2016-07-19 21:35:11 +08:00
|
|
|
; RUN: llc < %s -mtriple=arm64-eabi -aarch64-neon-syntax=apple -asm-verbose=false | FileCheck %s
|
2014-03-29 18:18:08 +08:00
|
|
|
|
|
|
|
define i32 @vmin_u8x8(<8 x i8> %a) nounwind ssp {
|
|
|
|
; CHECK-LABEL: vmin_u8x8:
|
|
|
|
; CHECK: uminv.8b b[[REG:[0-9]+]], v0
|
|
|
|
; CHECK: fmov [[REG2:w[0-9]+]], s[[REG]]
|
|
|
|
; CHECK-NOT: and
|
|
|
|
; CHECK: cbz [[REG2]],
|
|
|
|
entry:
|
2014-05-24 20:50:23 +08:00
|
|
|
%vminv.i = tail call i32 @llvm.aarch64.neon.uminv.i32.v8i8(<8 x i8> %a) nounwind
|
2014-03-29 18:18:08 +08:00
|
|
|
%tmp = trunc i32 %vminv.i to i8
|
|
|
|
%tobool = icmp eq i8 %tmp, 0
|
|
|
|
br i1 %tobool, label %return, label %if.then
|
|
|
|
|
|
|
|
if.then:
|
|
|
|
%call1 = tail call i32 bitcast (i32 (...)* @bar to i32 ()*)() nounwind
|
|
|
|
br label %return
|
|
|
|
|
|
|
|
return:
|
|
|
|
%retval.0 = phi i32 [ %call1, %if.then ], [ 0, %entry ]
|
|
|
|
ret i32 %retval.0
|
|
|
|
}
|
|
|
|
|
|
|
|
declare i32 @bar(...)
|
|
|
|
|
|
|
|
define i32 @vmin_u4x16(<4 x i16> %a) nounwind ssp {
|
|
|
|
; CHECK-LABEL: vmin_u4x16:
|
|
|
|
; CHECK: uminv.4h h[[REG:[0-9]+]], v0
|
|
|
|
; CHECK: fmov [[REG2:w[0-9]+]], s[[REG]]
|
|
|
|
; CHECK-NOT: and
|
|
|
|
; CHECK: cbz [[REG2]],
|
|
|
|
entry:
|
2014-05-24 20:50:23 +08:00
|
|
|
%vminv.i = tail call i32 @llvm.aarch64.neon.uminv.i32.v4i16(<4 x i16> %a) nounwind
|
2014-03-29 18:18:08 +08:00
|
|
|
%tmp = trunc i32 %vminv.i to i16
|
|
|
|
%tobool = icmp eq i16 %tmp, 0
|
|
|
|
br i1 %tobool, label %return, label %if.then
|
|
|
|
|
|
|
|
if.then:
|
|
|
|
%call1 = tail call i32 bitcast (i32 (...)* @bar to i32 ()*)() nounwind
|
|
|
|
br label %return
|
|
|
|
|
|
|
|
return:
|
|
|
|
%retval.0 = phi i32 [ %call1, %if.then ], [ 0, %entry ]
|
|
|
|
ret i32 %retval.0
|
|
|
|
}
|
|
|
|
|
|
|
|
define i32 @vmin_u8x16(<8 x i16> %a) nounwind ssp {
|
|
|
|
; CHECK-LABEL: vmin_u8x16:
|
|
|
|
; CHECK: uminv.8h h[[REG:[0-9]+]], v0
|
|
|
|
; CHECK: fmov [[REG2:w[0-9]+]], s[[REG]]
|
|
|
|
; CHECK-NOT: and
|
|
|
|
; CHECK: cbz [[REG2]],
|
|
|
|
entry:
|
2014-05-24 20:50:23 +08:00
|
|
|
%vminv.i = tail call i32 @llvm.aarch64.neon.uminv.i32.v8i16(<8 x i16> %a) nounwind
|
2014-03-29 18:18:08 +08:00
|
|
|
%tmp = trunc i32 %vminv.i to i16
|
|
|
|
%tobool = icmp eq i16 %tmp, 0
|
|
|
|
br i1 %tobool, label %return, label %if.then
|
|
|
|
|
|
|
|
if.then:
|
|
|
|
%call1 = tail call i32 bitcast (i32 (...)* @bar to i32 ()*)() nounwind
|
|
|
|
br label %return
|
|
|
|
|
|
|
|
return:
|
|
|
|
%retval.0 = phi i32 [ %call1, %if.then ], [ 0, %entry ]
|
|
|
|
ret i32 %retval.0
|
|
|
|
}
|
|
|
|
|
|
|
|
define i32 @vmin_u16x8(<16 x i8> %a) nounwind ssp {
|
|
|
|
; CHECK-LABEL: vmin_u16x8:
|
|
|
|
; CHECK: uminv.16b b[[REG:[0-9]+]], v0
|
|
|
|
; CHECK: fmov [[REG2:w[0-9]+]], s[[REG]]
|
|
|
|
; CHECK-NOT: and
|
|
|
|
; CHECK: cbz [[REG2]],
|
|
|
|
entry:
|
2014-05-24 20:50:23 +08:00
|
|
|
