2017-05-12 23:56:39 +08:00
|
|
|
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE2
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512F
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512BW
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512dq | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512DQ
|
|
|
|
|
|
|
|
|
|
; AVX1 has support for 256-bit bitwise logic because the FP variants were included.
|
|
|
|
|
; If using those ops requires extra insert/extract though, it's probably not worth it.
|
|
|
|
|
|
|
|
|
|
define <8 x i32> @PR32790(<8 x i32> %a, <8 x i32> %b, <8 x i32> %c, <8 x i32> %d) {
|
|
|
|
|
; SSE-LABEL: PR32790:
|
2017-12-05 01:18:51 +08:00
|
|
|
; SSE: # %bb.0:
|
2017-05-12 23:56:39 +08:00
|
|
|
; SSE-NEXT: paddd %xmm2, %xmm0
|
|
|
|
|
; SSE-NEXT: paddd %xmm3, %xmm1
|
|
|
|
|
; SSE-NEXT: pand %xmm5, %xmm1
|
|
|
|
|
; SSE-NEXT: pand %xmm4, %xmm0
|
|
|
|
|
; SSE-NEXT: psubd %xmm6, %xmm0
|
|
|
|
|
; SSE-NEXT: psubd %xmm7, %xmm1
|
|
|
|
|
; SSE-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX1-LABEL: PR32790:
|
2017-12-05 01:18:51 +08:00
|
|
|
; AVX1: # %bb.0:
|
[DAGCombiner] use narrow vector ops to eliminate concat/extract (PR32790)
In the best case:
extract (binop (concat X1, X2), (concat Y1, Y2)), N --> binop XN, YN
...we kill all of the extract/concat and just have narrow binops remaining.
If only one of the binop operands is amenable, this transform is still
worthwhile because we kill some of the extract/concat.
Optional bitcasting makes the code more complicated, but there doesn't
seem to be a way to avoid that.
The TODO about extending to more than bitwise logic is there because we really
will regress several x86 tests including madd, psad, and even a plain
integer-multiply-by-2 or shift-left-by-1. I don't think there's anything
fundamentally wrong with this patch that would cause those regressions; those
folds are just missing or brittle.
If we extend to more binops, I found that this patch will fire on at least one
non-x86 regression test. There's an ARM NEON test in
test/CodeGen/ARM/coalesce-subregs.ll with a pattern like:
t5: v2f32 = vector_shuffle<0,3> t2, t4
t6: v1i64 = bitcast t5
t8: v1i64 = BUILD_VECTOR Constant:i64<0>
t9: v2i64 = concat_vectors t6, t8
t10: v4f32 = bitcast t9
t12: v4f32 = fmul t11, t10
t13: v2i64 = bitcast t12
t16: v1i64 = extract_subvector t13, Constant:i32<0>
There was no functional change in the codegen from this transform from what I
could see though.
For the x86 test changes:
1. PR32790() is the closest call. We don't reduce the AVX1 instruction count in that case,
but we improve throughput. Also, on a core like Jaguar that double-pumps 256-bit ops,
there's an unseen win because two 128-bit ops have the same cost as the wider 256-bit op.
SSE/AVX2/AXV512 are not affected which is expected because only AVX1 has the extract/concat
ops to match the pattern.
2. do_not_use_256bit_op() is the best case. Everyone wins by avoiding the concat/extract.
Related bug for IR filed as: https://bugs.llvm.org/show_bug.cgi?id=33026
3. The SSE diffs in vector-trunc-math.ll are just scheduling/RA, so nothing real AFAICT.
4. The AVX1 diffs in vector-tzcnt-256.ll are all the same pattern: we reduced the instruction
count by one in each case by eliminating two insert/extract while adding one narrower logic op.
https://bugs.llvm.org/show_bug.cgi?id=32790
Differential Revision: https://reviews.llvm.org/D33137
llvm-svn: 303997
2017-05-26 23:33:18 +08:00
|
|
|
; AVX1-NEXT: vpaddd %xmm1, %xmm0, %xmm4
|
|
|
|
|
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
|
2017-05-12 23:56:39 +08:00
|
|
|
; AVX1-NEXT: vpaddd %xmm1, %xmm0, %xmm0
|
[DAGCombiner] use narrow vector ops to eliminate concat/extract (PR32790)
In the best case:
extract (binop (concat X1, X2), (concat Y1, Y2)), N --> binop XN, YN
...we kill all of the extract/concat and just have narrow binops remaining.
