maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/X86/vector-sext.ll

5011 lines
193 KiB
LLVM
Raw Normal View History

Make utils/update_llc_test_checks.py note that the assertions are autogenerated. Also update existing test cases which appear to be generated by it and weren't modified (other than addition of the header) by rerunning it. llvm-svn: 253917
2015-11-24 05:33:58 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
[X86][SSE] vector sext/zext tests - remove unnecessary mcpu arguments llvm-svn: 251233
2015-10-25 20:15:00 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE --check-prefix=SSE2
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+ssse3 | FileCheck %s --check-prefix=SSE --check-prefix=SSSE3
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE --check-prefix=SSE41
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512F
[X86][AVX512BW] Add sext/zext AVX512BW 512-bit vector tests llvm-svn: 277957
2016-08-07 20:41:36 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512BW
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
; Just one 32-bit run to make sure we do reasonable things there.
[X86][SSE] vector sext/zext tests - remove unnecessary mcpu arguments llvm-svn: 251233
2015-10-25 20:15:00 +08:00
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=X32-SSE41
Optimization for SIGN_EXTEND operation on AVX. Special handling was added for v4i32 -> v4i64 and v8i16 -> v8i32 extensions. llvm-svn: 149600
2012-02-02 17:10:43 +08:00
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
define <8 x i16> @sext_16i8_to_8i16(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i8_to_8i16:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: psraw $8, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i8_to_8i16:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: psraw $8, %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i8_to_8i16:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: sext_16i8_to_8i16:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxbw %xmm0, %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: sext_16i8_to_8i16:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbw %xmm0, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <16 x i8> %A, <16 x i8> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%C = sext <8 x i8> %B to <8 x i16>
ret <8 x i16> %C
}
define <16 x i16> @sext_16i8_to_16i16(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i8_to_16i16:
; SSE2: # BB#0: # %entry
; SSE2-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]
; SSE2-NEXT: psraw $8, %xmm2
[X86][SSE] Combine UNPCKL with vector_shuffle into UNPCKH to save one instruction for sext from v16i8 to v16i16 and v8i16 to v8i32. This patch is enabling combining UNPCKL with vector_shuffle that moves the upper half of a vector into the lower half, into a UNPCKH instruction. For example: t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1, undef:v16i8 t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2 will be combined to: t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1 Differential revision: http://reviews.llvm.org/D14399 llvm-svn: 253067
2015-11-14 03:47:43 +08:00
; SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSE2-NEXT: psraw $8, %xmm1
; SSE2-NEXT: movdqa %xmm2, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i8_to_16i16:
; SSSE3: # BB#0: # %entry
; SSSE3-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]
; SSSE3-NEXT: psraw $8, %xmm2
[X86][SSE] Combine UNPCKL with vector_shuffle into UNPCKH to save one instruction for sext from v16i8 to v16i16 and v8i16 to v8i32. This patch is enabling combining UNPCKL with vector_shuffle that moves the upper half of a vector into the lower half, into a UNPCKH instruction. For example: t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1, undef:v16i8 t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2 will be combined to: t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1 Differential revision: http://reviews.llvm.org/D14399 llvm-svn: 253067
2015-11-14 03:47:43 +08:00
; SSSE3-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSSE3-NEXT: psraw $8, %xmm1
; SSSE3-NEXT: movdqa %xmm2, %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i8_to_16i16:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbw %xmm0, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxbw %xmm0, %xmm1
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_16i8_to_16i16:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbw %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxbw %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_16i8_to_16i16:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxbw %xmm0, %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_16i8_to_16i16:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxbw %xmm0, %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; X32-SSE41-LABEL: sext_16i8_to_16i16:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbw %xmm0, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxbw %xmm0, %xmm1
; X32-SSE41-NEXT: movdqa %xmm2, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = sext <16 x i8> %A to <16 x i16>
ret <16 x i16> %B
}
[X86][AVX512] Add sext/zext to 512-bit vector tests llvm-svn: 277956
2016-08-07 20:10:46 +08:00
define <32 x i16> @sext_32i8_to_32i16(<32 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_32i8_to_32i16:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
; SSE2-NEXT: psraw $8, %xmm4
; SSE2-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
; SSE2-NEXT: psraw $8, %xmm5
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSE2-NEXT: psraw $8, %xmm2
; SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
; SSE2-NEXT: psraw $8, %xmm3
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: movdqa %xmm5, %xmm1
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_32i8_to_32i16:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
; SSSE3-NEXT: psraw $8, %xmm4
; SSSE3-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
; SSSE3-NEXT: psraw $8, %xmm5
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSSE3-NEXT: psraw $8, %xmm2
; SSSE3-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
; SSSE3-NEXT: psraw $8, %xmm3
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: movdqa %xmm5, %xmm1
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_32i8_to_32i16:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbw %xmm0, %xmm5
; SSE41-NEXT: pmovsxbw %xmm1, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxbw %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
; SSE41-NEXT: pmovsxbw %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm5, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_32i8_to_32i16:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbw %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxbw %xmm2, %xmm2
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm2
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
; AVX1-NEXT: vpmovsxbw %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxbw %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm1
; AVX1-NEXT: vmovaps %ymm2, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_32i8_to_32i16:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxbw %xmm0, %ymm2
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm0
; AVX2-NEXT: vpmovsxbw %xmm0, %ymm1
; AVX2-NEXT: vmovdqa %ymm2, %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512BW] Add sext/zext AVX512BW 512-bit vector tests llvm-svn: 277957
2016-08-07 20:41:36 +08:00
; AVX512F-LABEL: sext_32i8_to_32i16:
; AVX512F: # BB#0: # %entry
; AVX512F-NEXT: vpmovsxbw %xmm0, %ymm2
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm0
; AVX512F-NEXT: vpmovsxbw %xmm0, %ymm1
[AVX-512] Simplify X86InstrInfo::copyPhysReg for 128/256-bit vectors with AVX512, but not VLX. We should use the VEX opcodes and trust the register allocator to not use the extended XMM/YMM register space. Previously we were extending to copying the whole ZMM register. The register allocator shouldn't use XMM16-31 or YMM16-31 in this configuration as the instructions to spill them aren't available. llvm-svn: 280648
2016-09-05 14:43:06 +08:00
; AVX512F-NEXT: vmovdqa %ymm2, %ymm0
[X86][AVX512BW] Add sext/zext AVX512BW 512-bit vector tests llvm-svn: 277957
2016-08-07 20:41:36 +08:00
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: sext_32i8_to_32i16:
; AVX512BW: # BB#0: # %entry
; AVX512BW-NEXT: vpmovsxbw %ymm0, %zmm0
; AVX512BW-NEXT: retq
[X86][AVX512] Add sext/zext to 512-bit vector tests llvm-svn: 277956
2016-08-07 20:10:46 +08:00
;
; X32-SSE41-LABEL: sext_32i8_to_32i16:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbw %xmm0, %xmm5
; X32-SSE41-NEXT: pmovsxbw %xmm1, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxbw %xmm0, %xmm4
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
; X32-SSE41-NEXT: pmovsxbw %xmm0, %xmm3
; X32-SSE41-NEXT: movdqa %xmm5, %xmm0
; X32-SSE41-NEXT: movdqa %xmm4, %xmm1
; X32-SSE41-NEXT: retl
entry:
%B = sext <32 x i8> %A to <32 x i16>
ret <32 x i16> %B
}
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
define <4 x i32> @sext_16i8_to_4i32(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i8_to_4i32:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: psrad $24, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i8_to_4i32:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: psrad $24, %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i8_to_4i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbd %xmm0, %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: sext_16i8_to_4i32:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxbd %xmm0, %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: sext_16i8_to_4i32:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbd %xmm0, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <16 x i8> %A, <16 x i8> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%C = sext <4 x i8> %B to <4 x i32>
