[llvm-bugs] [Bug 25342] New: [ppc] floating point var allocated to integer register

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=25342

            Bug ID: 25342
           Summary: [ppc] floating point var allocated to integer register
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: PowerPC
          Assignee: unassignedbugs at nondot.org
          Reporter: carrot at google.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Created attachment 15183
  --> https://llvm.org/bugs/attachment.cgi?id=15183&action=edit
testcase

Compile the attached source code with options

~/llvm/obj/bin/clang++ --target=powerpc64le-grtev4-linux-gnu -O2
-fno-strict-aliasing -m64 -mvsx -mcpu=power8 -ffp-contract=fast

llvm generates following code for the inner loop

        lfsu 1, 16(22)
        addi 24, 24, -2
        cmplwi 6, 24, 0
        lfsu 2, 16(23)
        lxsspx 4, 22, 5
        lxsspx 3, 23, 5
        stw 25, -112(1)           // A
        ori 2, 2, 0 
        lxsspx 5, 0, 6            // B
        stw 26, -108(1)           // C
        xsmulsp 7, 3, 4 
        ori 2, 2, 0 
        lxsspx 6, 0, 7            // D
        lxsspx 9, 23, 10
        lxsspx 8, 22, 10
        lxsspx 11, 23, 8
        lxsspx 12, 22, 8
        xsmaddasp 6, 2, 4 
        xsmulsp 10, 9, 8 
        xsmsubasp 7, 2, 1 
        xsmaddasp 6, 3, 1 
        xsmsubasp 10, 11, 12
        xsaddsp 5, 5, 7 
        xsmaddasp 6, 11, 8
        xsaddsp 13, 5, 10
        xsmaddasp 6, 9, 12
        stxsspx 13, 0, 11         // E
        ori 2, 2, 0 
        lwz 25, -100(1)           // F
        stxsspx 6, 0, 12          // G
        ori 2, 2, 0 
        lwz 26, -104(1)           // H
        bne      6, .LBB0_9

The variable std::complex<float> sum is allocated to integer registers r25,r26.
Before using it, instructions ABCD move it to floating point registers, after
generating new values, instructions EFGH move it back to integer registers.

GCC allocates sum to floating point registers directly.

On power8, llvm generated code is 4.5x slower than gcc.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20151028/6972a0b6/attachment.html>