https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114350
Bug ID: 114350
Summary: missing support for SVE widening floating point
conversion
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: tnfchris at gcc dot gnu.org
Target Milestone: ---
Target: aarch64*
The following example:
#include <arm_sve.h>
svfloat64_t widening (svint32_t a)
{
svbool_t pred = svptrue_b32 ();
svint64_t cvt = svreinterpret_s64_s32 (a);
svint64_t ext = svextw_s64_x (pred, cvt);
svfloat64_t res = svcvt_f64_s64_x (pred, ext);
return res;
}
compiled with -Ofast -march=armv9-a generates:
widening(__SVInt32_t):
ptrue p3.b, all
sxtw z0.d, p3/m, z0.d
scvtf z0.d, p3/m, z0.d
ret
but SVE has widening and narrowing floating point conversions, as such this
should generate:
widening(__SVInt32_t):
ptrue p3.b, all
scvtf z0.d, p3/m, z0.s
ret
The autovec equivalent is:
void f(int n, double *data) {
for (int i=0;i<n;i++) {
data[i] = i;
}
}
which generates:
.L5:
mov z29.d, z31.d
sxtw z29.d, p7/m, z29.d
scvtf z29.d, p7/m, z29.d
st1d z29.d, p6, [x1, x2, lsl 3]
add z31.s, z31.s, z30.s
incd x2
whilelo p6.d, w2, w0
b.any .L5
note that scalar has the widening variant as well (which we do use) and we
account for in vectorizer costing:
(double) i_14 1 times scalar_stmt costs 1 in epilogue
_4 1 times scalar_store costs 1 in epilogue
Adv. SIMD costing is right:
(double) i_14 2 times vec_promote_demote costs 4 in body
but SVE costing is wrong:
(double) i_14 2 times vector_stmt costs 2 in body
which makes SVE seem as expensive than Adv. SIMD.
Note that we can also use the widening instruction for Adv. SIMD on a SVE
capable system.