ret <4 x i32> %C
}
define <8 x i32> @sext_16i8_to_8i32(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i8_to_8i32:
; SSE2: # BB#0: # %entry
; SSE2-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]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE2-NEXT: psrad $24, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; SSE2-NEXT: psrad $24, %xmm1
; SSE2-NEXT: movdqa %xmm2, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i8_to_8i32:
; SSSE3: # BB#0: # %entry
[X86][SSE] Resolve target shuffle inputs to sentinels to permit more combines The combineX86ShufflesRecursively only supports unary shuffles, but was missing the opportunity to combine binary shuffles with a zero / undef second input. This patch resolves target shuffle inputs, converting the shuffle mask elements to SM_SentinelUndef/SM_SentinelZero where possible. It then resolves the updated mask to check if we have created a faux unary shuffle. Additionally, we now attempt to recursively call combineX86ShufflesRecursively for all input operands (we used to just recurse for unary integer shuffles and unary unpacks) - it safely returns early if its not a target shuffle. Differential Revision: http://reviews.llvm.org/D16683 llvm-svn: 260063
2016-02-08 06:51:06 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: psrad $24, %xmm0
; SSSE3-NEXT: pshufb {{.*#+}} xmm1 = xmm1[u,u,u,4,u,u,u,5,u,u,u,6,u,u,u,7]
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSSE3-NEXT: psrad $24, %xmm1
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i8_to_8i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbd %xmm0, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; SSE41-NEXT: pmovsxbd %xmm0, %xmm1
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_16i8_to_8i32:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbd %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; AVX1-NEXT: vpmovsxbd %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_16i8_to_8i32:
; AVX2: # BB#0: # %entry
[X86][AVX2] Fix SIGN_EXTEND vector handling on AVX2 targets. On AVX2 target we are poorly legalizing SIGN_EXTEND ops for which the input's legalized type doesn't have the same number of elements as the destination, resulting in an ANY_EXTEND followed by a SIGN_EXTEND_INREG. This patch uses the existing SIGN_EXTEND -> SIGN_EXTEND_VECTOR_INREG combine to extend the input to the size of the result and using SIGN_EXTEND_VECTOR_INREG instead. Differential Revision: http://reviews.llvm.org/D16994 llvm-svn: 260210
2016-02-09 16:19:19 +08:00
; AVX2-NEXT: vpmovsxbd %xmm0, %ymm0
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_16i8_to_8i32:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxbd %xmm0, %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; X32-SSE41-LABEL: sext_16i8_to_8i32:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbd %xmm0, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; X32-SSE41-NEXT: pmovsxbd %xmm0, %xmm1
; X32-SSE41-NEXT: movdqa %xmm2, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <16 x i8> %A, <16 x i8> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%C = sext <8 x i8> %B to <8 x i32>
ret <8 x i32> %C
}
[X86][AVX512] Add sext/zext to 512-bit vector tests llvm-svn: 277956
2016-08-07 20:10:46 +08:00
define <16 x i32> @sext_16i8_to_16i32(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i8_to_16i32:
; SSE2: # BB#0: # %entry
; SSE2-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]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; SSE2-NEXT: psrad $24, %xmm4
; SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE2-NEXT: psrad $24, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
; SSE2-NEXT: psrad $24, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
; SSE2-NEXT: psrad $24, %xmm3
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i8_to_16i32:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: psrad $24, %xmm0
; SSSE3-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm3[8],xmm1[9],xmm3[9],xmm1[10],xmm3[10],xmm1[11],xmm3[11],xmm1[12],xmm3[12],xmm1[13],xmm3[13],xmm1[14],xmm3[14],xmm1[15],xmm3[15]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSSE3-NEXT: psrad $24, %xmm2
; SSSE3-NEXT: movdqa %xmm3, %xmm1
; SSSE3-NEXT: pshufb {{.*#+}} xmm1 = xmm1[u,u,u,4,u,u,u,5,u,u,u,6,u,u,u,7]
; SSSE3-NEXT: psrad $24, %xmm1
; SSSE3-NEXT: pshufb {{.*#+}} xmm3 = xmm3[u,u,u,12,u,u,u,13,u,u,u,14,u,u,u,15]
; SSSE3-NEXT: psrad $24, %xmm3
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i8_to_16i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbd %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
; SSE41-NEXT: pmovsxbd %xmm1, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxbd %xmm2, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; SSE41-NEXT: pmovsxbd %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm4, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_16i8_to_16i32:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbd %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
; AVX1-NEXT: vpmovsxbd %xmm2, %xmm2
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm2
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxbd %xmm1, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; AVX1-NEXT: vpmovsxbd %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm1
; AVX1-NEXT: vmovaps %ymm2, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_16i8_to_16i32:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxbd %xmm0, %ymm2
; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX2-NEXT: vpmovsxbd %xmm0, %ymm1
; AVX2-NEXT: vmovdqa %ymm2, %ymm0
; AVX2-NEXT: retq
;
; AVX512-LABEL: sext_16i8_to_16i32:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxbd %xmm0, %zmm0
; AVX512-NEXT: retq
;
; X32-SSE41-LABEL: sext_16i8_to_16i32:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbd %xmm0, %xmm4
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
; X32-SSE41-NEXT: pmovsxbd %xmm1, %xmm1
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxbd %xmm2, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; X32-SSE41-NEXT: pmovsxbd %xmm0, %xmm3
; X32-SSE41-NEXT: movdqa %xmm4, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = sext <16 x i8> %A to <16 x i32>
ret <16 x i32> %B
}
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
define <2 x i64> @sext_16i8_to_2i64(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i8_to_2i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: psrad $24, %xmm0
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i8_to_2i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: psrad $24, %xmm0
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i8_to_2i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbq %xmm0, %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: sext_16i8_to_2i64:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxbq %xmm0, %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: sext_16i8_to_2i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbq %xmm0, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <16 x i8> %A, <16 x i8> undef, <2 x i32> <i32 0, i32 1>
%C = sext <2 x i8> %B to <2 x i64>
ret <2 x i64> %C
}
define <4 x i64> @sext_16i8_to_4i64(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i8_to_4i64:
; SSE2: # BB#0: # %entry
; SSE2-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]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: psrad $24, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
; SSE2-NEXT: psrld $16, %xmm0
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: psrad $31, %xmm0
; SSE2-NEXT: psrad $24, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE2-NEXT: movdqa %xmm2, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i8_to_4i64:
; SSSE3: # BB#0: # %entry
[X86][SSE] Added support for combining bit-shifts with shuffles. Bit-shifts by a whole number of bytes can be represented as a shuffle mask suitable for combining. Added a 'getFauxShuffleMask' function to allow us to create shuffle masks from other suitable operations. llvm-svn: 288040
2016-11-29 00:25:01 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: psrad $24, %xmm0
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSSE3-NEXT: pshufb {{.*#+}} xmm1 = xmm1[u,u,u,2,u,u,u,3,u,u,u],zero,xmm1[u,u,u],zero
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSSE3-NEXT: psrad $24, %xmm1
[X86][SSE] Added support for combining bit-shifts with shuffles. Bit-shifts by a whole number of bytes can be represented as a shuffle mask suitable for combining. Added a 'getFauxShuffleMask' function to allow us to create shuffle masks from other suitable operations. llvm-svn: 288040
2016-11-29 00:25:01 +08:00
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i8_to_4i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbq %xmm0, %xmm2
; SSE41-NEXT: psrld $16, %xmm0
; SSE41-NEXT: pmovsxbq %xmm0, %xmm1
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_16i8_to_4i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbq %xmm0, %xmm1
; AVX1-NEXT: vpsrld $16, %xmm0, %xmm0
; AVX1-NEXT: vpmovsxbq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_16i8_to_4i64:
; AVX2: # BB#0: # %entry