If only one of the binop operands is amenable, this transform is still
worthwhile because we kill some of the extract/concat.
Optional bitcasting makes the code more complicated, but there doesn't
seem to be a way to avoid that.
The TODO about extending to more than bitwise logic is there because we really
will regress several x86 tests including madd, psad, and even a plain
integer-multiply-by-2 or shift-left-by-1. I don't think there's anything
fundamentally wrong with this patch that would cause those regressions; those
folds are just missing or brittle.
If we extend to more binops, I found that this patch will fire on at least one
non-x86 regression test. There's an ARM NEON test in
test/CodeGen/ARM/coalesce-subregs.ll with a pattern like:
t5: v2f32 = vector_shuffle<0,3> t2, t4
t6: v1i64 = bitcast t5
t8: v1i64 = BUILD_VECTOR Constant:i64<0>
t9: v2i64 = concat_vectors t6, t8
t10: v4f32 = bitcast t9
t12: v4f32 = fmul t11, t10
t13: v2i64 = bitcast t12
t16: v1i64 = extract_subvector t13, Constant:i32<0>
There was no functional change in the codegen from this transform from what I
could see though.
For the x86 test changes:
1. PR32790() is the closest call. We don't reduce the AVX1 instruction count in that case,
but we improve throughput. Also, on a core like Jaguar that double-pumps 256-bit ops,
there's an unseen win because two 128-bit ops have the same cost as the wider 256-bit op.
SSE/AVX2/AXV512 are not affected which is expected because only AVX1 has the extract/concat
ops to match the pattern.
2. do_not_use_256bit_op() is the best case. Everyone wins by avoiding the concat/extract.
Related bug for IR filed as: https://bugs.llvm.org/show_bug.cgi?id=33026
3. The SSE diffs in vector-trunc-math.ll are just scheduling/RA, so nothing real AFAICT.
4. The AVX1 diffs in vector-tzcnt-256.ll are all the same pattern: we reduced the instruction
count by one in each case by eliminating two insert/extract while adding one narrower logic op.
https://bugs.llvm.org/show_bug.cgi?id=32790
Differential Revision: https://reviews.llvm.org/D33137
llvm-svn: 303997
2017-05-26 23:33:18 +08:00
|
|
|
; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpsubd %xmm1, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm4, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpsubd %xmm3, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
|
2017-05-12 23:56:39 +08:00
|
|
|
; AVX1-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX2-LABEL: PR32790:
|
2017-12-05 01:18:51 +08:00
|
|
|
; AVX2: # %bb.0:
|
2017-05-12 23:56:39 +08:00
|
|
|
; AVX2-NEXT: vpaddd %ymm1, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpand %ymm2, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpsubd %ymm3, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX512-LABEL: PR32790:
|
2017-12-05 01:18:51 +08:00
|
|
|
; AVX512: # %bb.0:
|
2017-05-12 23:56:39 +08:00
|
|
|
; AVX512-NEXT: vpaddd %ymm1, %ymm0, %ymm0
|
|
|
|
|
; AVX512-NEXT: vpand %ymm2, %ymm0, %ymm0
|
|
|
|
|
; AVX512-NEXT: vpsubd %ymm3, %ymm0, %ymm0
|
|
|
|
|
; AVX512-NEXT: retq
|
|
|
|
|
%add = add <8 x i32> %a, %b
|
|
|
|
|
%and = and <8 x i32> %add, %c
|
|
|
|
|
%sub = sub <8 x i32> %and, %d
|
|
|
|
|
ret <8 x i32> %sub
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; In a more extreme case, even the later AVX targets should avoid extract/insert just
|
|
|
|
|
; because 256-bit ops are supported.