%vminv.i = tail call i32 @llvm.aarch64.neon.uminv.i32.v16i8(<16 x i8> %a) nounwind
|
2014-03-29 18:18:08 +08:00
|
|
|
%tmp = trunc i32 %vminv.i to i8
|
|
|
|
%tobool = icmp eq i8 %tmp, 0
|
|
|
|
br i1 %tobool, label %return, label %if.then
|
|
|
|
|
|
|
|
if.then:
|
|
|
|
%call1 = tail call i32 bitcast (i32 (...)* @bar to i32 ()*)() nounwind
|
|
|
|
br label %return
|
|
|
|
|
|
|
|
return:
|
|
|
|
%retval.0 = phi i32 [ %call1, %if.then ], [ 0, %entry ]
|
|
|
|
ret i32 %retval.0
|
|
|
|
}
|
|
|
|
|
[AArch64] Avoid going through GPRs for across-vector instructions.
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
(v4i32 (uaddv ...))
is the same as
(v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
(v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
(i32 (int_aarch64_neon_uaddv ...)), ssub)
In a combine, we transform all such across-vector-lanes intrinsics to:
(i32 (extract_vector_elt (uaddv ...), 0))
This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs. Consider:
uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
return vmulq_n_u32(a, vaddvq_u32(b));
}
We now generate:
addv.4s s1, v1
mul.4s v0, v0, v1[0]
instead of the previous:
addv.4s s1, v1
fmov w8, s1
dup.4s v1, w8
mul.4s v0, v1, v0
rdar://20044838
llvm-svn: 231840
2015-03-11 04:45:38 +08:00
|
|
|
define <8 x i8> @test_vminv_u8_used_by_laneop(<8 x i8> %a1, <8 x i8> %a2) {
|
|
|
|
; CHECK-LABEL: test_vminv_u8_used_by_laneop:
|
|
|
|
; CHECK: uminv.8b b[[REGNUM:[0-9]+]], v1
|
2017-02-09 05:41:16 +08:00
|
|
|
; CHECK-NEXT: ins.b v0[3], v[[REGNUM]][0]
|
[AArch64] Avoid going through GPRs for across-vector instructions.
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
(v4i32 (uaddv ...))
is the same as
(v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
(v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
(i32 (int_aarch64_neon_uaddv ...)), ssub)
In a combine, we transform all such across-vector-lanes intrinsics to:
(i32 (extract_vector_elt (uaddv ...), 0))
This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs. Consider:
uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
return vmulq_n_u32(a, vaddvq_u32(b));
}
We now generate:
addv.4s s1, v1
mul.4s v0, v0, v1[0]
instead of the previous:
addv.4s s1, v1
fmov w8, s1
dup.4s v1, w8
mul.4s v0, v1, v0
rdar://20044838
llvm-svn: 231840
2015-03-11 04:45:38 +08:00
|
|
|
; CHECK-NEXT: ret
|
|
|
|
entry:
|
|
|
|
%0 = tail call i32 @llvm.aarch64.neon.uminv.i32.v8i8(<8 x i8> %a2)
|
|
|
|
%1 = trunc i32 %0 to i8
|
|
|
|
%2 = insertelement <8 x i8> %a1, i8 %1, i32 3
|
|
|
|
ret <8 x i8> %2
|
|
|
|
}
|
|
|
|
|
|
|
|
define <4 x i16> @test_vminv_u16_used_by_laneop(<4 x i16> %a1, <4 x i16> %a2) {
|
|
|
|
; CHECK-LABEL: test_vminv_u16_used_by_laneop:
|
|
|
|
; CHECK: uminv.4h h[[REGNUM:[0-9]+]], v1
|
2017-02-09 05:41:16 +08:00
|
|
|
; CHECK-NEXT: ins.h v0[3], v[[REGNUM]][0]
|
[AArch64] Avoid going through GPRs for across-vector instructions.
