[PATCH] Disable specific MSVC warnings so that W4 warnings can be enabled.

Andy Kaylor andrew.kaylor at intel.com
Tue Mar 24 11:29:53 PDT 2015


REPOSITORY
  rL LLVM

================
Comment at: cmake/modules/HandleLLVMOptions.cmake:286
@@ +285,3 @@
+    -wd4706 # Suppress 'assignment within conditional expression'
+    -wd4310 # Suppress 'cast truncates constant value'
+    -wd4701 # Suppress 'potentially uninitialized local variable'
----------------
dblaikie wrote:
> You mentioned this one fired ~200 times - this might actually be worth doing a small sample check to see what it's firing on. Perhaps these are things we should fix - unless they appear in dependent expressions in templates, in which case they're probably false positives.
> 
> The rest of these disables I'm totally fine with - if you want, please commit them separately/ahead of the rest of the review.
These look pretty harmless.  All of the ones I looked at were something like these:

```
uint8_t(~0U)
(char)0xFF
(uint16_t)-1U
```

================
Comment at: lib/Support/APFloat.cpp:1433
@@ -1432,3 +1432,3 @@
      an addition or subtraction.  */
-  subtract ^= sign ^ rhs.sign;
+  subtract ^= (sign ^ rhs.sign) != 0;
 
----------------
dblaikie wrote:
> andrew.w.kaylor wrote:
> > This addresses warning C4805: unsafe mix of type <X> and type <Y> in operation.
> This one's been fixed (slightly differently) by Aaron Ballman. We had some more discussions about the same issues/principles in that review thread too - I think he's OK with the idea of disabling this warning.
OK

================
Comment at: lib/Target/ARM/InstPrinter/ARMInstPrinter.cpp:641
@@ -640,3 +640,3 @@
   unsigned Op = ARM_AM::getAM5Op(MO2.getImm());
-  if (AlwaysPrintImm0 || ImmOffs || Op == ARM_AM::sub) {
+  if (ImmOffs || Op == ARM_AM::sub || AlwaysPrintImm0) {
     O << ", "
----------------
dblaikie wrote:
> andrew.w.kaylor wrote:
> > This addresses warning C4189: local variable is initialized but not referenced.
> > 
> > AlwaysPrintImm0 is a template parameter that can short circuit the condition so that ImmOffs and Op are never referenced.
> This seems like a step backwards, actually - we try to deliberately put the cheap tests first (and a compile time constant is the cheapest test). Granted any compiler worth its salt will kill the dead expressions, but there's no need to ask it to do that work.
> 
> I'm fairly strongly in favor of disabling this warning rather than making this code change.
If it were just cases like this covered by the warning I would definitely agree that it's preferable to disable the warning.  The problem, as Aaron Ballman mentioned, is that this same warning reports the basic case.
```
void foo() {
  int i = 0;
  // Doesn't use i
}
```
That's exactly the sort of problem that I want to see when compiling with MSVC that motivated me to undertake this change set.

I hear what you're saying about the change here moving backward, though the way the code is written the compiler still has to know that the calls above the condition have no side effects to eliminate them.  How about if instead I rewrite the condition like this?

```
  if (AlwaysPrintImm0 ||
      ARM_AM::getAM5Offset(MO2.getImm()) ||
      ARM_AM::getAM5Op(MO2.getImm()) == ARM_AM::sub) {
```

================
Comment at: tools/llvm-c-test/metadata.c:21
@@ -21,1 +20,3 @@
+  LLVMValueRef values[1];
+  values[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 
----------------
dblaikie wrote:
> andrew.w.kaylor wrote:
> > This addresses warning C4204: nonstandard extension used : non-constant aggregate initializer.
> Clang has a warning for this (-Wc99-extensions) but we don't enable it, so I'd say it's reasonable to disable MSVC's equivalent for consistency. Unless someone wants to look at turning Clang's version on & seeing what breaks... 
I have no objection to disabling this warning.

http://reviews.llvm.org/D8572

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/






More information about the llvm-commits mailing list