[CHERIoT] Fix divergence between frontend and backend as to whether CHERIoT counts as EABI.

It was always intended to be treated as such, but when upstream added support for EABI, we
missed updating the check in the frontend to handle CHERIoT as well.

This manifests as a bug only in one very specific scenario: clang maintains compatibility
with non-conforming versions of GCC that do not properly align doubles on the stack when
targeting EABI, aligning them instead to 4 bytes. CHERIoT does not specifically care whether
or not we are using the GCC-compatible version, but if the frontend and backend disagree
on which version we are using, then the caller and callee will end up with incompatible stack
layouts. This is because the caller layout is constructed by the backend, while the callee
layout is constructed by the frontend.

Author: resistor

clang/lib/CodeGen/CodeGenModule.cpp

@@ -244,6 +244,8 @@ createTargetCodeGenInfo(CodeGenModule &CGM) {
     else if (ABIStr.ends_with("d"))
       ABIFLen = 64;
     bool EABI = ABIStr.ends_with("e");
+    EABI |= ABIStr == "cheriot";
+    EABI |= ABIStr == "cheriot-baremetal";
     return createRISCVTargetCodeGenInfo(CGM, XLen, ABIFLen, EABI);
   }
 

clang/test/CodeGen/cheri/cheriot-variadic.c

@@ -14,9 +14,24 @@ typedef struct {
 
 extern int onward(void *, int, char *);
 
+// CHECK-LABEL: @foo
 int foo(va_list ap) {
   // Make sure that we don't see a memcpy in address space zero!
   // CHECK-NOT: p0i8
   bar_t x = va_arg(ap, bar_t);
   return onward(x.a, x.b, x.c);
 }
+
+// CHECK-LABEL: @bar
+double bar(const char* c, ...) {
+// Make sure that a double is *not* dynamically aligned up to 8 bytes.
+// This may change once the RVE spec is ratified and compatibility
+// with old versions of GCC are dropped in upstream.
+// CHECK-NOT: ptrmask
+    va_list ap;
+    va_start((ap), c);
+    int i1 = va_arg(ap, int);
+    double f = va_arg(ap, double);
+    va_end(ap);
+    return f;
+}