Stuart Hastings
2009-Jul-06 22:47 UTC
[LLVMdev] switching to llvm-g++ as the host compiler
IANACE (I Am Not A Configury Expert :-), but I've been asked to
reconfigure LLVM and LLVM-GCC to build with LLVM-GCC (perhaps more
accurately, LLVM-G++).
The patches to do this are short. In the LLVM tree, this suffices; it
looks for LLVM-GCC first, and if that fails, falls back to GCC:
Index: llvm.test/autoconf/configure.ac
==================================================================---
llvm.test/autoconf/configure.ac (revision 74729)
+++ llvm.test/autoconf/configure.ac (working copy)
@@ -601,8 +601,8 @@
dnl Check for compilation tools
AC_PROG_CPP
-AC_PROG_CC(gcc)
-AC_PROG_CXX(g++)
+AC_PROG_CC([llvm-gcc gcc])
+AC_PROG_CXX([llvm-g++ g++])
AC_PROG_NM
AC_SUBST(NM)
Likewise, in our LLVM-GCC tree, this seems to work for Darwin/x86:
Index: llvm-gcc.test/configure.in
==================================================================---
llvm-gcc.test/configure.in (revision 74729)
+++ llvm-gcc.test/configure.in (working copy)
@@ -993,10 +993,10 @@
# APPLE LOCAL begin dynamic-no-pic
i[[3456789]]86-*-darwin*)
host_makefile_frag="config/mh-x86-darwin"
- # gcc can default to x86_64 code generation, avoid that
+ # llvm-gcc can default to x86_64 code generation, avoid that
if test "${build}" = "${host}"; then
- CC="${CC-gcc} -m32"
- CXX="${CXX-g++} -m32"
+ CC="${CC-llvm-gcc} -m32"
+ CXX="${CXX-llvm-g++} -m32"
fi
;;
# APPLE LOCAL end dynamic-no-pic
I've done some light testing of these, and I'm convinced they will
work, at least for those of us on Darwin. The latter patch is Darwin/
x86-specific, and hopefully non-contentious. I would hope the former
patch wouldn't break anybody, but I thought I'd ask before applying it.
Of course, I'm omitting the diffs resulting from the re-auto-
confiscation of both trees. I was obliged to tweak all the version
checks in the LLVM file AutoRegen.sh in order to get it to run :-),
but no errors turned up after my hacking. (Should AutoRegen.sh be
updated too?)
Comments invited; if there's a better approach, please share.
stuart
On Mon, Jul 6, 2009 at 3:47 PM, Stuart Hastings<stuart at apple.com> wrote:> IANACE (I Am Not A Configury Expert :-), but I've been asked to > reconfigure LLVM and LLVM-GCC to build with LLVM-GCC (perhaps more > accurately, LLVM-G++). > > The patches to do this are short. In the LLVM tree, this suffices; it > looks for LLVM-GCC first, and if that fails, falls back to GCC:Would it make sense to have configure bootstrap the current sources instead of using whatever llvm-gcc is installed?> Of course, I'm omitting the diffs resulting from the re-auto- > confiscation of both trees. I was obliged to tweak all the version > checks in the LLVM file AutoRegen.sh in order to get it to run :-), > but no errors turned up after my hacking. (Should AutoRegen.sh be > updated too?)I was just asking about this in IRC. I couldn't grok why AutoRegen.sh wouldn't work on a stock Leopard install. Fixing this would be most welcome. deep
On Jul 6, 2009, at 3:47 PM, Stuart Hastings wrote:> The patches to do this are short. In the LLVM tree, this suffices; it > looks for LLVM-GCC first, and if that fails, falls back to GCCSeems reasonable.