|
|
|
|
|
|
|
|
|
|
define <4 x i32> @do_not_use_256bit_op(<4 x i32> %a, <4 x i32> %b, <4 x i32> %c, <4 x i32> %d) {
|
|
|
|
|
; SSE-LABEL: do_not_use_256bit_op:
|
2017-12-05 01:18:51 +08:00
|
|
|
; SSE: # %bb.0:
|
2017-05-12 23:56:39 +08:00
|
|
|
; SSE-NEXT: pand %xmm2, %xmm0
|
[DAGCombiner] use narrow vector ops to eliminate concat/extract (PR32790)
In the best case:
extract (binop (concat X1, X2), (concat Y1, Y2)), N --> binop XN, YN
...we kill all of the extract/concat and just have narrow binops remaining.
If only one of the binop operands is amenable, this transform is still
worthwhile because we kill some of the extract/concat.
Optional bitcasting makes the code more complicated, but there doesn't
seem to be a way to avoid that.
The TODO about extending to more than bitwise logic is there because we really
will regress several x86 tests including madd, psad, and even a plain
integer-multiply-by-2 or shift-left-by-1. I don't think there's anything
fundamentally wrong with this patch that would cause those regressions; those
folds are just missing or brittle.
If we extend to more binops, I found that this patch will fire on at least one
non-x86 regression test. There's an ARM NEON test in
test/CodeGen/ARM/coalesce-subregs.ll with a pattern like:
t5: v2f32 = vector_shuffle<0,3> t2, t4
t6: v1i64 = bitcast t5
t8: v1i64 = BUILD_VECTOR Constant:i64<0>
t9: v2i64 = concat_vectors t6, t8
t10: v4f32 = bitcast t9
t12: v4f32 = fmul t11, t10
t13: v2i64 = bitcast t12
t16: v1i64 = extract_subvector t13, Constant:i32<0>
There was no functional change in the codegen from this transform from what I
could see though.
For the x86 test changes:
1. PR32790() is the closest call. We don't reduce the AVX1 instruction count in that case,
but we improve throughput. Also, on a core like Jaguar that double-pumps 256-bit ops,
there's an unseen win because two 128-bit ops have the same cost as the wider 256-bit op.
SSE/AVX2/AXV512 are not affected which is expected because only AVX1 has the extract/concat
ops to match the pattern.
2. do_not_use_256bit_op() is the best case. Everyone wins by avoiding the concat/extract.
Related bug for IR filed as: https://bugs.llvm.org/show_bug.cgi?id=33026
3. The SSE diffs in vector-trunc-math.ll are just scheduling/RA, so nothing real AFAICT.
4. The AVX1 diffs in vector-tzcnt-256.ll are all the same pattern: we reduced the instruction
count by one in each case by eliminating two insert/extract while adding one narrower logic op.
https://bugs.llvm.org/show_bug.cgi?id=32790
Differential Revision: https://reviews.llvm.org/D33137
llvm-svn: 303997
2017-05-26 23:33:18 +08:00
|
|
|
; SSE-NEXT: pand %xmm3, %xmm1
|
2017-05-12 23:56:39 +08:00
|
|
|
; SSE-NEXT: psubd %xmm1, %xmm0
|
|
|
|
|
; SSE-NEXT: retq
|
|
|
|
|
;
|
[DAGCombiner] use narrow vector ops to eliminate concat/extract (PR32790)
In the best case:
extract (binop (concat X1, X2), (concat Y1, Y2)), N --> binop XN, YN
...we kill all of the extract/concat and just have narrow binops remaining.
If only one of the binop operands is amenable, this transform is still
worthwhile because we kill some of the extract/concat.
