Variable not visible at Og

[FiorellaArtuso]

Bugzilla Link	51559
Version	trunk
OS	Linux
CC	@dwblaikie,@JDevlieghere,@jmorse,@walkerkd,@pogo59

Extended Description

Comment:
When stepping on line 4 the variable g_3012 is not visible.

Steps to reproduce bug:

root@e60e0a2bba1a:/home/stepping/output# clang -v
clang version 14.0.0 (https://github.com/llvm/llvm-project.git 9573343)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /usr/local/bin
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/7
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/7.5.0
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/8
Selected GCC installation: /usr/lib/gcc/x86_64-linux-gnu/7.5.0
Candidate multilib: .;@m64
Selected multilib: .;@m64

root@e60e0a2bba1a:/home/stepping/output# lldb -v
lldb version 14.0.0 (https://github.com/llvm/llvm-project.git revision 9573343)
clang revision 9573343
llvm revision 9573343

root@e60e0a2bba1a:/home/stepping/output# cat a.c
int g_97 ;
static short g_3012 ;
void func_30() {
g_97 = g_3012;
}
int main() {
{
func_30();
}
}

root@e60e0a2bba1a:/home/stepping/output# cat -n a.c
1 int g_97 ;
2 static short g_3012 ;
3 void func_30() {
4 g_97 = g_3012;
5 }
6 int main() {
7 {
8 func_30();
9 }
10 }

root@e60e0a2bba1a:/home/stepping/output# clang -g -Og a.c -o opt

root@e60e0a2bba1a:/home/stepping/output# lldb opt
(lldb) target create "opt"
Current executable set to '/home/stepping/output/opt' (x86_64).
(lldb) b main
Breakpoint 1: where = opt`main [inlined] func_30 at a.c:4:15, address = 0x0000000000400490
(lldb) r
Process 68731 launched: '/home/stepping/output/opt' (x86_64)
Process 68731 stopped

• thread #​1, name = 'opt', stop reason = breakpoint 1.1
  frame #​0: 0x0000000000400490 opt`main [inlined] func_30 at a.c:4:15
  1 int g_97 ;
  2 static short g_3012 ;
  3 void func_30() {
  -> 4 g_97 = g_3012;
  5 }
  6 int main() {
  7 {
  (lldb) p g_3012
  error: expression failed to parse:
  error: <user expression 0>:1:1: use of undeclared identifier 'g_3012'
  g_3012

[dwblaikie]

Looks like IPSCCPPass removes the global variable & doesn't create a constant expression for the debug representation (DIGlobalVariableExpression) representing the value.

[llvmbot]

Yes, this is a bug in ISCCP as Dave said.

[hashemthomas1]

I recently encountered a crash which was a result of the patch that closed this issue.

The crash happens in the function getExpressionForConstant that SCCP pass utilizes, when reaching the following code section that deals with floating point values.

if (Ty.isFloatTy() || Ty.isDoubleTy()) {
    const APFloat &APF = cast<ConstantFP>(&C)->getValueAPF();
    return DIB.createConstantValueExpression(
        APF.bitcastToAPInt().getZExtValue());
  }

It specifically happens in the cast operation when the value of the float is undef.

A very simple OpenCL reproducer of the crash with x86 architecture: https://godbolt.org/z/sTqaW6MvE

[Endilll]

I mentioned the comment above in #66745

[fhahn]

Fixed the crash in b46638d. @CarlosAlbertoEnciso please could you check if there's more appropriate debug data for undef/poison. Also, the original code only checked for float and double FP types, but there are others, like half. Could you extend the test coverage to include those types as well?

[CarlosAlbertoEnciso]

@fhahn Thanks for fixing the crash. I just returned this week from PTO.

[CarlosAlbertoEnciso]

Uploaded a patch to extend the test coverage for half, bfloat and undef.
#77901