generate better fallback code for vmin/vmax intrinsics

db4
Joe Groff 2009-10-20 18:45:35 -05:00
parent 6f72c3ca24
commit cb36111a3c
3 changed files with 17 additions and 4 deletions

View File

@ -163,8 +163,8 @@ IN: compiler.cfg.intrinsics
{ math.vectors.simd.intrinsics:(simd-v*) [ [ ^^mul-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vs*) [ [ ^^saturated-mul-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-v/) [ [ ^^div-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmin) [ [ ^^min-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmax) [ [ ^^max-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmin) [ [ generate-min-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmax) [ [ generate-max-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-v.) [ [ ^^dot-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vabs) [ [ generate-abs-vector ] emit-unary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vsqrt) [ [ ^^sqrt-vector ] emit-unary-vector-op ] }

View File

@ -265,3 +265,16 @@ MACRO: if-literals-match ( quots -- )
]
} cond ;
: generate-min-vector ( src1 src2 rep -- dst )
dup %min-vector-reps member?
[ ^^min-vector ] [
[ cc< generate-compare-vector ]
[ generate-blend-vector ] 3bi
] if ;
: generate-max-vector ( src1 src2 rep -- dst )
dup %max-vector-reps member?
[ ^^max-vector ] [
[ cc> generate-compare-vector ]
[ generate-blend-vector ] 3bi
] if ;

View File

@ -163,8 +163,8 @@ M: vector-rep supported-simd-op?
{ \ (simd-v*) [ %mul-vector-reps ] }
{ \ (simd-vs*) [ %saturated-mul-vector-reps ] }
{ \ (simd-v/) [ %div-vector-reps ] }
{ \ (simd-vmin) [ %min-vector-reps ] }
{ \ (simd-vmax) [ %max-vector-reps ] }
{ \ (simd-vmin) [ %min-vector-reps cc< %compare-vector-reps union ] }
{ \ (simd-vmax) [ %max-vector-reps cc> %compare-vector-reps union ] }
{ \ (simd-v.) [ %dot-vector-reps ] }
{ \ (simd-vsqrt) [ %sqrt-vector-reps ] }
{ \ (simd-sum) [ %horizontal-add-vector-reps ] }