Optional bitcasting makes the code more complicated, but there doesn't
seem to be a way to avoid that.
The TODO about extending to more than bitwise logic is there because we really
will regress several x86 tests including madd, psad, and even a plain
integer-multiply-by-2 or shift-left-by-1. I don't think there's anything
fundamentally wrong with this patch that would cause those regressions; those
folds are just missing or brittle.
If we extend to more binops, I found that this patch will fire on at least one
non-x86 regression test. There's an ARM NEON test in
test/CodeGen/ARM/coalesce-subregs.ll with a pattern like:
t5: v2f32 = vector_shuffle<0,3> t2, t4
t6: v1i64 = bitcast t5
t8: v1i64 = BUILD_VECTOR Constant:i64<0>
t9: v2i64 = concat_vectors t6, t8
t10: v4f32 = bitcast t9
t12: v4f32 = fmul t11, t10
t13: v2i64 = bitcast t12
t16: v1i64 = extract_subvector t13, Constant:i32<0>
There was no functional change in the codegen from this transform from what I
could see though.
For the x86 test changes:
1. PR32790() is the closest call. We don't reduce the AVX1 instruction count in that case,
but we improve throughput. Also, on a core like Jaguar that double-pumps 256-bit ops,
there's an unseen win because two 128-bit ops have the same cost as the wider 256-bit op.
SSE/AVX2/AXV512 are not affected which is expected because only AVX1 has the extract/concat
ops to match the pattern.
2. do_not_use_256bit_op() is the best case. Everyone wins by avoiding the concat/extract.
Related bug for IR filed as: https://bugs.llvm.org/show_bug.cgi?id=33026
3. The SSE diffs in vector-trunc-math.ll are just scheduling/RA, so nothing real AFAICT.
4. The AVX1 diffs in vector-tzcnt-256.ll are all the same pattern: we reduced the instruction
count by one in each case by eliminating two insert/extract while adding one narrower logic op.
https://bugs.llvm.org/show_bug.cgi?id=32790
Differential Revision: https://reviews.llvm.org/D33137
llvm-svn: 303997
2017-05-26 23:33:18 +08:00
|
|
|
; AVX-LABEL: do_not_use_256bit_op:
|
2017-12-05 01:18:51 +08:00
|
|
|
; AVX: # %bb.0:
|
[DAGCombiner] use narrow vector ops to eliminate concat/extract (PR32790)
In the best case:
extract (binop (concat X1, X2), (concat Y1, Y2)), N --> binop XN, YN
...we kill all of the extract/concat and just have narrow binops remaining.
If only one of the binop operands is amenable, this transform is still
worthwhile because we kill some of the extract/concat.
Optional bitcasting makes the code more complicated, but there doesn't
seem to be a way to avoid that.
The TODO about extending to more than bitwise logic is there because we really
will regress several x86 tests including madd, psad, and even a plain
integer-multiply-by-2 or shift-left-by-1. I don't think there's anything
fundamentally wrong with this patch that would cause those regressions; those
folds are just missing or brittle.
If we extend to more binops, I found that this patch will fire on at least one
non-x86 regression test. There's an ARM NEON test in
test/CodeGen/ARM/coalesce-subregs.ll with a pattern like:
t5: v2f32 = vector_shuffle<0,3> t2, t4
t6: v1i64 = bitcast t5
t8: v1i64 = BUILD_VECTOR Constant:i64<0>
t9: v2i64 = concat_vectors t6, t8
t10: v4f32 = bitcast t9
t12: v4f32 = fmul t11, t10
t13: v2i64 = bitcast t12
t16: v1i64 = extract_subvector t13, Constant:i32<0>
There was no functional change in the codegen from this transform from what I
could see though.
For the x86 test changes:
1. PR32790() is the closest call. We don't reduce the AVX1 instruction count in that case,
but we improve throughput. Also, on a core like Jaguar that double-pumps 256-bit ops,
there's an unseen win because two 128-bit ops have the same cost as the wider 256-bit op.
