[llvm-bugs] [Bug 32564] New: AVX2 - aggressive broadcast generation instead of memory operands

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=32564

            Bug ID: 32564
           Summary: AVX2 - aggressive broadcast generation instead of
                    memory operands
           Product: clang
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: normal
          Priority: P
         Component: LLVM Codegen
          Assignee: unassignedclangbugs at nondot.org
          Reporter: regis.portalez at gmail.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 18247
  --> https://bugs.llvm.org/attachment.cgi?id=18247&action=edit
open zip to see code reproducer (clang vs gcc and icc)

Following resolution of bug #20054
(https://bugs.llvm.org/show_bug.cgi?id=20054).

llvm codegen (x86 - avx2) now always generates broadcast instructions for splat
values, instead of using memory operands. 

See this reproducer : 

#include <immintrin.h>

__m256d mulconst(__m256d x) {
        const __m256d a = { 15.0, 15.0, 15.0, 15.0 };
        return _mm256_mul_pd(x, a);
}

generates [ -O3 -g  -S -mavx2 -mavx -mfma  ]

.LCPI0_0:
        .quad   4624633867356078080     # double 15
mulconst(double __vector(4)):                       # @mulconst(double
__vector(4))
        vbroadcastsd    ymm1, qword ptr [rip + .LCPI0_0]
        vmulpd  ymm0, ymm0, ymm1
        ret

This is legitimate when optimizing for code size, but not for speed.

Indeed: 
   vbroadcastsd is a supplemental instruction,
   the result consumes an extra register (which can further generate spilling)
   this prevents any use of memory operands, even with inline assembly. 


See attached larger reproducer to spot unnecessary spills (and compared
assemble between gcc 6.2 and clang 4.0.

-- 
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/20170407/a3cea5fc/attachment-0001.html>