commit c04a05d898982614a2df80d928b97ed4f8c49b60 broke mcmodel=medium on windows

[erik895]

int foo = 32;

int* go()
{
    return &foo;
}

compiled with clang on windows:

clang.exe -mcmodel=medium -Wl,/SUBSYSTEM:CONSOLE -o a.exe -nostdlib -Wl,-entry:go test.cpp

used to generate:

llvm-objdump.exe  -d a.exe

Disassembly of section .text:

0000000140001000 <.text>:
140001000: 48 b8 00 30 00 40 01 00 00 00        movabsq $0x140003000, %rax # imm = 0x140003000
14000100a: c3                           retq

After commit c04a05d
more specific this chunk:

-  CodeModel::Model M = getTargetMachine().getCodeModel();
+  // The following OpFlags under RIP-rel PIC use RIP.
   if (Subtarget.isPICStyleRIPRel() &&
-      (M == CodeModel::Small || M == CodeModel::Kernel))
+      (OpFlags == X86II::MO_NO_FLAG || OpFlags == X86II::MO_COFFSTUB ||
+       OpFlags == X86II::MO_DLLIMPORT))
     return X86ISD::WrapperRIP;

The output changes to

0000000140001000 <.text>:
140001000: 48 8d 05 f9 1f 00 00         leaq    0x1ff9(%rip), %rax      # 0x140003000
140001007: c3                           retq

I think its wrong to generate rip relative with mcmodel=medium.

[nikic]

cc @aeubanks

[aeubanks]

I don't think the x86-64 medium code model is well defined on windows (correct me if I'm wrong), but I'm not sure what's wrong with having it follow the ELF medium code model where it respects -mlarge-data-threshold (default 2^16) and use rip-relative to access "small" globals. -mlarge-data-threshold=0 will make all globals considered "large".

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (erik895)

``` int foo = 32;

int* go()
{
return &foo;
}

compiled with clang on windows:

clang.exe -mcmodel=medium -Wl,/SUBSYSTEM:CONSOLE -o a.exe -nostdlib -Wl,-entry:go test.cpp

used to generate:

llvm-objdump.exe -d a.exe

Disassembly of section .text:

0000000140001000 <.text>:
140001000: 48 b8 00 30 00 40 01 00 00 00 movabsq $0x140003000, %rax # imm = 0x140003000
14000100a: c3 retq

After commit c04a05d898982614a2df80d928b97ed4f8c49b60 
  more specific this chunk:

• CodeModel::Model M = getTargetMachine().getCodeModel();

• // The following OpFlags under RIP-rel PIC use RIP.
  if (Subtarget.isPICStyleRIPRel() &&

•  (M == CodeModel::Small || M == CodeModel::Kernel))
  

•  (OpFlags == X86II::MO_NO_FLAG || OpFlags == X86II::MO_COFFSTUB ||
  

•   OpFlags == X86II::MO_DLLIMPORT))
  

  return X86ISD::WrapperRIP;

The output changes to

0000000140001000 <.text>:
140001000: 48 8d 05 f9 1f 00 00 leaq 0x1ff9(%rip), %rax # 0x140003000
140001007: c3 retq

I think its wrong to generate rip relative with mcmodel=medium.
</details>

[erik895]

Same behaviour as ELF sounds good to me. Problem now is I cant get medium to not use rip even if I add -mlarge-data-threshold=0.

Maybe I blame the wrong commit but thats where it stopped working for me (I did a git bisect)

Here's an updated test output:

C:\temp>go.bat 

echo on

type test.cpp
int foo = 32;

int* go()
{
    return &foo;
}
clang.exe --version
clang version 19.1.0
Target: x86_64-pc-windows-msvc
Thread model: posix
InstalledDir: C:\Program Files\LLVM\bin

clang.exe -mcmodel=medium -mlarge-data-threshold=0 -Wl,/SUBSYSTEM:CONSOLE -o a.exe -nostdlib -Wl,-entry:go test.cpp

llvm-objdump.exe -d a.exe 

a.exe:  file format coff-x86-64

Disassembly of section .text:

0000000140001000 <.text>:
140001000: 48 8d 05 f9 1f 00 00         leaq    0x1ff9(%rip), %rax      # 0x140003000
140001007: c3                           retq
C:\temp>