https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645
Bug ID: 92645
Summary: Hand written vector code is 450 times slower when
compiled with GCC compared to Clang
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: hubicka at gcc dot gnu.org
Target Milestone: ---
Hi,
the attached are preprocessed files for Skia where Clang ifdefs was removed so
we get roughly same file for GCC and Clang. The internal loop of
_ZN3hsw16blit_row_color32EPjPKjij, _ZN3hsw16blit_row_color32EPjPKjij,
_ZN3hsw16blit_row_color32EPjPKjij and _ZN3hsw16blit_row_color32EPjPKjij
looks a lot worse when compiled by GCC then by clang.
I also added flatten to eliminate the inlining difference. Clang has heuristics
that makes functions with hand written vector code hot.
GCC code packs via stack:
0.43 â mov %ax,0xae(%rsp)
0.03 â movzbl 0x78(%rsp),%eax
0.02 â mov %cx,0xd8(%rsp)
0.02 â mov %ax,0xb0(%rsp)
0.54 â vpextrb $0x9,%xmm5,%eax
0.16 â mov %ax,0xb2(%rsp)
0.51 â vpextrb $0xa,%xmm5,%eax
0.21 â mov %ax,0xb4(%rsp)
0.16 â vpextrb $0xb,%xmm5,%eax
0.46 â mov %ax,0xb6(%rsp)
0.24 â vpextrb $0xc,%xmm5,%eax
0.28 â mov %ax,0xb8(%rsp)
0.41 â vpextrb $0xd,%xmm5,%eax
0.20 â mov %ax,0xba(%rsp)
0.47 â vpextrb $0xe,%xmm5,%eax
0.92 â mov %ax,0xbc(%rsp)
0.72 â vpextrb $0xf,%xmm5,%eax
1.24 â mov %ax,0xbe(%rsp)
10.94 â vmovdqa 0xa0(%rsp),%ymm4
0.02 â mov %cx,0xda(%rsp)
0.00 â mov %cx,0xdc(%rsp)
â mov %cx,0xde(%rsp)
10.34 â vpmullw 0xc0(%rsp),%ymm4,%ymm0
2.05 â vpaddw %ymm1,%ymm0,%ymm0
0.50 â vpaddw %ymm3,%ymm0,%ymm0
0.00 â mov %r9,0x58(%rsp)
0.52 â vpsrlw $0x8,%ymm0,%ymm0
0.39 â vpextrw $0x0,%xmm0,%eax
0.69 â mov %al,%r8b
0.17 â vpextrw $0x1,%xmm0,%eax
0.51 â mov %r8,0x50(%rsp)
6.87 â vmovdqa 0x50(%rsp),%xmm5
1.08 â vpinsrb $0x1,%eax,%xmm5,%xmm1
0.00 â vpextrw $0x2,%xmm0,%eax
0.73 â vpinsrb $0x2,%eax,%xmm1,%xmm1
0.02 â vpextrw $0x3,%xmm0,%eax
0.75 â vpinsrb $0x3,%eax,%xmm1,%xmm1
0.10 â vpextrw $0x4,%xmm0,%eax
0.98 â vpinsrb $0x4,%eax,%xmm1,%xmm1
0.16 â vpextrw $0x5,%xmm0,%eax
1.00 â vpinsrb $0x5,%eax,%xmm1,%xmm1
0.22 â vpextrw $0x6,%xmm0,%eax
1.10 â vpinsrb $0x6,%eax,%xmm1,%xmm1
0.30 â vpextrw $0x7,%xmm0,%eax
0.31 â vextracti128 $0x1,%ymm0,%xmm0
0.90 â vpinsrb $0x7,%eax,%xmm1,%xmm6
0.21 â vpextrw $0x0,%xmm0,%eax
0.35 â vmovaps %xmm6,0x50(%rsp)
1.15 â mov 0x58(%rsp),%r9
0.13 â mov 0x50(%rsp),%r8
0.29 â mov %al,%r9b
0.49 â mov %r8,0x50(%rsp)
0.07 â vpextrw $0x1,%xmm0,%eax
0.45 â mov %r9,0x58(%rsp)
7.08 â vmovdqa 0x50(%rsp),%xmm7
1.19 â vpinsrb $0x9,%eax,%xmm7,%xmm1
0.00 â vpextrw $0x2,%xmm0,%eax
0.78 â vpinsrb $0xa,%eax,%xmm1,%xmm1
0.00 â vpextrw $0x3,%xmm0,%eax
0.77 â vpinsrb $0xb,%eax,%xmm1,%xmm1
0.01 â vpextrw $0x4,%xmm0,%eax
0.86 â vpinsrb $0xc,%eax,%xmm1,%xmm1
0.03 â vpextrw $0x5,%xmm0,%eax
0.88 â vpinsrb $0xd,%eax,%xmm1,%xmm1
0.04 â vpextrw $0x6,%xmm0,%eax
0.97 â vpinsrb $0xe,%eax,%xmm1,%xmm1
0.08 â vpextrw $0x7,%xmm0,%eax
1.44 â vpinsrb $0xf,%eax,%xmm1,%xmm0
1.37 â vpextrd $0x1,%xmm0,%eax
0.13 â vinsertps $0xe,%xmm0,%xmm0,%xmm1
0.02 â vmovaps %xmm0,0x50(%rsp)
2.17 â vpinsrd $0x1,%eax,%xmm1,%xmm1
....