On Mon, Jul 6, 2009 at 3:47 PM, Stuart Hastings<stuart at apple.com> wrote:> IANACE (I Am Not A Configury Expert :-), but I've been asked to > reconfigure LLVM and LLVM-GCC to build with LLVM-GCC (perhaps more > accurately, LLVM-G++).The attached patches affect the *tests*, not the actual builds themselves; is that intentional? -Eli
Please make sure when we are running the llvm test suite the "native" executables are still built with gcc (or another compiler) unless it's explicitly specified to be llvm-gcc. Thanks. Evan On Jul 6, 2009, at 3:47 PM, Stuart Hastings wrote:> IANACE (I Am Not A Configury Expert :-), but I've been asked to > reconfigure LLVM and LLVM-GCC to build with LLVM-GCC (perhaps more > accurately, LLVM-G++). > > The patches to do this are short. In the LLVM tree, this suffices; it > looks for LLVM-GCC first, and if that fails, falls back to GCC: > > Index: llvm.test/autoconf/configure.ac > ==================================================================> --- llvm.test/autoconf/configure.ac (revision 74729) > +++ llvm.test/autoconf/configure.ac (working copy) > @@ -601,8 +601,8 @@ > > dnl Check for compilation tools > AC_PROG_CPP > -AC_PROG_CC(gcc) > -AC_PROG_CXX(g++) > +AC_PROG_CC([llvm-gcc gcc]) > +AC_PROG_CXX([llvm-g++ g++]) > > AC_PROG_NM > AC_SUBST(NM) > > > Likewise, in our LLVM-GCC tree, this seems to work for Darwin/x86: > > Index: llvm-gcc.test/configure.in > ==================================================================> --- llvm-gcc.test/configure.in (revision 74729) > +++ llvm-gcc.test/configure.in (working copy) > @@ -993,10 +993,10 @@ > # APPLE LOCAL begin dynamic-no-pic > i[[3456789]]86-*-darwin*) > host_makefile_frag="config/mh-x86-darwin" > - # gcc can default to x86_64 code generation, avoid that > + # llvm-gcc can default to x86_64 code generation, avoid that > if test "${build}" = "${host}"; then > - CC="${CC-gcc} -m32" > - CXX="${CXX-g++} -m32" > + CC="${CC-llvm-gcc} -m32" > + CXX="${CXX-llvm-g++} -m32" > fi > ;; > # APPLE LOCAL end dynamic-no-pic > > > I've done some light testing of these, and I'm convinced they will > work, at least for those of us on Darwin. The latter patch is Darwin/ > x86-specific, and hopefully non-contentious. I would hope the former > patch wouldn't break anybody, but I thought I'd ask before applying > it. > > Of course, I'm omitting the diffs resulting from the re-auto- > confiscation of both trees. I was obliged to tweak all the version > checks in the LLVM file AutoRegen.sh in order to get it to run :-), > but no errors turned up after my hacking. (Should AutoRegen.sh be > updated too?) > > Comments invited; if there's a better approach, please share. > > stuart > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
Shouldn't it be using the compiler specified by --with-llvmgcc/--with- llvmgxx if specified? Normally that just gives a compiler driver for the test suite, but it seems appropriate to use here as well, since most people won't have llvm-gcc-4.2 built with --program-prefix=llvm- unless they do special configuration there as well. Or maybe llvm-gcc-4.2 should be defaulting to --program-prefix=llvm- if --enable-llvm is used? Shantonu Sent from my MacBook On Jul 6, 2009, at 5:34 PM, Evan Cheng wrote:> Please make sure when we are running the llvm test suite the "native" > executables are still built with gcc (or another compiler) unless it's > explicitly specified to be llvm-gcc. Thanks. > > Evan > > On Jul 6, 2009, at 3:47 PM, Stuart Hastings wrote: > >> IANACE (I Am Not A Configury Expert :-), but I've been asked to >> reconfigure LLVM and LLVM-GCC to build with LLVM-GCC (perhaps more >> accurately, LLVM-G++). >> >> The patches to do this are short. In the LLVM tree, this suffices; >> it >> looks for LLVM-GCC first, and if that fails, falls back to GCC: >> >> Index: llvm.test/autoconf/configure.ac >> ==================================================================>> --- llvm.test/autoconf/configure.ac (revision 74729) >> +++ llvm.test/autoconf/configure.ac (working copy) >> @@ -601,8 +601,8 @@ >> >> dnl Check for compilation tools >> AC_PROG_CPP >> -AC_PROG_CC(gcc) >> -AC_PROG_CXX(g++) >> +AC_PROG_CC([llvm-gcc gcc]) >> +AC_PROG_CXX([llvm-g++ g++]) >> >> AC_PROG_NM >> AC_SUBST(NM) >> >> >> Likewise, in our LLVM-GCC tree, this seems to