SSE/AVX2/AXV512 are not affected which is expected because only AVX1 has the extract/concat
ops to match the pattern.
2. do_not_use_256bit_op() is the best case. Everyone wins by avoiding the concat/extract.
Related bug for IR filed as: https://bugs.llvm.org/show_bug.cgi?id=33026
3. The SSE diffs in vector-trunc-math.ll are just scheduling/RA, so nothing real AFAICT.
4. The AVX1 diffs in vector-tzcnt-256.ll are all the same pattern: we reduced the instruction
count by one in each case by eliminating two insert/extract while adding one narrower logic op.
https://bugs.llvm.org/show_bug.cgi?id=32790
Differential Revision: https://reviews.llvm.org/D33137
llvm-svn: 303997
2017-05-26 23:33:18 +08:00
|
|
|
; AVX-NEXT: vpand %xmm2, %xmm0, %xmm0
|
|
|
|
|
; AVX-NEXT: vpand %xmm3, %xmm1, %xmm1
|
|
|
|
|
; AVX-NEXT: vpsubd %xmm1, %xmm0, %xmm0
|
|
|
|
|
; AVX-NEXT: retq
|
2017-05-12 23:56:39 +08:00
|
|
|
%concat1 = shufflevector <4 x i32> %a, <4 x i32> %b, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
|
|
|
|
|
%concat2 = shufflevector <4 x i32> %c, <4 x i32> %d, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
|
|
|
|
|
%and = and <8 x i32> %concat1, %concat2
|
|
|
|
|
%extract1 = shufflevector <8 x i32> %and, <8 x i32> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
|
|
|
|
|
%extract2 = shufflevector <8 x i32> %and, <8 x i32> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
|
|
|
|
|
%sub = sub <4 x i32> %extract1, %extract2
|
|
|
|
|
ret <4 x i32> %sub
|
|
|
|
|
}
|
|
|
|
|
|
2018-11-01 23:41:12 +08:00
|
|
|
; When extracting from a vector binop, the source width should be a multiple of the destination width.
|
|
|
|
|
; https://bugs.llvm.org/show_bug.cgi?id=39511
|
|
|
|
|
|
|
|
|
|
define <3 x float> @PR39511(<4 x float> %t0, <3 x float>* %b) {
|
|
|
|
|
; SSE-LABEL: PR39511:
|
|
|
|
|
; SSE: # %bb.0:
|
|
|
|
|
; SSE-NEXT: addps {{.*}}(%rip), %xmm0
|
|
|
|
|
; SSE-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX-LABEL: PR39511:
|
|
|
|
|
; AVX: # %bb.0:
|
|
|
|
|
; AVX-NEXT: vaddps {{.*}}(%rip), %xmm0, %xmm0
|
|
|
|
|
; AVX-NEXT: retq
|
|
|
|
|
%add = fadd <4 x float> %t0, <float 1.0, float 2.0, float 3.0, float 4.0>
|
|
|
|
|
%ext = shufflevector <4 x float> %add, <4 x float> undef, <3 x i32> <i32 0, i32 1, i32 2>
|
|
|
|
|
ret <3 x float> %ext
|
|
|
|
|
}
|
|
|
|
|
|
2018-12-06 01:10:30 +08:00
|
|
|
; When extracting from a vector binop, we need to be extracting
|
|
|
|
|
; by a width of at least 1 of the original vector elements.
