[AVX] Cannot select: v8f32 = X86ISD::VZEXT_MOVL

[llvmbot]

Bugzilla Link	10844
Resolution	FIXED
Resolved on	Sep 08, 2011 20:49
Version	unspecified
OS	All
Attachments	test case
Reporter	LLVM Bugzilla Contributor
CC	@bcardosolopes

Extended Description

With the attached test case, I'm seeing the following crash:

% llc -O0 bug.bc -filetype=obj -o x.o -mattr=+avx
LLVM ERROR: Cannot select: 0x7fd50907a310: v8f32 = X86ISD::VZEXT_MOVL 0x7fd50907ab10 [ID=96]
0x7fd50907ab10: v8f32 = insert_subvector 0x7fd50907aa10, 0x7fd50907a910, 0x7fd50902f410 [ID=91]
0x7fd50907aa10: v8f32 = undef [ID=35]
0x7fd50907a910: v4f32 = scalar_to_vector 0x7fd509029e10 [ID=85]
0x7fd509029e10: f32,ch = load 0x10dfe7bd8, 0x7fd509027d10, 0x7fd50902ed10<LD4[ConstantPool]> [ID=75]
0x7fd509027d10: i64 = X86ISD::WrapperRIP 0x7fd509028010 [ID=50]
0x7fd509028010: i64 = TargetConstantPool<float 0xFFFFFFFFE0000000> 0 [ID=27]
0x7fd50902ed10: i64 = undef [ORD=181] [ID=3]
0x7fd50902f410: i32 = Constant<0> [ORD=185] [ID=8]

(If I try to reduce it to a simpler test case with bugpoint, then it ends up reducing it to something that hits 10842 or 10843.)

[bcardosolopes]

Fixed in TOT. Not sure though when it was solved! :)

[bcardosolopes]

Actually I forgot to use -O0 to test it, reopening...

[bcardosolopes]

Hi, fixed in r139304. Is there any reason why are you using llc with -O0? This opt level doesn't even fold loads, and unless you don't have any specific reason I'd recommend you using the default level (-O2)

[llvmbot]

Excellent!

(Usually, I end up running the equivalent of -O2, but when generating debugging info, for example, will run essentially no optimization passes. If there are a set of optimizations I can still apply when doing debugging that won't rearrange instructions with respect to their source positions, I'd be happy to run those, but otherwise want to be conservative about that...)