https://llvm.org/bugs/show_bug.cgi?id=31690
Bug ID: 31690
Summary: Simple addition reduction not vectorised when limit is
48 or less.
Product: libraries
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: Loop Optimizer
Assignee: unassignedb...@nondot.org
Reporter: drr...@gmail.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Consider this simple loop
float f(float x[]) {
float p = 1.0;
for (int i = 0; i < 48; i++)
p += x[i];
return p;
}
clang 3.9.1 gives
.LCPI0_0:
.long 1065353216 # float 1
f: # @f
vmovss xmm0, dword ptr [rdi] # xmm0 = mem[0],zero,zero,zero
vaddss xmm0, xmm0, dword ptr [rip + .LCPI0_0]
vaddss xmm0, xmm0, dword ptr [rdi + 4]
vaddss xmm0, xmm0, dword ptr [rdi + 8]
vaddss xmm0, xmm0, dword ptr [rdi + 12]
vaddss xmm0, xmm0, dword ptr [rdi + 16]
vaddss xmm0, xmm0, dword ptr [rdi + 20]
vaddss xmm0, xmm0, dword ptr [rdi + 24]
vaddss xmm0, xmm0, dword ptr [rdi + 28]
vaddss xmm0, xmm0, dword ptr [rdi + 32]
vaddss xmm0, xmm0, dword ptr [rdi + 36]
vaddss xmm0, xmm0, dword ptr [rdi + 40]
vaddss xmm0, xmm0, dword ptr [rdi + 44]
vaddss xmm0, xmm0, dword ptr [rdi + 48]
vaddss xmm0, xmm0, dword ptr [rdi + 52]
vaddss xmm0, xmm0, dword ptr [rdi + 56]
vaddss xmm0, xmm0, dword ptr [rdi + 60]
vaddss xmm0, xmm0, dword ptr [rdi + 64]
vaddss xmm0, xmm0, dword ptr [rdi + 68]
vaddss xmm0, xmm0, dword ptr [rdi + 72]
vaddss xmm0, xmm0, dword ptr [rdi + 76]
vaddss xmm0, xmm0, dword ptr [rdi + 80]
vaddss xmm0, xmm0, dword ptr [rdi + 84]
vaddss xmm0, xmm0, dword ptr [rdi + 88]
vaddss xmm0, xmm0, dword ptr [rdi + 92]
vaddss xmm0, xmm0, dword ptr [rdi + 96]
vaddss xmm0, xmm0, dword ptr [rdi + 100]
vaddss xmm0, xmm0, dword ptr [rdi + 104]
vaddss xmm0, xmm0, dword ptr [rdi + 108]
vaddss xmm0, xmm0, dword ptr [rdi + 112]
vaddss xmm0, xmm0, dword ptr [rdi + 116]
vaddss xmm0, xmm0, dword ptr [rdi + 120]
vaddss xmm0, xmm0, dword ptr [rdi + 124]
vaddss xmm0, xmm0, dword ptr [rdi + 128]
vaddss xmm0, xmm0, dword ptr [rdi + 132]
vaddss xmm0, xmm0, dword ptr [rdi + 136]
vaddss xmm0, xmm0, dword ptr [rdi + 140]
vaddss xmm0, xmm0, dword ptr [rdi + 144]
vaddss xmm0, xmm0, dword ptr [rdi + 148]
vaddss xmm0, xmm0, dword ptr [rdi + 152]
vaddss xmm0, xmm0, dword ptr [rdi + 156]
vaddss xmm0, xmm0, dword ptr [rdi + 160]
vaddss xmm0, xmm0, dword ptr [rdi + 164]
vaddss xmm0, xmm0, dword ptr [rdi + 168]
vaddss xmm0, xmm0, dword ptr [rdi + 172]
vaddss xmm0, xmm0, dword ptr [rdi + 176]
vaddss xmm0, xmm0, dword ptr [rdi + 180]
vaddss xmm0, xmm0, dword ptr [rdi + 184]
vaddss xmm0, xmm0, dword ptr [rdi + 188]
ret
This is fully unrolled but not vectorized.
However, using icc we get:
f:
movups xmm13, XMMWORD PTR [rdi] #4.10
movups xmm0, XMMWORD PTR [16+rdi] #4.10
movups xmm6, XMMWORD PTR [32+rdi] #4.10
movups xmm1, XMMWORD PTR [48+rdi] #4.10
movups xmm9, XMMWORD PTR [64+rdi] #4.10
movups xmm2, XMMWORD PTR [80+rdi] #4.10
movups xmm7, XMMWORD PTR [96+rdi] #4.10
movups xmm3, XMMWORD PTR [112+rdi] #4.10
movups xmm10, XMMWORD PTR [128+rdi] #4.10
movups xmm4, XMMWORD PTR [144+rdi] #4.10
movups xmm8, XMMWORD PTR [160+rdi] #4.10
movups xmm5, XMMWORD PTR [176+rdi] #4.10
addps xmm13, xmm0 #2.11
addps xmm6, xmm1 #2.11
addps xmm9, xmm2 #2.11
addps xmm7, xmm3 #2.11
addps xmm10, xmm4 #2.11
addps xmm8, xmm5 #2.11
addps xmm13, xmm6 #2.11
addps xmm9, xmm7 #2.11
addps xmm10, xmm8 #2.11
addps xmm13, xmm9 #2.11
addps xmm13, xmm10 #2.11
movaps xmm11, xmm13 #2.11
movhlps xmm11, xmm13 #2.11
addps xmm13, xmm11 #2.11
movaps xmm12, xmm13 #2.11
shufps xmm12, xmm13, 245 #2.11
addss xmm13, xmm12 #2.11
addss xmm13, DWORD PTR .L_2il0floatpacket.0[rip] #2.11
movaps xmm0, xmm13 #5.10
ret #5.10
.L_2il0floatpacket.0:
.long 0x3f800000
This is both vectorised and unrolled.
If you increase the loop limit from 48 to 64, say, then clang/llvm stops
unrolling and does manage to vectorise the code.
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