[X86][AVX2] Fix SIGN_EXTEND vector handling on AVX2 targets. On AVX2 target we are poorly legalizing SIGN_EXTEND ops for which the input's legalized type doesn't have the same number of elements as the destination, resulting in an ANY_EXTEND followed by a SIGN_EXTEND_INREG. This patch uses the existing SIGN_EXTEND -> SIGN_EXTEND_VECTOR_INREG combine to extend the input to the size of the result and using SIGN_EXTEND_VECTOR_INREG instead. Differential Revision: http://reviews.llvm.org/D16994 llvm-svn: 260210
2016-02-09 16:19:19 +08:00
; AVX2-NEXT: vpmovsxbq %xmm0, %ymm0
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_16i8_to_4i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxbq %xmm0, %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; X32-SSE41-LABEL: sext_16i8_to_4i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbq %xmm0, %xmm2
; X32-SSE41-NEXT: psrld $16, %xmm0
; X32-SSE41-NEXT: pmovsxbq %xmm0, %xmm1
; X32-SSE41-NEXT: movdqa %xmm2, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <16 x i8> %A, <16 x i8> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%C = sext <4 x i8> %B to <4 x i64>
ret <4 x i64> %C
}
[X86][SSE] Added 16i8 -> 8i64 sext test Shows poor codegen for AVX2 llvm-svn: 266560
2016-04-17 23:10:42 +08:00
define <8 x i64> @sext_16i8_to_8i64(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i8_to_8i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: psrad $24, %xmm0
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,1,2,3]
; SSE2-NEXT: psrld $16, %xmm1
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: psrad $24, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm2, %xmm4
; SSE2-NEXT: psrad $31, %xmm4
; SSE2-NEXT: psrad $24, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]
; SSE2-NEXT: psrld $16, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm3, %xmm4
; SSE2-NEXT: psrad $31, %xmm4
; SSE2-NEXT: psrad $24, %xmm3
; SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i8_to_8i64:
; SSSE3: # BB#0: # %entry
[X86][SSE] Added support for combining bit-shifts with shuffles. Bit-shifts by a whole number of bytes can be represented as a shuffle mask suitable for combining. Added a 'getFauxShuffleMask' function to allow us to create shuffle masks from other suitable operations. llvm-svn: 288040
2016-11-29 00:25:01 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = <u,u,u,2,u,u,u,3,u,u,u,255,u,u,u,255>
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,2,3]
[X86][SSE] Added 16i8 -> 8i64 sext test Shows poor codegen for AVX2 llvm-svn: 266560
2016-04-17 23:10:42 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm1
[X86][SSE] Added support for combining bit-shifts with shuffles. Bit-shifts by a whole number of bytes can be represented as a shuffle mask suitable for combining. Added a 'getFauxShuffleMask' function to allow us to create shuffle masks from other suitable operations. llvm-svn: 288040
2016-11-29 00:25:01 +08:00
; SSSE3-NEXT: pshufb %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: psrad $31, %xmm0
; SSSE3-NEXT: psrad $24, %xmm1
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; SSSE3-NEXT: movdqa %xmm0, %xmm4
; SSSE3-NEXT: psrad $31, %xmm4
[X86][SSE] Added 16i8 -> 8i64 sext test Shows poor codegen for AVX2 llvm-svn: 266560
2016-04-17 23:10:42 +08:00
; SSSE3-NEXT: psrad $24, %xmm0
[X86][SSE] Added support for combining bit-shifts with shuffles. Bit-shifts by a whole number of bytes can be represented as a shuffle mask suitable for combining. Added a 'getFauxShuffleMask' function to allow us to create shuffle masks from other suitable operations. llvm-svn: 288040
2016-11-29 00:25:01 +08:00
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; SSSE3-NEXT: pshufb %xmm2, %xmm3
; SSSE3-NEXT: movdqa %xmm3, %xmm2
[X86][SSE] Added 16i8 -> 8i64 sext test Shows poor codegen for AVX2 llvm-svn: 266560
2016-04-17 23:10:42 +08:00
; SSSE3-NEXT: psrad $31, %xmm2
[X86][SSE] Added support for combining bit-shifts with shuffles. Bit-shifts by a whole number of bytes can be represented as a shuffle mask suitable for combining. Added a 'getFauxShuffleMask' function to allow us to create shuffle masks from other suitable operations. llvm-svn: 288040
2016-11-29 00:25:01 +08:00
; SSSE3-NEXT: psrad $24, %xmm3
; SSSE3-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
[X86][SSE] Added 16i8 -> 8i64 sext test Shows poor codegen for AVX2 llvm-svn: 266560
2016-04-17 23:10:42 +08:00
; SSSE3-NEXT: movdqa %xmm2, %xmm4
; SSSE3-NEXT: psrad $31, %xmm4
; SSSE3-NEXT: psrad $24, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i8_to_8i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbq %xmm0, %xmm4
; SSE41-NEXT: movdqa %xmm0, %xmm1
; SSE41-NEXT: psrld $16, %xmm1
; SSE41-NEXT: pmovsxbq %xmm1, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
; SSE41-NEXT: pmovsxbq %xmm2, %xmm2
; SSE41-NEXT: psrlq $48, %xmm0
; SSE41-NEXT: pmovsxbq %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm4, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_16i8_to_8i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbq %xmm0, %xmm1
; AVX1-NEXT: vpsrld $16, %xmm0, %xmm2
; AVX1-NEXT: vpmovsxbq %xmm2, %xmm2
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm2
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
; AVX1-NEXT: vpmovsxbq %xmm1, %xmm1
; AVX1-NEXT: vpsrlq $48, %xmm0, %xmm0
; AVX1-NEXT: vpmovsxbq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm1
; AVX1-NEXT: vmovaps %ymm2, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_16i8_to_8i64:
; AVX2: # BB#0: # %entry
[X86][AVX2] Improve sign/zero extension on AVX2 targets Split extensions to large vectors into 256-bit chunks - the equivalent of what we do with pre-AVX2 into 128-bit chunks llvm-svn: 277939
2016-08-07 05:21:12 +08:00
; AVX2-NEXT: vpmovsxbq %xmm0, %ymm2
[X86][SSE] Added 16i8 -> 8i64 sext test Shows poor codegen for AVX2 llvm-svn: 266560
2016-04-17 23:10:42 +08:00
; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
[X86][AVX2] Improve sign/zero extension on AVX2 targets Split extensions to large vectors into 256-bit chunks - the equivalent of what we do with pre-AVX2 into 128-bit chunks llvm-svn: 277939
2016-08-07 05:21:12 +08:00
; AVX2-NEXT: vpmovsxbq %xmm0, %ymm1
[X86][SSE] Added 16i8 -> 8i64 sext test Shows poor codegen for AVX2 llvm-svn: 266560
2016-04-17 23:10:42 +08:00
; AVX2-NEXT: vmovdqa %ymm2, %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_16i8_to_8i64:
; AVX512: # BB#0: # %entry
[AVX-512] Add support for creating SIGN_EXTEND_VECTOR_INREG and ZERO_EXTEND_VECTOR_INREG for 512-bit vectors to support vpmovzxbq and vpmovsxbq. Summary: The one tricky thing about this is that the sign/zero_extend_inreg uses v64i8 as an input type which isn't legal without BWI support. Though the vpmovsxbq and vpmovzxbq instructions themselves don't require BWI. To support this we need to add custom lowering for ZERO_EXTEND_VECTOR_INREG with v64i8 input. This can mostly reuse the existing sign extend code with a couple checks for sign extend vs zero extend added. Reviewers: delena, RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D25594 llvm-svn: 285053
2016-10-25 12:00:29 +08:00
; AVX512-NEXT: vpmovsxbq %xmm0, %zmm0
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-NEXT: retq
;
[X86][SSE] Added 16i8 -> 8i64 sext test Shows poor codegen for AVX2 llvm-svn: 266560
2016-04-17 23:10:42 +08:00
; X32-SSE41-LABEL: sext_16i8_to_8i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxbq %xmm0, %xmm4
; X32-SSE41-NEXT: movdqa %xmm0, %xmm1
; X32-SSE41-NEXT: psrld $16, %xmm1
; X32-SSE41-NEXT: pmovsxbq %xmm1, %xmm1
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
; X32-SSE41-NEXT: pmovsxbq %xmm2, %xmm2
; X32-SSE41-NEXT: psrlq $48, %xmm0
; X32-SSE41-NEXT: pmovsxbq %xmm0, %xmm3
; X32-SSE41-NEXT: movdqa %xmm4, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <16 x i8> %A, <16 x i8> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%C = sext <8 x i8> %B to <8 x i64>
ret <8 x i64> %C
}
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
define <4 x i32> @sext_8i16_to_4i32(<8 x i16> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_8i16_to_4i32:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: psrad $16, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_8i16_to_4i32:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: psrad $16, %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_8i16_to_4i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxwd %xmm0, %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: sext_8i16_to_4i32:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxwd %xmm0, %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: sext_8i16_to_4i32:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxwd %xmm0, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <8 x i16> %A, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%C = sext <4 x i16> %B to <4 x i32>
ret <4 x i32> %C
}
Optimization for SIGN_EXTEND operation on AVX. Special handling was added for v4i32 -> v4i64 and v8i16 -> v8i32 extensions. llvm-svn: 149600
2012-02-02 17:10:43 +08:00
define <8 x i32> @sext_8i16_to_8i32(<8 x i16> %A) nounwind uwtable readnone ssp {
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-LABEL: sext_8i16_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0: # %entry
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: psrad $16, %xmm2
[X86][SSE] Combine UNPCKL with vector_shuffle into UNPCKH to save one instruction for sext from v16i8 to v16i16 and v8i16 to v8i32. This patch is enabling combining UNPCKL with vector_shuffle that moves the upper half of a vector into the lower half, into a UNPCKH instruction. For example: t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1, undef:v16i8 t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2 will be combined to: t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1 Differential revision: http://reviews.llvm.org/D14399 llvm-svn: 253067