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
(v4i32 (uaddv ...))
is the same as
(v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
(v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
(i32 (int_aarch64_neon_uaddv ...)), ssub)
In a combine, we transform all such across-vector-lanes intrinsics to:
(i32 (extract_vector_elt (uaddv ...), 0))
This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs. Consider:
uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
return vmulq_n_u32(a, vaddvq_u32(b));
}
We now generate:
addv.4s s1, v1
mul.4s v0, v0, v1[0]
instead of the previous:
addv.4s s1, v1
fmov w8, s1
dup.4s v1, w8
mul.4s v0, v1, v0
rdar://20044838
llvm-svn: 231840
2015-03-11 04:45:38 +08:00
|
|
|
; CHECK-NEXT: ret
|
|
|
|
entry:
|
|
|
|
%0 = tail call i32 @llvm.aarch64.neon.uminv.i32.v4i16(<4 x i16> %a2)
|
|
|
|
%1 = trunc i32 %0 to i16
|
|
|
|
%2 = insertelement <4 x i16> %a1, i16 %1, i32 3
|
|
|
|
ret <4 x i16> %2
|
|
|
|
}
|
|
|
|
|
|
|
|
define <2 x i32> @test_vminv_u32_used_by_laneop(<2 x i32> %a1, <2 x i32> %a2) {
|
|
|
|
; CHECK-LABEL: test_vminv_u32_used_by_laneop:
|
|
|
|
; CHECK: uminp.2s v[[REGNUM:[0-9]+]], v1, v1
|
2017-02-09 05:41:16 +08:00
|
|
|
; CHECK-NEXT: ins.s v0[1], v[[REGNUM]][0]
|
[AArch64] Avoid going through GPRs for across-vector instructions.
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
(v4i32 (uaddv ...))
is the same as
(v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
(v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
(i32 (int_aarch64_neon_uaddv ...)), ssub)
In a combine, we transform all such across-vector-lanes intrinsics to:
(i32 (extract_vector_elt (uaddv ...), 0))
This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs. Consider:
uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
return vmulq_n_u32(a, vaddvq_u32(b));
}
We now generate:
addv.4s s1, v1
mul.4s v0, v0, v1[0]
instead of the previous:
addv.4s s1, v1
fmov w8, s1
dup.4s v1, w8
mul.4s v0, v1, v0
rdar://20044838
llvm-svn: 231840
2015-03-11 04:45:38 +08:00
|
|
|
; CHECK-NEXT: ret
|
|
|
|
entry:
|
|
|
|
%0 = tail call i32 @llvm.aarch64.neon.uminv.i32.v2i32(<2 x i32> %a2)
|
|
|
|
%1 = insertelement <2 x i32> %a1, i32 %0, i32 1
|
|
|
|
ret <2 x i32> %1
|
|
|
|
}
|
|
|
|
|
|
|
|
define <16 x i8> @test_vminvq_u8_used_by_laneop(<16 x i8> %a1, <16 x i8> %a2) {
|
|
|
|
; CHECK-LABEL: test_vminvq_u8_used_by_laneop:
|
|
|
|
; CHECK: uminv.16b b[[REGNUM:[0-9]+]], v1
|
2017-02-09 05:41:16 +08:00
|
|
|
; CHECK-NEXT: ins.b v0[3], v[[REGNUM]][0]
|
[AArch64] Avoid going through GPRs for across-vector instructions.
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
(v4i32 (uaddv ...))
is the same as
(v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
(v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
(i32 (int_aarch64_neon_uaddv ...)), ssub)
In a combine, we transform all such across-vector-lanes intrinsics to:
(i32 (extract_vector_elt (uaddv ...), 0))
This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs. Consider:
uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
return vmulq_n_u32(a, vaddvq_u32(b));
}
We now generate:
addv.4s s1, v1
mul.4s v0, v0, v1[0]
instead of the previous:
addv.4s s1, v1
fmov w8, s1
dup.4s v1, w8
mul.4s v0, v1, v0
rdar://20044838
llvm-svn: 231840
2015-03-11 04:45:38 +08:00
|
|
|
; CHECK-NEXT: ret
|
|
|
|
entry:
|
|
|
|
%0 = tail call i32 @llvm.aarch64.neon.uminv.i32.v16i8(<16 x i8> %a2)
|
|
|
|
%1 = trunc i32 %0 to i8
|
|
|
|
%2 = insertelement <16 x i8> %a1, i8 %1, i32 3
|
|
|
|
ret <16 x i8> %2
|
|
|
|
}
|
|
|
|
|
|
|
|
define <8 x i16> @test_vminvq_u16_used_by_laneop(<8 x i16> %a1, <8 x i16> %a2) {
|
|
|
|
; CHECK-LABEL: test_vminvq_u16_used_by_laneop:
|
|
|
|
; CHECK: uminv.8h h[[REGNUM:[0-9]+]], v1
|
2017-02-09 05:41:16 +08:00
|
|
|
; CHECK-NEXT: ins.h v0[3], v[[REGNUM]][0]
|
[AArch64] Avoid going through GPRs for across-vector instructions.
