Linaro GCC

ice in slp pass for WIDEN_LSHIFT_EXPR

Bug #960274 reported by mickael guene on 2012-03-20

This bug report is a duplicate of: Bug #960283: slp pass assert when compiler configure with --enable-checking. Edit Remove

This bug affects 1 person

Affects		Status	Importance	Assigned to	Milestone
	Linaro GCC	In Progress	High	Ulrich Weigand

Bug Description

Compiler crash when compiling small.c file with following options:
I'm using 4.6-2012.03 release.
gcc -c small.c -msoft-float -march=armv7-a -mfloat-abi=softfp -mfpu=neon -mvectorize-with-neon-quad -ftree-vectorize -O2

It appear that problem is due to the fact that we deference random pointer. This is due to the fact that in vectorizable_type_promotion() in case of slp and with code == WIDEN_LSHIFT_EXPR no operand is push into vec_oprnds1.

Attach patch fix this problem but then we have crash later on.

See original description

Revision history for this message

mickael guene (mickael-guene) wrote on 2012-03-20:

source file that cause crash Edit (481 bytes, text/plain)

Revision history for this message

mickael guene (mickael-guene) wrote on 2012-03-20:

small.patch Edit (613 bytes, text/plain)

description:

updated

Revision history for this message

Ulrich Weigand (uweigand) wrote on 2012-03-20:

Thanks for the bug report. I can confirm this also with current mainline. I'll have a look.

Changed in gcc-linaro:
status:	New → Confirmed
importance:	Undecided → High
assignee:	nobody → Ulrich Weigand (uweigand)

Ulrich Weigand (uweigand) on 2012-03-21

Changed in gcc-linaro:
status:	Confirmed → In Progress

Revision history for this message

Ulrich Weigand (uweigand) wrote on 2012-03-23:

It turns out this patch isn't actually needed once the fix for bug #960283 is in; the number of operands, once computed correctly, is only half of slp_node->vec_stmts_size, so the off-by-one doesn't actually matter.

Marking this bug as a duplicate.