[llvm-bugs] [Bug 45378] New: [X86] recent codegen regression for vector load and reduce

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=45378

            Bug ID: 45378
           Summary: [X86] recent codegen regression for vector load and
                    reduce
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: fedor.v.sergeev at gmail.com
                CC: craig.topper at gmail.com, llvm-bugs at lists.llvm.org,
                    llvm-dev at redking.me.uk, spatel+llvm at rotateright.com

The following IR testcase started producing suboptimal codegen on recent X86
hardware (haswell+):

] cat load-and-reduce.ll
declare i64 @llvm.experimental.vector.reduce.or.v2i64(<2 x i64>) 
declare void @TrapFunc(i64)

define void @parseHeaders(i64 * %ptr) {
  %vptr = bitcast i64 * %ptr to <2 x i64> *
  %vload = load <2 x i64>, <2 x i64> * %vptr, align 8
  %vreduce = call i64 @llvm.experimental.vector.reduce.or.v2i64(<2 x i64>
%vload)
  %vcheck = icmp eq i64 %vreduce, 0
  br i1 %vcheck, label %ret, label %trap
trap:
  %v2 = extractelement <2 x i64> %vload, i32 1
  call void @TrapFunc(i64 %v2)
  ret void
ret:
  ret void
}
] bin/llc -filetype=asm load-and-reduce.ll -mtriple=x86_64-unknown-linux-gnu
-mcpu=skylake -o - | grep -A8 movdq
        vmovdqu (%rdi), %xmm0
        vpbroadcastq    8(%rdi), %xmm1
        vpor    %xmm1, %xmm0, %xmm1
        vmovq   %xmm1, %rax
        testq   %rax, %rax
        je      .LBB0_2
# %bb.1:                                # %trap
        vpextrq $1, %xmm0, %rdi
        callq   TrapFunc
]

There are two loads here, which does not seem to be efficient solution here.

Note, that for ivybridge- it generates load+shuffle instead of a load+load,
which looks like a better codegen:

] bin/llc -filetype=asm load-and-reduce.ll -mtriple=x86_64-unknown-linux-gnu
-mcpu=ivybridge -o - | grep -A8 movdq
        vmovdqu (%rdi), %xmm0
        vpshufd $78, %xmm0, %xmm1       # xmm1 = xmm0[2,3,0,1]
        vpor    %xmm1, %xmm0, %xmm1
        vmovq   %xmm1, %rax
        testq   %rax, %rax
        je      .LBB0_2
# %bb.1:                                # %trap
        vpextrq $1, %xmm0, %rdi
        callq   TrapFunc
]

This appears to be a rather recent change in behavior, caused by 
commit e48b536be66b60b05fa77b85258e6cf2ec417220
Author: Sanjay Patel <spatel at rotateright.com>
Date:   Sun Feb 16 10:32:56 2020 -0500

    [x86] form broadcast of scalar memop even with >1 use
----
Before that commit all X86 targets generate the same load+shuffle sequence.

-- 
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/20200331/add581a3/attachment.html>