Subzero. ARM32. Nonsfi.

crosstest/test_cast_to_u1.ll

Index: crosstest/test_cast_to_u1.ll
diff --git a/crosstest/test_cast_to_u1.ll b/crosstest/test_cast_to_u1.ll
index 31c1f077e3458ad2cbb84d470ad6cd85aec086d0..3e4f07bef52ef42b95c5b803a9b5423ac418f35c 100644
--- a/crosstest/test_cast_to_u1.ll
+++ b/crosstest/test_cast_to_u1.ll
@@ -76,15 +76,17 @@ entry:
 define i32 @_Z4castIdbET0_T_(double %a) {
 entry:
 ;  %tobool = fcmp une double %a, 0.000000e+00
-  %tobool = fptoui double %a to i1
-  %tobool.ret_ext = zext i1 %tobool to i32
+  %tobool = fptoui double %a to i32
+  %tobool.i1 = trunc i32 %tobool to i1

[Jim Stichnoth, 2016/02/10 06:35:59]

I'd like to understand more about this change.

I expect that one problem is that we are testing undefined behavior, i.e. in
most cases a double doesn't truncate "cleanly" to i1 and results in undefined
behavior.

But truncating (fptoui) a double to i32 also may have undefined behavior, so
this change isn't making the test bulletproof or anything.

In fact, when possible, Subzero's goal is to mimic llc's undefined behavior.  In
theory, this minimizes wasting time investigating bug reports that turn out to
be the program relying on undefined behavior.

All that said, what on earth does this have to do with nonsfi, and why didn't we
see it before?

[John, 2016/02/10 15:41:13]

the nonsfi test_cast_main.cpp is ocmpiled with -O2, while {nacl,native} are both
compiled with -O0.

For both O0 and O2, pnacl-clang does the following

ret.llc = cast<float, bool>()
ret.llc.i1 = reg & 1
ret.sz = Subzero_::cast<float, bool>()
ret.sz.i1 = reg & 1

now, here's the  difference. For O0, pnacl-clang does

memcmp(&ret.llc.i1, &ret.sz.i1)

and, for O2,

cmp ret.llc, ret.sz


For X86, we see the same lowering for calling the cast functions, and the same
memcmp call for O0. For O2, it does

cmp.i32 = xor ret.llc, reg.sz
cmp.i1 = cmp.i32 & 1

In essence, arm32 nonsfi just exposed something that's "wrong" with subzero in
general. Even for x86, llvm does not & 1 when converting from from f?? to i1,
but subzero does.

What should we do here? I can see us leaving this change here, or we could
change the crosstest.py script, or we could even update the targets to not "& 1"
these fptoui conversions (this last approach would probably be better off if
implemented in a different CL, and we would still need a stop gap solution.)

+  %tobool.ret_ext = zext i1 %tobool.i1 to i32
   ret i32 %tobool.ret_ext
 }
 
 define i32 @_Z4castIfbET0_T_(float %a) {
 entry:
 ;  %tobool = fcmp une float %a, 0.000000e+00
-  %tobool = fptoui float %a to i1
-  %tobool.ret_ext = zext i1 %tobool to i32
+  %tobool = fptoui float %a to i32
+  %tobool.i1 = trunc i32 %tobool to i1
+  %tobool.ret_ext = zext i1 %tobool.i1 to i32
   ret i32 %tobool.ret_ext
 }