2015-11-14 03:47:43 +08:00
; SSE2-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: psrad $16, %xmm1
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE2-NEXT: movdqa %xmm2, %xmm0
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_8i16_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0: # %entry
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: psrad $16, %xmm2
[X86][SSE] Combine UNPCKL with vector_shuffle into UNPCKH to save one instruction for sext from v16i8 to v16i16 and v8i16 to v8i32. This patch is enabling combining UNPCKL with vector_shuffle that moves the upper half of a vector into the lower half, into a UNPCKH instruction. For example: t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1, undef:v16i8 t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2 will be combined to: t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1 Differential revision: http://reviews.llvm.org/D14399 llvm-svn: 253067
2015-11-14 03:47:43 +08:00
; SSSE3-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: psrad $16, %xmm1
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSSE3-NEXT: movdqa %xmm2, %xmm0
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_8i16_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0: # %entry
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE41-NEXT: pmovsxwd %xmm0, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxwd %xmm0, %xmm1
; SSE41-NEXT: movdqa %xmm2, %xmm0
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
; AVX1-LABEL: sext_8i16_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX1: # BB#0: # %entry
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX1-NEXT: vpmovsxwd %xmm0, %xmm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX1-NEXT: vpmovsxwd %xmm0, %xmm0
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX2-LABEL: sext_8i16_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX2: # BB#0: # %entry
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX2-NEXT: vpmovsxwd %xmm0, %ymm0
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; AVX2-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_8i16_to_8i32:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxwd %xmm0, %ymm0
; AVX512-NEXT: retq
;
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-LABEL: sext_8i16_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0: # %entry
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; X32-SSE41-NEXT: pmovsxwd %xmm0, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxwd %xmm0, %xmm1
; X32-SSE41-NEXT: movdqa %xmm2, %xmm0
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: retl
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
entry:
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
%B = sext <8 x i16> %A to <8 x i32>
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
ret <8 x i32> %B
}
[X86][AVX512] Add sext/zext to 512-bit vector tests llvm-svn: 277956
2016-08-07 20:10:46 +08:00
define <16 x i32> @sext_16i16_to_16i32(<16 x i16> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_16i16_to_16i32:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; SSE2-NEXT: psrad $16, %xmm4
; SSE2-NEXT: punpckhwd {{.*#+}} xmm5 = xmm5[4],xmm0[4],xmm5[5],xmm0[5],xmm5[6],xmm0[6],xmm5[7],xmm0[7]
; SSE2-NEXT: psrad $16, %xmm5
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE2-NEXT: psrad $16, %xmm2
; SSE2-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; SSE2-NEXT: psrad $16, %xmm3
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: movdqa %xmm5, %xmm1
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_16i16_to_16i32:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; SSSE3-NEXT: psrad $16, %xmm4
; SSSE3-NEXT: punpckhwd {{.*#+}} xmm5 = xmm5[4],xmm0[4],xmm5[5],xmm0[5],xmm5[6],xmm0[6],xmm5[7],xmm0[7]
; SSSE3-NEXT: psrad $16, %xmm5
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSSE3-NEXT: psrad $16, %xmm2
; SSSE3-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; SSSE3-NEXT: psrad $16, %xmm3
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: movdqa %xmm5, %xmm1
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_16i16_to_16i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxwd %xmm0, %xmm5
; SSE41-NEXT: pmovsxwd %xmm1, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxwd %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
; SSE41-NEXT: pmovsxwd %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm5, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_16i16_to_16i32:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxwd %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxwd %xmm2, %xmm2
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm2
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
; AVX1-NEXT: vpmovsxwd %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxwd %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm1
; AVX1-NEXT: vmovaps %ymm2, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_16i16_to_16i32:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxwd %xmm0, %ymm2
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm0
; AVX2-NEXT: vpmovsxwd %xmm0, %ymm1
; AVX2-NEXT: vmovdqa %ymm2, %ymm0
; AVX2-NEXT: retq
;
; AVX512-LABEL: sext_16i16_to_16i32:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxwd %ymm0, %zmm0
; AVX512-NEXT: retq
;
; X32-SSE41-LABEL: sext_16i16_to_16i32:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxwd %xmm0, %xmm5
; X32-SSE41-NEXT: pmovsxwd %xmm1, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxwd %xmm0, %xmm4
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
; X32-SSE41-NEXT: pmovsxwd %xmm0, %xmm3
; X32-SSE41-NEXT: movdqa %xmm5, %xmm0
; X32-SSE41-NEXT: movdqa %xmm4, %xmm1
; X32-SSE41-NEXT: retl
entry:
%B = sext <16 x i16> %A to <16 x i32>
ret <16 x i32> %B
}
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
define <2 x i64> @sext_8i16_to_2i64(<8 x i16> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_8i16_to_2i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: psrad $16, %xmm0
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_8i16_to_2i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: psrad $16, %xmm0
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_8i16_to_2i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxwq %xmm0, %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: sext_8i16_to_2i64:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxwq %xmm0, %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: sext_8i16_to_2i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxwq %xmm0, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <8 x i16> %A, <8 x i16> undef, <2 x i32> <i32 0, i32 1>
%C = sext <2 x i16> %B to <2 x i64>
ret <2 x i64> %C
}
define <4 x i64> @sext_8i16_to_4i64(<8 x i16> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_8i16_to_4i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: psrad $16, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: psrad $31, %xmm0
; SSE2-NEXT: psrad $16, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE2-NEXT: movdqa %xmm2, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_8i16_to_4i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: psrad $16, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: psrad $31, %xmm0
; SSSE3-NEXT: psrad $16, %xmm1
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSSE3-NEXT: movdqa %xmm2, %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_8i16_to_4i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxwq %xmm0, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; SSE41-NEXT: pmovsxwq %xmm0, %xmm1
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_8i16_to_4i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxwq %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; AVX1-NEXT: vpmovsxwq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_8i16_to_4i64:
; AVX2: # BB#0: # %entry
[X86][AVX2] Fix SIGN_EXTEND vector handling on AVX2 targets. On AVX2 target we are poorly legalizing SIGN_EXTEND ops for which the input's legalized type doesn't have the same number of elements as the destination, resulting in an ANY_EXTEND followed by a SIGN_EXTEND_INREG. This patch uses the existing SIGN_EXTEND -> SIGN_EXTEND_VECTOR_INREG combine to extend the input to the size of the result and using SIGN_EXTEND_VECTOR_INREG instead. Differential Revision: http://reviews.llvm.org/D16994 llvm-svn: 260210
2016-02-09 16:19:19 +08:00
; AVX2-NEXT: vpmovsxwq %xmm0, %ymm0
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_8i16_to_4i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxwq %xmm0, %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; X32-SSE41-LABEL: sext_8i16_to_4i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxwq %xmm0, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; X32-SSE41-NEXT: pmovsxwq %xmm0, %xmm1
; X32-SSE41-NEXT: movdqa %xmm2, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <8 x i16> %A, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%C = sext <4 x i16> %B to <4 x i64>
ret <4 x i64> %C
}
[X86][AVX512] Add sext/zext to 512-bit vector tests llvm-svn: 277956
2016-08-07 20:10:46 +08:00
define <8 x i64> @sext_8i16_to_8i64(<8 x i16> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_8i16_to_8i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; SSE2-NEXT: movdqa %xmm4, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: psrad $16, %xmm4
; SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]
; SSE2-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: psrad $16, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: psrad $31, %xmm3
; SSE2-NEXT: psrad $16, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
