Incorrect generation of movapd for unaligned load

[hfinkel]

Bugzilla Link	11398
Resolution	INVALID
Resolved on	Nov 18, 2011 16:09
Version	trunk
OS	Linux
Attachments	Example
CC	@efriedma-quic

Extended Description

For x86_64, an instruction like:

%0 = load <2 x double>* %whatever, align 8, !tbaa !​0

can get assembled into a movapd instruction (using trunk r144649). This, however, causes a segmentation fault because movapd requires aligned data. I've attached an example file that suffers from this problem (see the very first instruction in the main function).

[efriedma-quic]

%1 = load <2 x double>* bitcast (double* getelementptr inbounds ([5 x %struct.planet]* @​bodies, i64 0, i64 0, i32 5) to <2 x double>*), align 16, !tbaa !​0

It looks like your testcase isn't correct.

[hfinkel]

The test case is correct. The main function starts with:
define i32 @​main(i32 %argc, i8** nocapture %argv) nounwind uwtable {
entry:
%0 = load <2 x double>* bitcast (double* getelementptr inbounds ([5 x %struct.planet]* @​bodies, i64 0, i64 0, i32 3) to <2 x double>), align 8, !tbaa !​0
%1 = load <2 x double> bitcast (double* getelementptr inbounds ([5 x %struct.planet]* @​bodies, i64 0, i64 0, i32 5) to <2 x double>*), align 16, !tbaa !​0

Indeed, the second load has the correct alignment, but the first load is unaligned. Furthermore, running the attached file under lli confirms the problem, and gdb gives the location of the crash:

$ gdb lli
...
(gdb) r -O3 ~/tmp/nbo.ll
...
Program received signal SIGSEGV, Segmentation fault.
0x00007ffff7f46062 in main ()
(gdb) disassemble
Dump of assembler code for function main:
0x00007ffff7f46050 <+0>: push %rbp
0x00007ffff7f46051 <+1>: push %r14
0x00007ffff7f46053 <+3>: push %rbx
0x00007ffff7f46054 <+4>: sub $0x10,%rsp
0x00007ffff7f46058 <+8>: movabs $0x7ffff7fe9010,%rax
=> 0x00007ffff7f46062 <+18>: movapd 0x28(%rax),%xmm6
0x00007ffff7f46067 <+23>: movapd 0x60(%rax),%xmm0
0x00007ffff7f4606c <+28>: movapd 0x98(%rax),%xmm5
0x00007ffff7f46074 <+36>: movapd 0xd0(%rax),%xmm10

[efriedma-quic]

"movapd 0x28(%rax),%xmm6" corresponds to "%1 = load <2 x double>* bitcast (double* getelementptr inbounds ([5 x %struct.planet]* @​bodies, i64 0, i64 0, i32 5) to <2 x double>*), align 16", I think... 5 * 8 == 0x28. Your code is getting rescheduled.

[hfinkel]

You are correct, the alignment tag is incorrect.