http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56083
Bug #: 56083
Summary: Vectorizer uses xor/movlps/movhps rather than movups
Classification: Unclassified
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: minor
Priority: P3
Component: c
AssignedTo: [email protected]
ReportedBy: [email protected]
Unnecessarily complex machine code is generated on x86-64. Perhaps there is a
reason for this but to me it seems like the compiler is failing to optimize
properly. Asm code labels changed and comments added, other than that they are
are produced by the respective compilers for this C code:
void negate(float* fvec) {
fvec[0] *= -1;
fvec[1] *= -1;
fvec[2] *= -1;
fvec[3] *= -1;
}
.LC0:
.long 2147483648 # float -0.000000e+00
.long 2147483648 # float -0.000000e+00
.long 2147483648 # float -0.000000e+00
.long 2147483648 # float -0.000000e+00
.LC1:
.long 3212836864 # float -1.000000e+00
.long 3212836864 # float -1.000000e+00
.long 3212836864 # float -1.000000e+00
.long 3212836864 # float -1.000000e+00
# Hand optimized
movups (%rdi), %xmm0 # Load fvec into xmm0
xorps .LC0(%rip), %xmm0 # Invert sign bits of each float (xor -0.0f)
movups %xmm0, (%rdi) # Store xmm0 to fvec
# Clang -O3 -mllvm -vectorize
movups (%rdi), %xmm0 # Load fvec
mulps .LC1(%rip), %xmm0 # Multiply by negative ones
movups %xmm0, (%rdi) # Store fvec
# GCC 4.7.2 -O3
xorps %xmm0, %xmm0 # Zero xmm0 (unnecessary)
movlps (%rdi), %xmm0 # Load two lower floats
movhps 8(%rdi), %xmm0 # Load two higher floats
xorps .LC0(%rip), %xmm0 # Invert sign bits
movlps %xmm0, (%rdi) # Store two lower floats
movhps %xmm0, 8(%rdi) # Store two higher floats
