::ConvertType(clang::QualType) (in
/tmp/llvm-project/build/bin/clang-10)
==766== by 0x2701D5A:
clang::CodeGen::CodeGenModule::GetAddrOfFunction(clang::GlobalDecl,
llvm::Type*, bool, bool, clang::CodeGen::ForDefinition_t) (in
/tmp/llvm-project/build/bin/clang-10)
==766== by 0x28472E2:
EmitFunctionDeclPointer(clang::CodeGen::CodeGenModule&, clang::FunctionDecl
const*) (in /tmp/llvm-project/build/bin/clang-10)
==766== by 0x2861F2C:
clang::CodeGen::CodeGenFunction::EmitCallee(clang::Expr const*) (in
/tmp/llvm-project/build/bin/clang-10)
==766== by 0x28621F2:
clang::CodeGen::CodeGenFunction::EmitCallExpr(clang::CallExpr const*,
clang::CodeGen::ReturnValueSlot) (in /tmp/llvm-project/build/bin/clang-10)
==766== by 0x2897478: (anonymous
namespace)::ScalarExprEmitter::VisitCallExpr(clang::CallExpr const*) (in
/tmp/llvm-project/build/bin/clang-10)
7) The wrong output is then:
marxin@marxinbox:/tmp/llvm-project/build> /tmp/llvm-project/build/bin/clang
-cc1 -triple lanai-unknown-unknown -mregparm 4
/tmp/llvm-project/clang/test/CodeGen/lanai-regparm.c -emit-llvm
before: 0
after: I.info: 1
marxin@marxinbox:/tmp/llvm-project/build> /tmp/llvm-project/build/bin/clang
-cc1 -triple lanai-unknown-unknown -mregparm 4
/tmp/llvm-project/clang/test/CodeGen/lanai-regparm.c -emit-llvm
before: 0
after: I.info: 0
So I.info.getInReg() is sometimes 0 and sometimes 1.
I'm going to carry on ..