|
|
|
|
|
; https://bugs.llvm.org/show_bug.cgi?id=39893
|
|
|
|
|
|
|
|
|
|
define <2 x i8> @PR39893(<2 x i32> %x, <8 x i8> %y) {
|
|
|
|
|
; SSE-LABEL: PR39893:
|
|
|
|
|
; SSE: # %bb.0:
|
|
|
|
|
; SSE-NEXT: pxor %xmm2, %xmm2
|
|
|
|
|
; SSE-NEXT: psubd %xmm0, %xmm2
|
2019-02-24 03:51:32 +08:00
|
|
|
; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
|
|
|
|
|
; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1],xmm1[2,3]
|
|
|
|
|
; SSE-NEXT: movaps %xmm2, %xmm0
|
2018-12-06 01:10:30 +08:00
|
|
|
; SSE-NEXT: retq
|
|
|
|
|
;
|
2019-03-28 03:54:41 +08:00
|
|
|
; AVX1-LABEL: PR39893:
|
|
|
|
|
; AVX1: # %bb.0:
|
|
|
|
|
; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
|
|
|
|
|
; AVX1-NEXT: vpsubd %xmm0, %xmm2, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2],zero,xmm0[3],zero,xmm0[2],zero,xmm0[3],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero
|
|
|
|
|
; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
|
|
|
|
|
; AVX1-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX2-LABEL: PR39893:
|
|
|
|
|
; AVX2: # %bb.0:
|
|
|
|
|
; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
|
|
|
|
|
; AVX2-NEXT: vpsubd %xmm0, %xmm2, %xmm0
|
|
|
|
|
; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2],zero,xmm0[3],zero,xmm0[2],zero,xmm0[3],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero
|
|
|
|
|
; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
|
|
|
|
|
; AVX2-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX512-LABEL: PR39893:
|
|
|
|
|
; AVX512: # %bb.0:
|
|
|
|
|
; AVX512-NEXT: vpxor %xmm2, %xmm2, %xmm2
|
|
|
|
|
; AVX512-NEXT: vpsubd %xmm0, %xmm2, %xmm0
|
|
|
|
|
; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2],zero,xmm0[3],zero,xmm0[2],zero,xmm0[3],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero
|
|
|
|
|
; AVX512-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
|
|
|
|
|
; AVX512-NEXT: retq
|
2018-12-06 01:10:30 +08:00
|
|
|
%sub = sub <2 x i32> <i32 0, i32 undef>, %x
|
|
|
|
|
%bc = bitcast <2 x i32> %sub to <8 x i8>
|
|
|
|
|
%shuffle = shufflevector <8 x i8> %y, <8 x i8> %bc, <2 x i32> <i32 10, i32 4>
|
|
|
|
|
ret <2 x i8> %shuffle
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
define <2 x i8> @PR39893_2(<2 x float> %x) {
|
|
|
|
|
; SSE-LABEL: PR39893_2:
|
|
|
|
|
; SSE: # %bb.0:
|
|
|
|
|
; SSE-NEXT: xorps %xmm1, %xmm1
|
|
|
|
|
; SSE-NEXT: subps %xmm0, %xmm1
|
|
|
|
|
; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
|
2019-02-22 09:49:53 +08:00
|
|
|
; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
|
|
|
|
|
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,1,1,3]
|
2018-12-06 01:10:30 +08:00
|
|
|
; SSE-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX-LABEL: PR39893_2:
|
|
|
|
|
; AVX: # %bb.0:
|
|
|
|
|
; AVX-NEXT: vxorps %xmm1, %xmm1, %xmm1
|
|
|
|
|
; AVX-NEXT: vsubps %xmm0, %xmm1, %xmm0
|
2019-02-22 09:49:53 +08:00
|
|
|
; AVX-NEXT: vpmovzxbq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero,zero,zero,zero
|