Clang code:
Percentâ vpmullw %ymm0,%ymm2,%ymm2
â vpaddw %ymm1,%ymm2,%ymm2
â vpsrlw $0x8,%ymm2,%ymm2
â vextracti128 $0x1,%ymm2,%xmm3
â vpackuswb %xmm3,%xmm2,%xmm2
â vmovdqu %xmm2,(%rdi)
â add $0x10,%rsi
â add $0x10,%rdi
â mov %r9d,%eax
â cmp $0x4,%r9d
â â jae 39179b0 <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0xa0>
â â jmp 3917a02 <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0xf2>
â mov %edx,%eax
0.29 â cmp $0x4,%r9d
0.00 â â jb 3917a02 <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0xf2>
0.07 â nop
3.95 â vpmovzxbw (%rsi),%ymm2
13.41 â vpmullw %ymm0,%ymm2,%ymm2
13.87 â vpaddw %ymm1,%ymm2,%ymm2
2.93 â vpsrlw $0x8,%ymm2,%ymm2
0.84 â vextracti128 $0x1,%ymm2,%xmm3
9.98 â vpackuswb %xmm3,%xmm2,%xmm2
6.89 â vmovdqu %xmm2,(%rdi)
0.57 â vpmovzxbw 0x10(%rsi),%ymm2
4.02 â vpmullw %ymm0,%ymm2,%ymm2
12.15 â vpaddw %ymm1,%ymm2,%ymm2
2.02 â vpsrlw $0x8,%ymm2,%ymm2
1.22 â vextracti128 $0x1,%ymm2,%xmm3
8.04 â vpackuswb %xmm3,%xmm2,%xmm2
7.09 â vmovdqu %xmm2,0x10(%rdi)
0.19 â add $0x20,%rsi
0.19 â add $0x20,%rdi
0.57 â add $0xfffffff8,%eax
7.26 â cmp $0x3,%eax
0.29 â â jg 39179b0 <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0xa0>
0.02 â and $0x3,%edx
Percentâ test %edx,%edx
0.04 â â jle 3917af1 <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0x1e1>
â mov %ecx,%eax
â shr $0x18,%eax
â shr $0x1f,%ecx
â add %eax,%ecx
â vmovd %ecx,%xmm0
â vpbroadcastb %xmm0,%xmm0
â mov %r8,%rax
â shl $0x20,%rax
â or %r8,%rax
â vmovq %rax,%xmm1
â vpbroadcastq %xmm1,%xmm1
â vpmovzxbw %xmm1,%ymm1
â vpmovzxbw %xmm0,%ymm0
â vpsllw $0x8,%ymm1,%ymm1
â vpor 0x1f3eb35(%rip),%ymm1,%ymm1 # 5856580
<SkNamedGamut::kAdobeRGB+0x858>
â test $0x1,%dl
â â jne 3917a5c <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0x14c>
â mov %edx,%eax
â cmp $0x1,%edx
â â jne 3917a90 <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0x180>
â â jmpq 3917af1 <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0x1e1>
â lea -0x1(%rdx),%eax
â mov (%rsi),%ecx
â add $0x4,%rsi
â vmovq %rcx,%xmm2
â vpmovzxbw %xmm2,%xmm2
â vpmullw %xmm0,%xmm2,%xmm2
â vpaddw %xmm1,%xmm2,%xmm2
â vpsrlw $0x8,%xmm2,%xmm2
â vpackuswb %xmm0,%xmm2,%xmm2
â vmovq %xmm2,%rcx
â mov %ecx,(%rdi)
â add $0x4,%rdi
â cmp $0x1,%edx
â xor %ecx,%ecx
â nop
â nop
â mov (%rsi,%rcx,1),%edx
â vmovq %rdx,%xmm2
â vpmovzxbw %xmm2,%xmm2
â vpmullw %xmm0,%xmm2,%xmm2
â vpaddw %xmm1,%xmm2,%xmm2
â vpsrlw $0x8,%xmm2,%xmm2
â vpackuswb %xmm0,%xmm2,%xmm2
â vmovd %xmm2,(%rdi,%rcx,1)
â mov 0x4(%rsi,%rcx,1),%edx
â vmovq %rdx,%xmm2
â vpmovzxbw %xmm2,%xmm2
â vpmullw %xmm0,%xmm2,%xmm2
â vpaddw %xmm1,%xmm2,%xmm2
â vpsrlw $0x8,%xmm2,%xmm2
â vpackuswb %xmm0,%xmm2,%xmm2
â vmovd %xmm2,0x4(%rdi,%rcx,1)
â add $0x8,%rcx
â add $0xfffffffe,%eax
â â jg 3917aa0 <hsw::blit_row_color32(unsigned int*,
unsigned int const*, int, unsigned int)+0x190>
0.38 â vzeroupper
0.26 â â retq
