gcc generates invalid vmov.f64 asm call with -march=armv7-a -mfpu=neon -marm -O3

[Bernhard Rosenkraenzer]

Building ffmpeg with a current linaro-gcc 4.6 snapshot results in

[bero@matterhorn ffmpeg]$ /opt/android-toolchain-eabi/bin/arm-linux-androideabi-gcc -march=armv7-a -mfloat-abi=softfp -mfpu=neon -fPIC -marm -O3 -o flvenc.o -c flvenc.i
/tmp/ccWYA5Nm.s: Assembler messages:
/tmp/ccWYA5Nm.s:989: Error: garbage following instruction -- `vmov.f64 s16,r3,r3'
/tmp/ccWYA5Nm.s:1052: Error: garbage following instruction -- `vmov.f64 s16,r3,r3'

Looks like the correct asm would be "vmov.f64 s16,r3" (without the second ,r3 - there doesn't seem to be a variant of vmov.f64 that takes 3 parameters).

[Bernhard Rosenkraenzer (berolinux)]

Attaching reduced test case

[bero@matterhorn bero]$ arm-linux-androideabi-gcc -msoft-float -march=armv7-a -mfloat-abi=softfp -mfpu=neon -marm -O3 -c -o test.o testcase900426.c
/tmp/ccSZDij0.s: Assembler messages:
/tmp/ccSZDij0.s:25: Error: garbage following instruction -- `vmov.f64 s16,r3,r3'

[Ramana Radhakrishnan (ramana)]

This is my last commit for vcvt fixed to float . There is an obvious fix for it .

Ramana

[Ramana Radhakrishnan (ramana)]

I have reverted the offending commit for the 2011.12 release. We will remerge the feature correctly in a later release.

Ramana

[Michael Hope (michaelh1)]

The updated version of the vcvt patch landed last month.