[Driver][Sparc] Default to -mcpu=v9 for 32-bit Linux/sparc64

[rorth]

While working on supporting PR #109101 on Linux/sparc64, I was reminded that clang -m32 still defaults to generating V8 code, although the 64-bit kernel requires a V9 CPU.

This patch corrects that on V9-only distros (Debian, Gentoo).

Tested on sparc64-unknown-linux-gnu, x86_64-pc-linux-gnu, sparcv9-sun-solaris2.11, and amd64-pc-solaris2.11.

A previous version of this patch was submitted as [Driver][Sparc] Default to -mcpu=v9 for SparcV8 on Linux, but got stalled for 2+ years.

[rorth]

@glaubitz

[llvmbot]

@llvm/pr-subscribers-clang
@llvm/pr-subscribers-backend-sparc

@llvm/pr-subscribers-clang-driver

Author: Rainer Orth (rorth)

Changes

While working on supporting PR #109101 on Linux/sparc64, I was reminded that clang -m32 still defaults to generating V8 code, although the 64-bit kernel requires a V9 CPU.

This patch corrects that on V9-only distros (Debian, Gentoo).

Tested on sparc64-unknown-linux-gnu, x86_64-pc-linux-gnu, sparcv9-sun-solaris2.11, and amd64-pc-solaris2.11.

A previous version of this patch was submitted as [Driver][Sparc] Default to -mcpu=v9 for SparcV8 on Linux, but got stalled for 2+ years.

---

Full diff: https://github.com/llvm/llvm-project/pull/109278.diff

3 Files Affected:

• (modified) clang/lib/Driver/ToolChains/Arch/Sparc.cpp (+4-1)
• (modified) clang/test/Preprocessor/predefined-arch-macros.c (+5-2)
• (modified) clang/test/lit.cfg.py (+4)

diff --git a/clang/lib/Driver/ToolChains/Arch/Sparc.cpp b/clang/lib/Driver/ToolChains/Arch/Sparc.cpp
index f7f0a265fef68b..44eddc7603b4bd 100644
--- a/clang/lib/Driver/ToolChains/Arch/Sparc.cpp
+++ b/clang/lib/Driver/ToolChains/Arch/Sparc.cpp
@@ -7,6 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "Sparc.h"
+#include "clang/Driver/Distro.h"
 #include "clang/Driver/Driver.h"
 #include "clang/Driver/DriverDiagnostic.h"
 #include "clang/Driver/Options.h"
@@ -125,7 +126,9 @@ std::string sparc::getSparcTargetCPU(const Driver &D, const ArgList &Args,
     return std::string(CPUName);
   }
 
-  if (Triple.getArch() == llvm::Triple::sparc && Triple.isOSSolaris())
+  Distro Dist(D.getVFS(), Triple);
+  if (Triple.getArch() == llvm::Triple::sparc &&
+      (Triple.isOSSolaris() || Dist.IsDebian() || Dist.IsGentoo()))
     return "v9";
   return "";
 }
diff --git a/clang/test/Preprocessor/predefined-arch-macros.c b/clang/test/Preprocessor/predefined-arch-macros.c
index a149c69ee0cdb2..2e576546b45c7a 100644
--- a/clang/test/Preprocessor/predefined-arch-macros.c
+++ b/clang/test/Preprocessor/predefined-arch-macros.c
@@ -4131,13 +4131,16 @@
 
 // RUN: %clang -E -dM %s -o - 2>&1 \
 // RUN:     -target sparc-unknown-linux \
-// RUN:   | FileCheck -match-full-lines %s -check-prefix=CHECK_SPARC
+// RUN:   | FileCheck -match-full-lines %s -check-prefixes=CHECK_SPARC,CHECK_SPARC%if sparc64-distro %{_V9%} %else %{_V8%}
 // CHECK_SPARC: #define __BIG_ENDIAN__ 1
 // CHECK_SPARC: #define __sparc 1
 // CHECK_SPARC: #define __sparc__ 1
 // CHECK_SPARC-NOT: #define __sparcv9 1
 // CHECK_SPARC-NOT: #define __sparcv9__ 1
-// CHECK_SPARC: #define __sparcv8 1
+// CHECK_SPARC_V8: #define __sparcv8 1
+// CHECK_SPARC_V8-NOT: #define __sparc_v9__ 1
+// CHECK_SPARC_V9: #define __sparc_v9__ 1
+// CHECK_SPARC_V9-NOT: #define __sparcv8 1
 // CHECK_SPARC-NOT: #define __sparcv9 1
 // CHECK_SPARC-NOT: #define __sparcv9__ 1
 