; SSE2-NEXT: movdqa %xmm3, %xmm0
; SSE2-NEXT: psrad $31, %xmm0
; SSE2-NEXT: psrad $16, %xmm3
; SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_8i16_to_8i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; SSSE3-NEXT: movdqa %xmm4, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: psrad $16, %xmm4
; SSSE3-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]
; SSSE3-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: psrad $16, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: psrad $31, %xmm3
; SSSE3-NEXT: psrad $16, %xmm1
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: psrad $31, %xmm0
; SSSE3-NEXT: psrad $16, %xmm3
; SSSE3-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_8i16_to_8i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxwq %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
; SSE41-NEXT: pmovsxwq %xmm1, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxwq %xmm2, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; SSE41-NEXT: pmovsxwq %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm4, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_8i16_to_8i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxwq %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
; AVX1-NEXT: vpmovsxwq %xmm2, %xmm2
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm2
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxwq %xmm1, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; AVX1-NEXT: vpmovsxwq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm1
; AVX1-NEXT: vmovaps %ymm2, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_8i16_to_8i64:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxwq %xmm0, %ymm2
; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX2-NEXT: vpmovsxwq %xmm0, %ymm1
; AVX2-NEXT: vmovdqa %ymm2, %ymm0
; AVX2-NEXT: retq
;
; AVX512-LABEL: sext_8i16_to_8i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxwq %xmm0, %zmm0
; AVX512-NEXT: retq
;
; X32-SSE41-LABEL: sext_8i16_to_8i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxwq %xmm0, %xmm4
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
; X32-SSE41-NEXT: pmovsxwq %xmm1, %xmm1
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxwq %xmm2, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; X32-SSE41-NEXT: pmovsxwq %xmm0, %xmm3
; X32-SSE41-NEXT: movdqa %xmm4, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = sext <8 x i16> %A to <8 x i64>
ret <8 x i64> %B
}
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
define <2 x i64> @sext_4i32_to_2i64(<4 x i32> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_4i32_to_2i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_4i32_to_2i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_4i32_to_2i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxdq %xmm0, %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: sext_4i32_to_2i64:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: sext_4i32_to_2i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm0
; X32-SSE41-NEXT: retl
entry:
%B = shufflevector <4 x i32> %A, <4 x i32> undef, <2 x i32> <i32 0, i32 1>
%C = sext <2 x i32> %B to <2 x i64>
ret <2 x i64> %C
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
}
define <4 x i64> @sext_4i32_to_4i64(<4 x i32> %A) nounwind uwtable readnone ssp {
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-LABEL: sext_4i32_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0: # %entry
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_4i32_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0: # %entry
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_4i32_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0: # %entry
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE41-NEXT: pmovsxdq %xmm0, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxdq %xmm0, %xmm1
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
; AVX1-LABEL: sext_4i32_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX1: # BB#0: # %entry
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_4i32_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX2: # BB#0: # %entry
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX2-NEXT: vpmovsxdq %xmm0, %ymm0
; AVX2-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_4i32_to_4i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxdq %xmm0, %ymm0
; AVX512-NEXT: retq
;
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-LABEL: sext_4i32_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0: # %entry
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm1
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: movdqa %xmm2, %xmm0
; X32-SSE41-NEXT: retl
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
entry:
Optimization for SIGN_EXTEND operation on AVX. Special handling was added for v4i32 -> v4i64 and v8i16 -> v8i32 extensions. llvm-svn: 149600
2012-02-02 17:10:43 +08:00
%B = sext <4 x i32> %A to <4 x i64>
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
ret <4 x i64> %B
Optimization for SIGN_EXTEND operation on AVX. Special handling was added for v4i32 -> v4i64 and v8i16 -> v8i32 extensions. llvm-svn: 149600
2012-02-02 17:10:43 +08:00
}
Optimized load + SIGN_EXTEND patterns in the X86 backend. llvm-svn: 170506
2012-12-19 15:50:20 +08:00
[X86][AVX512] Add sext/zext to 512-bit vector tests llvm-svn: 277956
2016-08-07 20:10:46 +08:00
define <8 x i64> @sext_8i32_to_8i64(<8 x i32> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_8i32_to_8i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: movdqa %xmm0, %xmm3
; SSE2-NEXT: psrad $31, %xmm3
; SSE2-NEXT: movdqa %xmm2, %xmm4
; SSE2-NEXT: psrad $31, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: psrad $31, %xmm3
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[2,3,0,1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]
; SSE2-NEXT: movdqa %xmm3, %xmm4
; SSE2-NEXT: psrad $31, %xmm4
; SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_8i32_to_8i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: psrad $31, %xmm3
; SSSE3-NEXT: movdqa %xmm2, %xmm4
; SSSE3-NEXT: psrad $31, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: psrad $31, %xmm3
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm2[2,3,0,1]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]
; SSSE3-NEXT: movdqa %xmm3, %xmm4
; SSSE3-NEXT: psrad $31, %xmm4
; SSSE3-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_8i32_to_8i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxdq %xmm0, %xmm5
; SSE41-NEXT: pmovsxdq %xmm1, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxdq %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
; SSE41-NEXT: pmovsxdq %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm5, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: sext_8i32_to_8i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxdq %xmm2, %xmm2
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm2
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm1
; AVX1-NEXT: vmovaps %ymm2, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_8i32_to_8i64:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxdq %xmm0, %ymm2
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm0
; AVX2-NEXT: vpmovsxdq %xmm0, %ymm1
; AVX2-NEXT: vmovdqa %ymm2, %ymm0
; AVX2-NEXT: retq
;
; AVX512-LABEL: sext_8i32_to_8i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxdq %ymm0, %zmm0
; AVX512-NEXT: retq
;
; X32-SSE41-LABEL: sext_8i32_to_8i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm5
; X32-SSE41-NEXT: pmovsxdq %xmm1, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm4
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm3
; X32-SSE41-NEXT: movdqa %xmm5, %xmm0
; X32-SSE41-NEXT: movdqa %xmm4, %xmm1
; X32-SSE41-NEXT: retl
entry:
%B = sext <8 x i32> %A to <8 x i64>
ret <8 x i64> %B
}
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
define <2 x i64> @load_sext_2i1_to_2i64(<2 x i1> *%ptr) {
; SSE-LABEL: load_sext_2i1_to_2i64:
; SSE: # BB#0: # %entry
; SSE-NEXT: movzbl (%rdi), %eax
; SSE-NEXT: movq %rax, %rcx
; SSE-NEXT: shlq $62, %rcx
; SSE-NEXT: sarq $63, %rcx
; SSE-NEXT: movd %rcx, %xmm1
; SSE-NEXT: shlq $63, %rax
; SSE-NEXT: sarq $63, %rax
; SSE-NEXT: movd %rax, %xmm0
; SSE-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX1-LABEL: load_sext_2i1_to_2i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: movzbl (%rdi), %eax
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $62, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vmovq %rcx, %xmm0
; AVX1-NEXT: shlq $63, %rax
; AVX1-NEXT: sarq $63, %rax
; AVX1-NEXT: vmovq %rax, %xmm1
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_2i1_to_2i64:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: movzbl (%rdi), %eax
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $62, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vmovq %rcx, %xmm0
; AVX2-NEXT: shlq $63, %rax
; AVX2-NEXT: sarq $63, %rax
; AVX2-NEXT: vmovq %rax, %xmm1
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX2-NEXT: retq
;
[X86][AVX512BW] Add sext/zext AVX512BW 512-bit vector tests llvm-svn: 277957
2016-08-07 20:41:36 +08:00
; AVX512F-LABEL: load_sext_2i1_to_2i64:
; AVX512F: # BB#0: # %entry
; AVX512F-NEXT: movzbl (%rdi), %eax
; AVX512F-NEXT: kmovw %eax, %k1
; AVX512F-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512F-NEXT: # kill: %XMM0<def> %XMM0<kill> %ZMM0<kill>
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: load_sext_2i1_to_2i64:
; AVX512BW: # BB#0: # %entry
; AVX512BW-NEXT: movzbl (%rdi), %eax
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512BW-NEXT: # kill: %XMM0<def> %XMM0<kill> %ZMM0<kill>
; AVX512BW-NEXT: retq
