[AArch64] no-FP ABI check blames caller's declaration

[jroelofs]

$ cat >vaargs.c <<EOF
#include <stdarg.h>
#include <stdbool.h>

double takes_double(double);

static bool takes_va_list_ptr(va_list * ap) {
  double val = va_arg(*ap, double);
  return takes_double(val);
}

int main() {
  va_list ap;
  takes_va_list_ptr(&ap);
}
EOF
$ ./bin/clang vaargs.c -c -o /dev/null -march=armv8-a+nofp --target=arm64-apple-ios
vaargs.c:6:13: error: 'takes_va_list_ptr' requires 'double' type support, but ABI 'darwinpcs' does not support it
    6 | static bool takes_va_list_ptr(va_list * ap) {
      |             ^
vaargs.c:6:13: error: 'takes_va_list_ptr' requires 'double' type support, but ABI 'darwinpcs' does not support it
2 errors generated.

It would be less confusing if the diagnostic pointed at the call site for takes_double, with a note pointing at takes_double's declaration.

[jroelofs]

cc @ostannard

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Jon Roelofs (jroelofs)

``` $ cat >vaargs.c <<EOF #include <stdarg.h> #include <stdbool.h>

double takes_double(double);

static bool takes_va_list_ptr(va_list * ap) {
double val = va_arg(*ap, double);
return takes_double(val);
}

float foo(float);

int main() {
va_list ap;
takes_va_list_ptr(&ap);
}
EOF
$ ./bin/clang vaargs.c -c -o /dev/null -march=armv8-a+nofp --target=arm64-apple-ios
vaargs.c:6:13: error: 'takes_va_list_ptr' requires 'double' type support, but ABI 'darwinpcs' does not support it
6 | static bool takes_va_list_ptr(va_list * ap) {
| ^
vaargs.c:6:13: error: 'takes_va_list_ptr' requires 'double' type support, but ABI 'darwinpcs' does not support it
2 errors generated.

It would be less confusing if the diagnostic pointed at the call site for `takes_double`, with a note pointing at `takes_double`'s declaration.
</details>

[jroelofs]

diff --git a/clang/lib/CodeGen/Targets/AArch64.cpp b/clang/lib/CodeGen/Targets/AArch64.cpp
index 800a55e1b962..5b6c1e98b043 100644
--- a/clang/lib/CodeGen/Targets/AArch64.cpp
+++ b/clang/lib/CodeGen/Targets/AArch64.cpp
@@ -931,11 +931,11 @@ void AArch64TargetCodeGenInfo::checkFunctionCallABISoftFloat(
     return;
 
   diagnoseIfNeedsFPReg(CGM.getDiags(), TI.getABI(), ABIInfo, ReturnType,
-                       Caller);
+                       Callee);
 
   for (const CallArg &Arg : Args)
     diagnoseIfNeedsFPReg(CGM.getDiags(), TI.getABI(), ABIInfo, Arg.getType(),
-                         Caller);
+                         Callee);
 }
 
 void AArch64TargetCodeGenInfo::checkFunctionCallABI(CodeGenModule &CGM,

I'll put up a PR with tests tomorrow.

[porglezomp]

     diagnoseIfNeedsFPReg(CGM.getDiags(), TI.getABI(), ABIInfo, Arg.getType(),
-                         Caller);
+                         Callee);

Hm, that patch doesn't look right—we want to diagnose the caller—the issue is just that Caller is the function declaration not the call-site, right?

[porglezomp]

The actual point I came here to make: This should probably be backported to 19.x because this is a new diagnostic in 19.x from @ostannard, 1fd196c.