http://llvm.org/bugs/show_bug.cgi?id=18846

            Bug ID: 18846
           Summary: needless avx spill/reload
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Register Allocator
          Assignee: [email protected]
          Reporter: [email protected]
                CC: [email protected]
    Classification: Unclassified

Created attachment 12063
  --> http://llvm.org/bugs/attachment.cgi?id=12063&action=edit
The LLVM-IR file to reproduce this bug

When running 'llc spill-reload.ll' the generated AVX assembly contains
the following code eight times (once for each spill):

        vmovups 704(%rsp), %ymm8        # 32-byte Folded Reload
        vmovups %ymm8, 704(%rsp)        # 32-byte Folded Spill

This is similar to PR10070, which was already fixed. Spilling the value
beforehand
and reloading it at this very position is a good decission. Even spilling the
same
value again is the right choice. However, I do not see why we need to have an
explicit copy back to the stack. We just loaded the value from the stack, so we
know that the very same value is already on the stack. I would expect LLVM to
not emmit the 'vmovups %ymm8, 704(%rsp)        # 32-byte Folded Spill'
instruction.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
LLVMbugs mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/llvmbugs

Reply via email to