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
;
; X32-SSE41-LABEL: load_sext_2i1_to_2i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: movzbl (%eax), %eax
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $31, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: movd %ecx, %xmm0
; X32-SSE41-NEXT: pinsrd $1, %ecx, %xmm0
; X32-SSE41-NEXT: shll $30, %eax
; X32-SSE41-NEXT: sarl $31, %eax
; X32-SSE41-NEXT: pinsrd $2, %eax, %xmm0
; X32-SSE41-NEXT: pinsrd $3, %eax, %xmm0
; X32-SSE41-NEXT: retl
entry:
%X = load <2 x i1>, <2 x i1>* %ptr
%Y = sext <2 x i1> %X to <2 x i64>
ret <2 x i64> %Y
}
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
define <2 x i64> @load_sext_2i8_to_2i64(<2 x i8> *%ptr) {
; SSE2-LABEL: load_sext_2i8_to_2i64:
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; SSE2: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE2-NEXT: movzwl (%rdi), %eax
; SSE2-NEXT: movd %eax, %xmm0
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: psrad $24, %xmm0
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; SSE2-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-LABEL: load_sext_2i8_to_2i64:
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; SSSE3: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-NEXT: movzwl (%rdi), %eax
; SSSE3-NEXT: movd %eax, %xmm0
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: psrad $24, %xmm0
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; SSSE3-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-LABEL: load_sext_2i8_to_2i64:
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; SSE41: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-NEXT: pmovsxbq (%rdi), %xmm0
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; SSE41-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX-LABEL: load_sext_2i8_to_2i64:
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; AVX: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX-NEXT: vpmovsxbq (%rdi), %xmm0
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; AVX-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: load_sext_2i8_to_2i64:
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; X32-SSE41: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxbq (%eax), %xmm0
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
; X32-SSE41-NEXT: retl
entry:
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
%X = load <2 x i8>, <2 x i8>* %ptr
%Y = sext <2 x i8> %X to <2 x i64>
ret <2 x i64> %Y
[X86][SSE] Sign extension for target vector sizes less than 128 bits (pt1) This patch adds support for sign extension for sub 128-bit vectors, such as to v2i32. It concatenates with UNDEF subvectors up to 128-bits, performs the sign extension (i.e. as v4i32) and then extracts the target subvector. Patch 1/2 of D10589 - the second patch covers the conversion of v2i8/v2i16 to v2f64. llvm-svn: 241323
2015-07-03 15:51:01 +08:00
}
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
define <4 x i32> @load_sext_4i1_to_4i32(<4 x i1> *%ptr) {
; SSE2-LABEL: load_sext_4i1_to_4i32:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movzbl (%rdi), %eax
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $60, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $62, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $61, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-NEXT: shlq $63, %rax
; SSE2-NEXT: sarq $63, %rax
; SSE2-NEXT: movd %eax, %xmm0
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_4i1_to_4i32:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movzbl (%rdi), %eax
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $60, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $62, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm1
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $61, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-NEXT: shlq $63, %rax
; SSSE3-NEXT: sarq $63, %rax
; SSSE3-NEXT: movd %eax, %xmm0
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_4i1_to_4i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: movzbl (%rdi), %eax
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $62, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: movq %rax, %rdx
; SSE41-NEXT: shlq $63, %rdx
; SSE41-NEXT: sarq $63, %rdx
; SSE41-NEXT: movd %edx, %xmm0
; SSE41-NEXT: pinsrd $1, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $61, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrd $2, %ecx, %xmm0
; SSE41-NEXT: shlq $60, %rax
; SSE41-NEXT: sarq $63, %rax
; SSE41-NEXT: pinsrd $3, %eax, %xmm0
; SSE41-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX1-LABEL: load_sext_4i1_to_4i32:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: movzbl (%rdi), %eax
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $62, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: movq %rax, %rdx
; AVX1-NEXT: shlq $63, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vmovd %edx, %xmm0
; AVX1-NEXT: vpinsrd $1, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $61, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrd $2, %ecx, %xmm0, %xmm0
; AVX1-NEXT: shlq $60, %rax
; AVX1-NEXT: sarq $63, %rax
; AVX1-NEXT: vpinsrd $3, %eax, %xmm0, %xmm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_4i1_to_4i32:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: movzbl (%rdi), %eax
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $62, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: movq %rax, %rdx
; AVX2-NEXT: shlq $63, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vmovd %edx, %xmm0
; AVX2-NEXT: vpinsrd $1, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $61, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrd $2, %ecx, %xmm0, %xmm0
; AVX2-NEXT: shlq $60, %rax
; AVX2-NEXT: sarq $63, %rax
; AVX2-NEXT: vpinsrd $3, %eax, %xmm0, %xmm0
; AVX2-NEXT: retq
;
[X86][AVX512BW] Add sext/zext AVX512BW 512-bit vector tests llvm-svn: 277957
2016-08-07 20:41:36 +08:00
; AVX512F-LABEL: load_sext_4i1_to_4i32:
; AVX512F: # BB#0: # %entry
; AVX512F-NEXT: movzbl (%rdi), %eax
; AVX512F-NEXT: kmovw %eax, %k1
; AVX512F-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512F-NEXT: vpmovqd %zmm0, %ymm0
; AVX512F-NEXT: # kill: %XMM0<def> %XMM0<kill> %YMM0<kill>
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: load_sext_4i1_to_4i32:
; AVX512BW: # BB#0: # %entry
; AVX512BW-NEXT: movzbl (%rdi), %eax
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512BW-NEXT: vpmovqd %zmm0, %ymm0
; AVX512BW-NEXT: # kill: %XMM0<def> %XMM0<kill> %YMM0<kill>
; AVX512BW-NEXT: retq
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
;
; X32-SSE41-LABEL: load_sext_4i1_to_4i32:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: movl (%eax), %eax
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $30, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: movl %eax, %edx
; X32-SSE41-NEXT: shll $31, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: movd %edx, %xmm0
; X32-SSE41-NEXT: pinsrd $1, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $29, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrd $2, %ecx, %xmm0
; X32-SSE41-NEXT: shll $28, %eax
; X32-SSE41-NEXT: sarl $31, %eax
; X32-SSE41-NEXT: pinsrd $3, %eax, %xmm0
; X32-SSE41-NEXT: retl
entry:
%X = load <4 x i1>, <4 x i1>* %ptr
%Y = sext <4 x i1> %X to <4 x i32>
ret <4 x i32> %Y
}
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
define <4 x i32> @load_sext_4i8_to_4i32(<4 x i8> *%ptr) {
; SSE2-LABEL: load_sext_4i8_to_4i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE2-NEXT: psrad $24, %xmm0
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-LABEL: load_sext_4i8_to_4i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-NEXT: psrad $24, %xmm0
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-LABEL: load_sext_4i8_to_4i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-NEXT: pmovsxbd (%rdi), %xmm0
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX-LABEL: load_sext_4i8_to_4i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX-NEXT: vpmovsxbd (%rdi), %xmm0
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: load_sext_4i8_to_4i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0: # %entry
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-NEXT: pmovsxbd (%eax), %xmm0
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: retl
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
entry:
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
%X = load <4 x i8>, <4 x i8>* %ptr
%Y = sext <4 x i8> %X to <4 x i32>
ret <4 x i32> %Y
Optimized load + SIGN_EXTEND patterns in the X86 backend. llvm-svn: 170506
2012-12-19 15:50:20 +08:00
}
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
define <4 x i64> @load_sext_4i1_to_4i64(<4 x i1> *%ptr) {
; SSE2-LABEL: load_sext_4i1_to_4i64:
; SSE2: # BB#0: # %entry
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSE2-NEXT: movl (%rdi), %eax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $3, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl %ecx
; SSE2-NEXT: movd %ecx, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSE2-NEXT: movd %eax, %xmm2
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: shrl $2, %eax
; SSE2-NEXT: movd %eax, %xmm0
; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSE2-NEXT: pand {{.*}}(%rip), %xmm2
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,1,1,3]
; SSE2-NEXT: psllq $63, %xmm0
; SSE2-NEXT: psrad $31, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[X86][SSE] Match zero/any extension shuffles that don't start from the first element This patch generalizes the lowering of shuffles as zero extensions to allow extensions that don't start from the first element. It now recognises extensions starting anywhere in the lower 128-bits or at the start of any higher 128-bit lane. The motivation was to reduce the number of high cost pshufb calls, but it also improves the SSE2 case as well. Differential Revision: http://reviews.llvm.org/D12561 llvm-svn: 248250
