labrinea added inline comments.
================
Comment at: lib/Basic/Targets.cpp:4222
@@ +4221,3 @@
+ setArchInfo(DefaultCPU);
+ ShouldUseInlineAtomic = (ArchISA == llvm::ARM::IK_ARM &&
+ ArchVersion >= 6) ||
----------------
rengolin wrote:
> ShouldUseInlineAtomic must be set up inside setAtomic(), too.
Moving it there triggers regression
(tools/clang/test/CodeGen/atomics-inlining.c). When triple does not define a
subArch, (and therefore llvm::ARMTargetParser::getDefaultCPU(ArchName) returns
null), the value of ShouldUseInlineAtomic should be **false** according to the
logic before my patch. That's because ShouldUseInlineAtomic was parsing the
arch name and returned true only if archName started with "arm**v**" or
"thumb**v**".
================
Comment at: lib/Basic/Targets.cpp:4228
@@ +4227,3 @@
+ else {
+ setArchInfo(CPU);
+ ShouldUseInlineAtomic = false;
----------------
If we are finally keeping the hard-coded string for CPU with a FIXME label,
there is no point in having setArchInfo(ArchName). It is redundant. We still
have to call setArchInfo(CPU) to avoid the regressions.
http://reviews.llvm.org/D10839
_______________________________________________
cfe-commits mailing list
cfe-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