2018-12-06 01:10:30 +08:00
|
|
|
; AVX-NEXT: retq
|
|
|
|
|
%fsub = fsub <2 x float> zeroinitializer, %x
|
|
|
|
|
%bc = bitcast <2 x float> %fsub to <8 x i8>
|
|
|
|
|
%shuffle = shufflevector <8 x i8> %bc, <8 x i8> undef, <2 x i32> <i32 0, i32 1>
|
|
|
|
|
ret <2 x i8> %shuffle
|
|
|
|
|
}
|
|
|
|
|
|
2019-05-13 21:32:16 +08:00
|
|
|
define <4 x double> @fmul_v2f64(<2 x double> %x, <2 x double> %y) {
|
|
|
|
|
; SSE-LABEL: fmul_v2f64:
|
|
|
|
|
; SSE: # %bb.0:
|
|
|
|
|
; SSE-NEXT: movapd %xmm1, %xmm2
|
|
|
|
|
; SSE-NEXT: unpcklpd {{.*#+}} xmm2 = xmm2[0],xmm0[0]
|
|
|
|
|
; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
|
|
|
|
|
; SSE-NEXT: mulpd %xmm0, %xmm0
|
|
|
|
|
; SSE-NEXT: mulpd %xmm2, %xmm2
|
|
|
|
|
; SSE-NEXT: addpd %xmm0, %xmm2
|
|
|
|
|
; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
|
|
|
; SSE-NEXT: movapd %xmm2, %xmm0
|
|
|
|
|
; SSE-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX1-LABEL: fmul_v2f64:
|
|
|
|
|
; AVX1: # %bb.0:
|
2019-05-13 22:31:14 +08:00
|
|
|
; AVX1-NEXT: vunpcklpd {{.*#+}} xmm2 = xmm1[0],xmm0[0]
|
|
|
|
|
; AVX1-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
|
|
|
|
|
; AVX1-NEXT: vmulpd %xmm0, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vmulpd %xmm2, %xmm2, %xmm1
|
|
|
|
|
; AVX1-NEXT: vaddpd %xmm0, %xmm1, %xmm0
|
2019-05-13 21:32:16 +08:00
|
|
|
; AVX1-NEXT: vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
|
|
|
|
|
; AVX1-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX2-LABEL: fmul_v2f64:
|
|
|
|
|
; AVX2: # %bb.0:
|
2019-05-13 22:31:14 +08:00
|
|
|
; AVX2-NEXT: vunpcklpd {{.*#+}} xmm2 = xmm1[0],xmm0[0]
|
|
|
|
|
; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
|
|
|
|
|
; AVX2-NEXT: vmulpd %xmm0, %xmm0, %xmm0
|
|
|
|
|
; AVX2-NEXT: vmulpd %xmm2, %xmm2, %xmm1
|
|
|
|
|
; AVX2-NEXT: vaddpd %xmm0, %xmm1, %xmm0
|
2019-05-13 21:32:16 +08:00
|
|
|
; AVX2-NEXT: vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
|
|
|
|
|
; AVX2-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX512-LABEL: fmul_v2f64:
|
|
|
|
|
; AVX512: # %bb.0:
|
|
|
|
|
; AVX512-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm0[1],xmm1[1]
|
|
|
|
|
; AVX512-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm1[0],xmm0[0]
|
2019-05-13 22:31:14 +08:00
|
|
|
; AVX512-NEXT: vmulpd %xmm0, %xmm0, %xmm0
|
|
|
|
|
; AVX512-NEXT: vfmadd231pd {{.*#+}} xmm0 = (xmm2 * xmm2) + xmm0
|
2019-05-13 21:32:16 +08:00
|
|
|
; AVX512-NEXT: vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
|
|
|
|
|
; AVX512-NEXT: retq
|
|
|
|
|
%s = shufflevector <2 x double> %x, <2 x double> %y, <4 x i32> <i32 2, i32 0, i32 1, i32 3>
|
|
|
|
|
%bo = fmul fast <4 x double> %s, %s
|
|
|
|
|
%ext = shufflevector <4 x double> %bo, <4 x double> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
|
|
|
|
|
%add = fadd fast <4 x double> %bo, %ext
|
|
|
|
|
%rdx = shufflevector <4 x double> %add, <4 x double> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
|
|
|
|
|
ret <4 x double> %rdx
|
|
|
|
|
}
|
|
|
|
|
|