work for Darwin/x86: >> >> Index: llvm-gcc.test/configure.in >> ==================================================================>> --- llvm-gcc.test/configure.in (revision 74729) >> +++ llvm-gcc.test/configure.in (working copy) >> @@ -993,10 +993,10 @@ >> # APPLE LOCAL begin dynamic-no-pic >> i[[3456789]]86-*-darwin*) >> host_makefile_frag="config/mh-x86-darwin" >> - # gcc can default to x86_64 code generation, avoid that >> + # llvm-gcc can default to x86_64 code generation, avoid that >> if test "${build}" = "${host}"; then >> - CC="${CC-gcc} -m32" >> - CXX="${CXX-g++} -m32" >> + CC="${CC-llvm-gcc} -m32" >> + CXX="${CXX-llvm-g++} -m32" >> fi >> ;; >> # APPLE LOCAL end dynamic-no-pic >> >> >> I've done some light testing of these, and I'm convinced they will >> work, at least for those of us on Darwin. The latter patch is >> Darwin/ >> x86-specific, and hopefully non-contentious. I would hope the former >> patch wouldn't break anybody, but I thought I'd ask before applying >> it. >> >> Of course, I'm omitting the diffs resulting from the re-auto- >> confiscation of both trees. I was obliged to tweak all the version >> checks in the LLVM file AutoRegen.sh in order to get it to run :-), >> but no errors turned up after my hacking. (Should AutoRegen.sh be >> updated too?) >> >> Comments invited; if there's a better approach, please share. >> >> stuart >> _______________________________________________ >> LLVM Developers mailing list >> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
On Mon, Jul 6, 2009 at 3:47 PM, Stuart Hastings<stuart at apple.com> wrote: ==================================================================> --- llvm-gcc.test/configure.in (revision 74729)> +++ llvm-gcc.test/configure.in (working copy) > @@ -993,10 +993,10 @@ > # APPLE LOCAL begin dynamic-no-pic > i[[3456789]]86-*-darwin*) > host_makefile_frag="config/mh-x86-darwin" > - # gcc can default to x86_64 code generation, avoid that > + # llvm-gcc can default to x86_64 code generation, avoid that > if test "${build}" = "${host}"; then > - CC="${CC-gcc} -m32" > - CXX="${CXX-g++} -m32" > + CC="${CC-llvm-gcc} -m32" > + CXX="${CXX-llvm-g++} -m32" > fi > ;; > # APPLE LOCAL end dynamic-no-pic >Stuart, Here we are making two independent LOCAL changes. 1. use llvm-gcc & llvm-g++, if available 2. Force -m32 Pl. mention both changes explicitly in the comment. - Devang
On Tue, Jul 7, 2009 at 11:50 AM, Devang Patel<devang.patel at gmail.com> wrote:> On Mon, Jul 6, 2009 at 3:47 PM, Stuart Hastings<stuart at apple.com> wrote: > > ==================================================================>> --- llvm-gcc.test/configure.in (revision 74729) >> +++ llvm-gcc.test/configure.in (working copy) >> @@ -993,10 +993,10 @@ >> # APPLE LOCAL begin dynamic-no-pic >> i[[3456789]]86-*-darwin*) >> host_makefile_frag="config/mh-x86-darwin" >> - # gcc can default to x86_64 code generation, avoid that >> + # llvm-gcc can default to x86_64 code generation, avoid that >> if test "${build}" = "${host}"; then >> - CC="${CC-gcc} -m32" >> - CXX="${CXX-g++} -m32" >> + CC="${CC-llvm-gcc} -m32" >> + CXX="${CXX-llvm-g++} -m32" >> fi >> ;; >> # APPLE LOCAL end dynamic-no-pic >> > > Stuart, > > Here we are making two independent LOCAL changes. > 1. use llvm-gcc & llvm-g++, if available > 2. Force -m32 > > Pl. mention both changes explicitly in the comment. >Also, we should fall back to a default of "gcc" and "g++" if LLVM isn't available. -bw
Stuart Hastings
2009-Jul-09 20:45 UTC
[LLVMdev] switching to llvm-g++ as the host compiler
On Jul 6, 2009, at 5:24 PM, Eli Friedman wrote:> On Mon, Jul 6, 2009 at 3:47 PM, Stuart Hastings<stuart at apple.com> > wrote: >> IANACE (I Am Not A Configury Expert :-), but I've been asked to >> reconfigure LLVM and LLVM-GCC to build with LLVM-GCC (perhaps more >> accurately, LLVM-G++). > > The attached patches affect the *tests*, not the actual builds > themselves; is that intentional?I'm missing something significant here. I tried my patch before posting it to the list, and I'm pretty sure that LLVM got compiled with a previously installed LLVM-GCC. And the tests I ran seemed "normal" too (i.e. the right thing got tested). Am I misunderstanding you? Could you be more specific? What, exactly, do you foresee going wrong? Thank you in advance, stuart