llvm: fix target triple #30554
llvm: fix target triple #30554
Conversation
broken by their move to cmake causing a switch away from the standard --host/--build autoconf fix #28046
|
If this is a sufficiently critical fix, it could go into 1.1.0 instead of 1.1.1. |
|
No, it’s not a new regression, although it’s certainly good to have it fixed soon |
StefanKarpinski
added
status:triage
JeffBezanson
removed
status:triage
broken by their move to cmake causing a switch away from the standard --host/--build autoconf fix #28046
broken by their move to cmake causing a switch away from the standard --host/--build autoconf fix #28046
broken by their move to cmake causing a switch away from the standard --host/--build autoconf fix #28046
broken by their move to cmake causing a switch away from the standard --host/--build autoconf fix #28046
broken by their move to cmake causing a switch away from the standard --host/--build autoconf fix #28046
| @@ -66,7 +66,7 @@ LLVM_CXXFLAGS += $(CXXFLAGS) | |||
| LLVM_CPPFLAGS += $(CPPFLAGS) | |||
| LLVM_LDFLAGS += $(LDFLAGS) | |||
| LLVM_CMAKE += -DLLVM_TARGETS_TO_BUILD:STRING="$(LLVM_TARGETS)" -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD="$(LLVM_EXPERIMENTAL_TARGETS)" -DCMAKE_BUILD_TYPE="$(LLVM_CMAKE_BUILDTYPE)" | |||
| LLVM_CMAKE += -DLLVM_ENABLE_ZLIB=OFF -DLLVM_ENABLE_LIBXML2=OFF | |||
| LLVM_CMAKE += -DLLVM_ENABLE_ZLIB=OFF -DLLVM_ENABLE_LIBXML2=OFF -DLLVM_HOST_TRIPLE="$(or $(XC_HOST),$(BUILD_MACHINE))" | |||
There was a problem hiding this comment.
This creates problems on Fedora armv7hl. Indeed, there BUILD_MACHINE=armv7hl-redhat-linux-gnueabi by default since that's what gcc -dumpmachine returns. Before this PR, LLVM used armv7l-unknown-linux-gnueabihf by default instead (as computed by deps/srccache/llvm-6.0.1/cmake/config.guess).
I don't really understand why Fedora arm uses gnueabi instead of gnueabihf, since it uses hard float (which is specified e.g. in CFLAGS via -mfloat-abi=hard), but it seems to be a conscious choice. I couldn't find official GCC/LLVM docs about what these mean. What I can see at least is that LLVM's config.guess script branches on __ARM_PCS_VFP | $CC_FOR_BUILD -E - 2>/dev/null | grep -q __ARM_PCS_VFP to decide whether to return gnueabi or gnueabihf. I guess we should do the same to make this more reliable? EDIT: or just avoid passing -DLLVM_HOST_TRIPLE when XC_HOST is unset.
broken by their move to cmake causing a switch away from the standard --host/--build autoconf fix #28046
fix #28046