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
(v4i32 (uaddv ...))
is the same as
(v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
(v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
(i32 (int_aarch64_neon_uaddv ...)), ssub)
In a combine, we transform all such across-vector-lanes intrinsics to:
(i32 (extract_vector_elt (uaddv ...), 0))
This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs. Consider:
uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
return vmulq_n_u32(a, vaddvq_u32(b));
}
We now generate:
addv.4s s1, v1
mul.4s v0, v0, v1[0]
instead of the previous:
addv.4s s1, v1
fmov w8, s1
dup.4s v1, w8
mul.4s v0, v1, v0
rdar://20044838
llvm-svn: 231840
2015-03-11 04:45:38 +08:00
|
|
|
; CHECK-NEXT: ret
|
|
|
|
entry:
|
|
|
|
%0 = tail call i32 @llvm.aarch64.neon.uminv.i32.v8i16(<8 x i16> %a2)
|
|
|
|
%1 = trunc i32 %0 to i16
|
|
|
|
%2 = insertelement <8 x i16> %a1, i16 %1, i32 3
|
|
|
|
ret <8 x i16> %2
|
|
|
|
}
|
|
|
|
|
|
|
|
define <4 x i32> @test_vminvq_u32_used_by_laneop(<4 x i32> %a1, <4 x i32> %a2) {
|
|
|
|
; CHECK-LABEL: test_vminvq_u32_used_by_laneop:
|
|
|
|
; CHECK: uminv.4s s[[REGNUM:[0-9]+]], v1
|
2017-02-09 05:41:16 +08:00
|
|
|
; CHECK-NEXT: ins.s v0[3], v[[REGNUM]][0]
|
[AArch64] Avoid going through GPRs for across-vector instructions.
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
(v4i32 (uaddv ...))
is the same as
(v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
(v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
(i32 (int_aarch64_neon_uaddv ...)), ssub)
In a combine, we transform all such across-vector-lanes intrinsics to:
(i32 (extract_vector_elt (uaddv ...), 0))
This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs. Consider:
uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
return vmulq_n_u32(a, vaddvq_u32(b));
}
We now generate:
addv.4s s1, v1
mul.4s v0, v0, v1[0]
instead of the previous:
addv.4s s1, v1
fmov w8, s1
dup.4s v1, w8
mul.4s v0, v1, v0
rdar://20044838
llvm-svn: 231840
2015-03-11 04:45:38 +08:00
|
|
|
; CHECK-NEXT: ret
|
|
|
|
entry:
|
|
|
|
%0 = tail call i32 @llvm.aarch64.neon.uminv.i32.v4i32(<4 x i32> %a2)
|
|
|
|
%1 = insertelement <4 x i32> %a1, i32 %0, i32 3
|
|
|
|
ret <4 x i32> %1
|
|
|
|
}
|
2014-05-24 20:50:23 +08:00
|
|
|
declare i32 @llvm.aarch64.neon.uminv.i32.v16i8(<16 x i8>) nounwind readnone
|
|
|
|
declare i32 @llvm.aarch64.neon.uminv.i32.v8i16(<8 x i16>) nounwind readnone
|
|
|
|
declare i32 @llvm.aarch64.neon.uminv.i32.v4i16(<4 x i16>) nounwind readnone
|
|
|
|
declare i32 @llvm.aarch64.neon.uminv.i32.v8i8(<8 x i8>) nounwind readnone
|
[AArch64] Avoid going through GPRs for across-vector instructions.
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
(v4i32 (uaddv ...))
is the same as
(v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
(v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
(i32 (int_aarch64_neon_uaddv ...)), ssub)
In a combine, we transform all such across-vector-lanes intrinsics to:
(i32 (extract_vector_elt (uaddv ...), 0))
This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs. Consider:
uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
return vmulq_n_u32(a, vaddvq_u32(b));
}
We now generate:
addv.4s s1, v1
mul.4s v0, v0, v1[0]
instead of the previous:
addv.4s s1, v1
fmov w8, s1
dup.4s v1, w8
mul.4s v0, v1, v0
rdar://20044838
llvm-svn: 231840
2015-03-11 04:45:38 +08:00
|
|
|
declare i32 @llvm.aarch64.neon.uminv.i32.v2i32(<2 x i32>) nounwind readnone
|
|
|
|
declare i32 @llvm.aarch64.neon.uminv.i32.v4i32(<4 x i32>) nounwind readnone
|