On Jun 16, 2007, at 12:35 AM, Duncan Sands wrote:>> Result compiled with llvm-g++ 2.0: >> tan float: -2.18504 >> tan double: 0.309336 > > This may be due to bug 1505.It fails on x86 using x87 floating point, with the inliner not run, because of 1505, yes. Gonsolo, is that your situation? (What happens is, there is a wrapper in the header file for std::tan (float), like this: inline float tan(float __x) { return __builtin_tanf(__x); } This wrapper is miscompiled due to 1505; when the inliner is run, no problem.) (But, when opt -std-compile-opts is run, the 'tan' call is evaluated at compile time, while 'tanf' resolves to a libc call. This is undesirable, since the whole reason tanf exists is that it's supposed to be faster than tan. Of course you have to be very careful about doing FP operations at compile time, but I assume this is being done right; it's getting the right answer here, anyway.)
Hi! Dale Johannesen schrieb:> > On Jun 16, 2007, at 12:35 AM, Duncan Sands wrote: > >>> Result compiled with llvm-g++ 2.0: >>> tan float: -2.18504 >>> tan double: 0.309336 >> >> This may be due to bug 1505. > > It fails on x86 using x87 floating point, with the inliner not run, > because of 1505, yes. Gonsolo, is that your situation? >I tried a simple "llvm-g++ -o simple tan.bug.cc" Output: tan float: -2.18504 tan double: 0.309336> (What happens is, there is a wrapper in the header file for std::tan(float), > like this: > > inline float > tan(float __x) > { return __builtin_tanf(__x); } > > This wrapper is miscompiled due to 1505; when the inliner is run, no problem.) > > (But, when opt -std-compile-opts is run, the 'tan' call is evaluated at compile > time, while 'tanf' resolves to a libc call. This is undesirable, since the whole > reason tanf exists is that it's supposed to be faster than tan. Of course you > have to be very careful about doing FP operations at compile time, but I > assume this is being done right; it's getting the right answer here, anyway.)I also tried: 1. g++ Output: tan float: 0.309336 tan double: 0.309336 Works. 2. lli Output: tan float: 0.309336 tan double: 0.309336 Works 3. lli with opt Output: tan float: 0.309336 tan double: 0.309336 Works. 4. opt + llc + as + g++ Output: tan float: 0.309336 tan double: 0.309336 Works 5. opt + llc + as + llvm-g++ Output: tan float: 0 tan double: 0 ???? In summary: Everything works except linking with llvm-g++! Gonsolo Appendix A: <tan_bug.cc> #include <iostream> #include <cmath> int main() { float a = 0.3; double b = 0.3; float result_a = std::tan( a ); float result_b = std::tan( b ); std::cout << "tan float: " << result_a << std::endl; std::cout << "tan double: " << result_b << std::endl; } </tan_bug.cc> Appendix B: <Makefile> all: test_g++ test_llvm-g++ test_interpreter test_interpreter_opt test_native_g++ test_native_llvm-g++ .PHONY: c clean test_simple test_interpreter test_interpreter_opt test_native_g++ test_native_llvm-g++ test_g++: simple_g++ ./simple_g++ test_llvm-g++: simple_llvm ./simple_llvm test_interpreter: emit.bc lli emit.bc test_interpreter_opt: opt.bc lli opt.bc test_native_g++: native_g++ ./native_g++ test_native_llvm-g++: native_llvm-g++ ./native_llvm-g++ simple_g++: tan_bug.cc g++ -o simple_g++ tan_bug.cc simple_llvm: tan_bug.cc llvm-g++ -o simple_llvm tan_bug.cc emit.bc: tan_bug.cc llvm-g++ -c --emit-llvm -o emit.bc tan_bug.cc opt.bc: emit.bc opt -std-compile-opts emit.bc -o opt.bc opt.s: opt.bc llc opt.bc -o opt.s opt.o: opt.s as opt.s -o opt.o native_g++: opt.o g++ -o native_g++ opt.o native_llvm-g++: opt.o llvm-g++ -o native_llvm-g++ opt.o c: clean clean: rm -f simple_g++ simple_llvm emit.bc opt.bc opt.s opt.o native_g++ native_llvm-g++ </Makefile> Appendix C: Thanks for your quick reply!!
On Jun 16, 2007, at 12:52 PM, Gonsolo wrote:> Hi! > > Dale Johannesen schrieb: >> >> On Jun 16, 2007, at 12:35 AM, Duncan Sands wrote: >> >>>> Result compiled with llvm-g++ 2.0: >>>> tan float: -2.18504 >>>> tan double: 0.309336 >>> >>> This may be due to bug 1505. >> >> It fails on x86 using x87 floating point, with the inliner not run, >> because of 1505, yes. Gonsolo, is that your situation? >> > > I tried a simple "llvm-g++ -o simple tan.bug.cc" > Output: > tan float: -2.18504 > tan double: 0.309336Thanks, but what I'm trying to get at is, what kind of machine are you running on? If x86, are you set up to use x87 or SSE for floating point? (You can tell by compiling with -save-temps and looking at the .s file. If you see any references to %xmm registers, that's SSE.)