diff --git a/clang/test/lit.cfg.py b/clang/test/lit.cfg.py
index 92a3361ce672e2..068975fac5a7a4 100644
--- a/clang/test/lit.cfg.py
+++ b/clang/test/lit.cfg.py
@@ -224,6 +224,10 @@ def have_host_clang_repl_cuda():
         "default-cxx-stdlib={}".format(config.clang_default_cxx_stdlib)
     )
 
+# clang -m32 defaults to -mcpu=v9 on Linux/sparc64 distros.
+if re.search(r"debian|gentoo", platform.version(), re.I):
+    config.available_features.add("sparc64-distro")
+
 # As of 2011.08, crash-recovery tests still do not pass on FreeBSD.
 if platform.system() not in ["FreeBSD"]:
     config.available_features.add("crash-recovery")

[koachan]

Pardon me for the basic question but why only for Debian and Gentoo and not for all Linux in general?

[glaubitz]

There is the very popular SPARC Leon CPU that is used by ESA among other important customers and that uses V8 as the baseline. And Linux is supported and actively maintained on Leon.

[rorth]

Exactly: so far I've just included distros that I know for certain are SPARC V9 only. This can always be augmented if more come up, of course.

[MaskRay]

We try to restrict distribution differences to things like default linker options and library paths. Affecting -mcpu= seems very unintuitive. There are many Debian derivatives. It's weird that Debian uses -mcpu=v9 while others use -mcpu=v8. We should not increase IsDebian or IsGentoo use.

If -mcpu=v9 seems the right thing for the majority of configurations, we can bump -mcpu=v9 for all Linux and ask other, older systems to use https://clang.llvm.org/docs/UsersManual.html#configuration-files to specify -mcpu=v8.

[thesamesam]

(I'm fine with maskray's idea as well.)

[rorth]

We try to restrict distribution differences to things like default linker options and library paths. Affecting -mcpu= seems very unintuitive. There are many Debian derivatives. It's weird that Debian uses -mcpu=v9 while others use -mcpu=v8. We should not increase IsDebian or IsGentoo use.

Ah, thanks. That simplifies things a lot.

If -mcpu=v9 seems the right thing for the majority of configurations, we can bump -mcpu=v9 for all Linux and ask other, older systems to use https://clang.llvm.org/docs/UsersManual.html#configuration-files to specify -mcpu=v8.

I'd been operating under the assumption that clang should work out of the box on either kind of distro. I'll update the patch accordingly and add a release notes entry about the change.

[rorth]

Just to be certain: the revised version of the patch is ok to commit?

[glaubitz]

Just to be certain: the revised version of the patch is ok to commit?

I think it would be good to get an ACK from at least one of the SPARC Leon maintainers.

CC @andreas-gaisler

[MaskRay]

Just to be certain: the revised version of the patch is ok to commit?

I think it would be good to get an ACK from at least one of the SPARC Leon maintainers.

CC @andreas-gaisler

Thanks for informing the maintainer. Note: this is done as a courtesy. When we bump a setting -mcpu= for good reasons, we do not add exception code like !isVendorA() && !isVendorB(). clangDriver is moving out of business to hard code such vendor details and requires adoption of Clang configuration files.

[rorth]

Good, that considerably simplifies non-native testing:

• The original patch ignored the issue because I had no idea how to handle it.
• The first version of this one used the hack of checking uname -v. While this works, it's still wrong since this checks the host distro, not the target one.
• It occured to me that one could use the vendor component of the target triple for this purpose. Right now, only SUSE is used here, but it could be extended to Debian, Gentoo and others if need be. AFAICS, the vendor field is usually just ignored.