2015-09-22 16:16:08 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,1,3,3]
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: psllq $63, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_4i1_to_4i64:
; SSSE3: # BB#0: # %entry
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSSE3-NEXT: movl (%rdi), %eax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $3, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl %ecx
; SSSE3-NEXT: movd %ecx, %xmm1
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSSE3-NEXT: movd %eax, %xmm2
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: shrl $2, %eax
; SSSE3-NEXT: movd %eax, %xmm0
; SSSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSSE3-NEXT: pand {{.*}}(%rip), %xmm2
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,1,1,3]
; SSSE3-NEXT: psllq $63, %xmm0
; SSSE3-NEXT: psrad $31, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[X86][SSE] Match zero/any extension shuffles that don't start from the first element This patch generalizes the lowering of shuffles as zero extensions to allow extensions that don't start from the first element. It now recognises extensions starting anywhere in the lower 128-bits or at the start of any higher 128-bit lane. The motivation was to reduce the number of high cost pshufb calls, but it also improves the SSE2 case as well. Differential Revision: http://reviews.llvm.org/D12561 llvm-svn: 248250
2015-09-22 16:16:08 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,1,3,3]
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: psllq $63, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_4i1_to_4i64:
; SSE41: # BB#0: # %entry
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSE41-NEXT: movl (%rdi), %eax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl %ecx
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSE41-NEXT: movd %eax, %xmm1
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrd $1, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $2, %ecx
; SSE41-NEXT: pinsrd $2, %ecx, %xmm1
; SSE41-NEXT: shrl $3, %eax
; SSE41-NEXT: pinsrd $3, %eax, %xmm1
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; SSE41-NEXT: pand {{.*}}(%rip), %xmm1
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm1[0],zero,xmm1[1],zero
; SSE41-NEXT: psllq $63, %xmm0
; SSE41-NEXT: psrad $31, %xmm0
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
; SSE41-NEXT: psllq $63, %xmm1
; SSE41-NEXT: psrad $31, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE41-NEXT: retq
;
; AVX1-LABEL: load_sext_4i1_to_4i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: movzbl (%rdi), %eax
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $62, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: movq %rax, %rdx
; AVX1-NEXT: shlq $63, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vmovd %edx, %xmm0
; AVX1-NEXT: vpinsrd $1, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $61, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrd $2, %ecx, %xmm0, %xmm0
; AVX1-NEXT: shlq $60, %rax
; AVX1-NEXT: sarq $63, %rax
; AVX1-NEXT: vpinsrd $3, %eax, %xmm0, %xmm0
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_4i1_to_4i64:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: movzbl (%rdi), %eax
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $60, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vmovq %rcx, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $61, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vmovq %rcx, %xmm1
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $62, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vmovq %rcx, %xmm1
; AVX2-NEXT: shlq $63, %rax
; AVX2-NEXT: sarq $63, %rax
; AVX2-NEXT: vmovq %rax, %xmm2
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512BW] Add sext/zext AVX512BW 512-bit vector tests llvm-svn: 277957
2016-08-07 20:41:36 +08:00
; AVX512F-LABEL: load_sext_4i1_to_4i64:
; AVX512F: # BB#0: # %entry
; AVX512F-NEXT: movzbl (%rdi), %eax
; AVX512F-NEXT: kmovw %eax, %k1
; AVX512F-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512F-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: load_sext_4i1_to_4i64:
; AVX512BW: # BB#0: # %entry
; AVX512BW-NEXT: movzbl (%rdi), %eax
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512BW-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
; AVX512BW-NEXT: retq
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
;
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-LABEL: load_sext_4i1_to_4i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: movzbl (%eax), %eax
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl %ecx
[X86][SSE] Vectorize a bit (AND/XOR/OR) op if a BUILD_VECTOR has the same op for all their scalar elements. If all a BUILD_VECTOR's source elements are the same bit (AND/XOR/OR) operation type and each has one constant operand, lower to a pair of BUILD_VECTOR and just apply the bit operation to the vectors. The constant operands will form a constant vector meaning that we still only have a single BUILD_VECTOR to lower and we will have replaced all the scalarized operations with a single SSE equivalent. Its not in our interest to start make a general purpose vectorizer from this, but I'm seeing enough of these scalar bit operations from the later legalization/scalarization stages to support them at least. Differential Revision: http://reviews.llvm.org/D18492 llvm-svn: 264666
2016-03-29 05:33:52 +08:00
; X32-SSE41-NEXT: movd %eax, %xmm1
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrd $1, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $2, %ecx
; X32-SSE41-NEXT: pinsrd $2, %ecx, %xmm1
; X32-SSE41-NEXT: shrl $3, %eax
; X32-SSE41-NEXT: pinsrd $3, %eax, %xmm1
[X86] Updated test checks script to generalise LCPI symbol refs The script now replace '.LCPI888_8' style asm symbols with the {{\.LCPI.*}} re pattern - this helps stop hardcoded symbols in 32-bit x86 tests changing with every edit of the file Refreshed some tests to demonstrate the new check llvm-svn: 272488
2016-06-12 04:39:21 +08:00
; X32-SSE41-NEXT: pand {{\.LCPI.*}}, %xmm1
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm1[0],zero,xmm1[1],zero
; X32-SSE41-NEXT: psllq $63, %xmm0
; X32-SSE41-NEXT: psrad $31, %xmm0
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
; X32-SSE41-NEXT: psllq $63, %xmm1
; X32-SSE41-NEXT: psrad $31, %xmm1
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; X32-SSE41-NEXT: retl
entry:
%X = load <4 x i1>, <4 x i1>* %ptr
%Y = sext <4 x i1> %X to <4 x i64>
ret <4 x i64> %Y
}
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
define <4 x i64> @load_sext_4i8_to_4i64(<4 x i8> *%ptr) {
; SSE2-LABEL: load_sext_4i8_to_4i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movsbq 1(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm1
; SSE2-NEXT: movsbq (%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm0
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE2-NEXT: movsbq 3(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm2
; SSE2-NEXT: movsbq 2(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm1
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_4i8_to_4i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movsbq 1(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm1
; SSSE3-NEXT: movsbq (%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm0
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSSE3-NEXT: movsbq 3(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm2
; SSSE3-NEXT: movsbq 2(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm1
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_4i8_to_4i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbq (%rdi), %xmm0
; SSE41-NEXT: pmovsxbq 2(%rdi), %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: load_sext_4i8_to_4i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbd (%rdi), %xmm0
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_4i8_to_4i64:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxbq (%rdi), %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_4i8_to_4i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxbq (%rdi), %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: load_sext_4i8_to_4i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxbq (%eax), %xmm0
; X32-SSE41-NEXT: pmovsxbq 2(%eax), %xmm1
; X32-SSE41-NEXT: retl
entry:
%X = load <4 x i8>, <4 x i8>* %ptr
%Y = sext <4 x i8> %X to <4 x i64>
ret <4 x i64> %Y
}
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
define <8 x i16> @load_sext_8i1_to_8i16(<8 x i1> *%ptr) {
; SSE2-LABEL: load_sext_8i1_to_8i16:
; SSE2: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: movsbq (%rdi), %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: shrq $7, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $60, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $58, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $62, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm1
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $57, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $61, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: movq %rax, %rcx
; SSE2-NEXT: shlq $59, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm3
; SSE2-NEXT: shlq $63, %rax
; SSE2-NEXT: sarq $63, %rax
; SSE2-NEXT: movd %eax, %xmm0
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_8i1_to_8i16:
