Context :

Aims : We present 139 spectra of 124 Type Ia supernovae ( SNe Ia ) that were observed at the ESO/VLT during the first three years of the Canada-France-Hawaï Telescope ( CFHT ) Supernova Legacy Survey ( SNLS ) .
This homogeneous data set is used to test for redshift evolution of SN Ia spectra , and will be used in the SNLS 3rd year cosmological analyses .

Methods : Spectra have been reduced and extracted with a dedicated pipeline that uses photometric information from deep CFHT Legacy Survey ( CFHT-LS ) reference images to trace , at sub-pixel accuracy , the position of the supernova on the spectrogram as a function of wavelength .
It also separates the supernova and its host light in \sim 60 % of cases .
The identification of the supernova candidates is performed using a spectrophotometric SN Ia model .

Results : A total of 124 SNe Ia , roughly 50 % of the overall SNLS spectroscopic sample , have been identified using the ESO/VLT during the first three years of the survey .
Their redshifts range from z = 0.149 to z = 1.031 .
The average redshift of the sample is z = 0.63 \pm 0.02 .
This constitutes the largest SN Ia spectral set to date in this redshift range .
The spectra are presented along with their best-fit spectral SN Ia model and a host model where relevant .
In the latter case , a host subtracted spectrum is also presented .
We produce average spectra for pre-maximum , maximum and post-maximum epochs for both z < 0.5 and z \geq 0.5 SNe Ia .
We find that z < 0.5 spectra have deeper intermediate mass element absorptions than z \geq 0.5 spectra .
The differences with redshift are consistent with the selection of brighter and bluer supernovae at higher redshift .

Conclusions :