https://llvm.org/bugs/show_bug.cgi?id=24394
Bug ID: 24394
Summary: Bad optimized LLVM IR generation after PowerPC
optimization using llopt -O2
Product: new-bugs
Version: 3.7
Hardware: PC
OS: Windows NT
Status: NEW
Severity: normal
Priority: P
Component: new bugs
Assignee: unassignedb...@nondot.org
Reporter: matt.da...@pgroup.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Created attachment 14709
--> https://llvm.org/bugs/attachment.cgi?id=14709&action=edit
Source, non-llopt, and llopt llvm IR
I have a simple TLS test case in C. This test runs to completion when
assembled in llvm 3.6 with llopt at -O2, or in 3.7 NOT using opt -O2. However,
when assembled when building the code using llc 3.7 with llopt -O2 a reference
to a local variable creates a segmentation fault.
I have the offending code in LLVM IR attached (code generated before and after
opt -O2). I also have the original C source that was used to generate the llvm
ir. I have stripped the llvm ir samples, making them more readable.
What you will see is that opt removed the assignment of the integer value, 10,
to the TLS variable "x"
Additionally, the TLS variable "x" has been converted from i32 to i1:
Before llopt:
@x = internal thread_local global i32 100, align 4
After llopt:
@x = internal thread_local unnamed_addr global i1 false
Works:
llvm/3.7/bin/llc ./test.ll -mcpu=native -O2 -post-RA-scheduler=0
-relocation-model=pic -o ./test.s
/usr/bin/as ./test.s -mpower8 -many -o test.o
Fails:
llvm/3.7/bin/opt -O2 ./test.ll -S -o ./test_opt.ll
llvm/3.7/bin/llc ./test_opt.ll -mcpu=native -O2 -post-RA-scheduler=0
-relocation-model=pic -o ./test_opt.s
/usr/bin/as ./test_opt.s -mpower8 -many -o test_opt.o
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
