OpenEuphoria: Ticket #903: sin() and ...

Euphoria Ticket #903: sin() and cos() returns incorrect results when translated

Reported by cargoan Sep 29, 2014
10 comments

--file sin.ex 
 
integer i = 1 
atom a = 1.0 
 
printf(1, "\n\nsin(%d) = %.10f\n", {i, sin(i)}) 
printf(1, "sin(%f) = %.10f\n", {a, sin(a)}) 
printf(1, "\ncos(%d) = %.10f\n", {i, cos(i)}) 
printf(1, "cos(%f) = %.10f\n\n", {a, cos(a)})

In terminal:

eui sine.ex && euc sine.ex && ./sine && ./sine 
 
sin(1) = 0.8414709848 
sin(1.000000) = 0.8414709848 
 
cos(1) = 0.5403023059 
cos(1.000000) = 0.5403023059 
 
Build directory: build-126339/ 
Translating code, pass: 1 2 3  generating 
Compiling with GCC 
Compiling  12% init-.c 
Compiling  50% sine.c 
Compiling  75% main-.c 
Linking 100% ../sine 
 
sin(1) = 1294776.0000000000 
sin(1.000000) = 0.8414709848 
 
cos(1) = 1294780.0000000000 
cos(1.000000) = 0.5403023059 
 
 
sin(1) = 1029048.0000000000 
sin(1.000000) = 0.8414709848 
 
cos(1) = 1029052.0000000000 
cos(1.000000) = 0.5403023059

if parameter is integer the result is wrong and it varies each call or execution. Sorry my english.

My system is Archlinux 64bits

Details

Type:	Bug Report	Severity:	Normal	Category:	Translator
Assigned To:	mattlewis	Status:	Accepted	Reported Release:	4.1.0
Fixed in SVN #:		View VCS:	none	Milestone:

1. Comment by SDPringle Sep 29, 2014

I cannot reproduce this on 32-bit Linux. Looking at the source code generated we have e_sin and unary_op(SIN, ..). I suggest: Put a break point in e_sin and see if a is correct, then look at the value returned and see if that is correct.

2. Comment by mattlewis Sep 30, 2014

I get different (but still bad) results on 64-bit Linux when translating a 64-bit binary. Results are correct for 32-bits.

3. Comment by mattlewis Sep 30, 2014

We need to add some stuff to euphoria.h for the translator. Without definitions, the compiler assumes they return 32-bit values which is wrong on a 64-bit architecture.

4. Comment by cargoan Sep 30, 2014

tan() is bad, sqrt(), arctan() and log() are ok.

5. Comment by cargoan Sep 30, 2014

solved adding to euphoria.h: ... ... object e_log(object a); object De_log(d_ptr a); /* added */ object e_sin(object a); object De_sin(d_ptr a); object e_cos(object a); object De_cos(d_ptr a); object e_tan(object a); object De_tan(d_ptr a); ....

6. Comment by mattlewis Oct 01, 2014

Added several imports:

See: hg:euphoria/rev/e9a250c9e114

7. Comment by DerekParnell Oct 01, 2014

Would it be worthwhile to check all the built-in function signatures to ensure that they are 64-bit ready, and not just the transcendentals.

8. Comment by mattlewis Oct 01, 2014

Probably worth-while. I just went through all of the stuff in be_runtime.c that are in the optable for binary ops to see what was used by the translator and missing from euphoria.h.

9. Comment by DerekParnell Oct 01, 2014

Further test cases should probably include things like ...

test_equal("sin(atom) = sin(int)", sin(1.000000), sin(1))

10. Comment by SDPringle Jan 31, 2015

See: hg:euphoria/rev/bb17603a0b00

changeset: 6298:bb17603a0b00 branch: 4.0 user: Shawn Pringle <shawn.pringle@gmail.com> date: Mon Jan 26 20:47:45 2015 -0300 files: tests/t_math.e description:

OpenEuphoria