; SSSE3: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: movsbq (%rdi), %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: shrq $7, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $60, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $58, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $62, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm1
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $57, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $61, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: movq %rax, %rcx
; SSSE3-NEXT: shlq $59, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm3
; SSSE3-NEXT: shlq $63, %rax
; SSSE3-NEXT: sarq $63, %rax
; SSSE3-NEXT: movd %eax, %xmm0
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_8i1_to_8i16:
; SSE41: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: movsbq (%rdi), %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $62, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: movq %rax, %rdx
; SSE41-NEXT: shlq $63, %rdx
; SSE41-NEXT: sarq $63, %rdx
; SSE41-NEXT: movd %edx, %xmm0
; SSE41-NEXT: pinsrw $1, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $61, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrw $2, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $60, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrw $3, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $59, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrw $4, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $58, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrw $5, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $57, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrw $6, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: shrq $7, %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrw $7, %eax, %xmm0
; SSE41-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX1-LABEL: load_sext_8i1_to_8i16:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: movsbq (%rdi), %rax
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $62, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: movq %rax, %rdx
; AVX1-NEXT: shlq $63, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vmovd %edx, %xmm0
; AVX1-NEXT: vpinsrw $1, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $61, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrw $2, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $60, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrw $3, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $59, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrw $4, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $58, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $57, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm0
; AVX1-NEXT: shrq $7, %rax
; AVX1-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_8i1_to_8i16:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: movsbq (%rdi), %rax
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $62, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: movq %rax, %rdx
; AVX2-NEXT: shlq $63, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vmovd %edx, %xmm0
; AVX2-NEXT: vpinsrw $1, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $61, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrw $2, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $60, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrw $3, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $59, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrw $4, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $58, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $57, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm0
; AVX2-NEXT: shrq $7, %rax
; AVX2-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0
; AVX2-NEXT: retq
;
[X86][AVX512BW] Add sext/zext AVX512BW 512-bit vector tests llvm-svn: 277957
2016-08-07 20:41:36 +08:00
; AVX512F-LABEL: load_sext_8i1_to_8i16:
; AVX512F: # BB#0: # %entry
; AVX512F-NEXT: movzbl (%rdi), %eax
; AVX512F-NEXT: kmovw %eax, %k1
; AVX512F-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512F-NEXT: vpmovqw %zmm0, %xmm0
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: load_sext_8i1_to_8i16:
; AVX512BW: # BB#0: # %entry
; AVX512BW-NEXT: movzbl (%rdi), %eax
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512BW-NEXT: vpmovqw %zmm0, %xmm0
; AVX512BW-NEXT: retq
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
;
; X32-SSE41-LABEL: load_sext_8i1_to_8i16:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: movsbl (%eax), %eax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $30, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: movl %eax, %edx
; X32-SSE41-NEXT: shll $31, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: movd %edx, %xmm0
; X32-SSE41-NEXT: pinsrw $1, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $29, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrw $2, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $28, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrw $3, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $27, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrw $4, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $26, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrw $5, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $25, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrw $6, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: shrl $7, %eax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrw $7, %eax, %xmm0
; X32-SSE41-NEXT: retl
entry:
%X = load <8 x i1>, <8 x i1>* %ptr
%Y = sext <8 x i1> %X to <8 x i16>
ret <8 x i16> %Y
}
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
define <8 x i16> @load_sext_8i8_to_8i16(<8 x i8> *%ptr) {
; SSE2-LABEL: load_sext_8i8_to_8i16:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: psraw $8, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_8i8_to_8i16:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: psraw $8, %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_8i8_to_8i16:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbw (%rdi), %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: load_sext_8i8_to_8i16:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxbw (%rdi), %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: load_sext_8i8_to_8i16:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxbw (%eax), %xmm0
; X32-SSE41-NEXT: retl
entry:
%X = load <8 x i8>, <8 x i8>* %ptr
%Y = sext <8 x i8> %X to <8 x i16>
ret <8 x i16> %Y
}
[X86][SSE] Added 8i8 to 8i64 sext/zext tests llvm-svn: 258868
2016-01-27 06:19:22 +08:00
define <8 x i64> @load_sext_8i8_to_8i64(<8 x i8> *%ptr) {
; SSE2-LABEL: load_sext_8i8_to_8i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movsbq 1(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm1
; SSE2-NEXT: movsbq (%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm0
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE2-NEXT: movsbq 3(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm2
; SSE2-NEXT: movsbq 2(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm1
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; SSE2-NEXT: movsbq 5(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm3
; SSE2-NEXT: movsbq 4(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm2
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
; SSE2-NEXT: movsbq 7(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm4
; SSE2-NEXT: movsbq 6(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm3
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm4[0]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_8i8_to_8i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movsbq 1(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm1
; SSSE3-NEXT: movsbq (%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm0
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSSE3-NEXT: movsbq 3(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm2
; SSSE3-NEXT: movsbq 2(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm1
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; SSSE3-NEXT: movsbq 5(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm3
; SSSE3-NEXT: movsbq 4(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm2
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
; SSSE3-NEXT: movsbq 7(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm4
; SSSE3-NEXT: movsbq 6(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm3
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm4[0]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_8i8_to_8i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxbq (%rdi), %xmm0
; SSE41-NEXT: pmovsxbq 2(%rdi), %xmm1
; SSE41-NEXT: pmovsxbq 4(%rdi), %xmm2
; SSE41-NEXT: pmovsxbq 6(%rdi), %xmm3
; SSE41-NEXT: retq
;
; AVX1-LABEL: load_sext_8i8_to_8i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbd (%rdi), %xmm0
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: vpmovsxbd 4(%rdi), %xmm1
; AVX1-NEXT: vpmovsxdq %xmm1, %xmm2
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,0,1]
; AVX1-NEXT: vpmovsxdq %xmm1, %xmm1
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_8i8_to_8i64:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxbq (%rdi), %ymm0
; AVX2-NEXT: vpmovsxbq 4(%rdi), %ymm1
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_8i8_to_8i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxbq (%rdi), %zmm0
; AVX512-NEXT: retq
;
[X86][SSE] Added 8i8 to 8i64 sext/zext tests llvm-svn: 258868
2016-01-27 06:19:22 +08:00
; X32-SSE41-LABEL: load_sext_8i8_to_8i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxbq (%eax), %xmm0
; X32-SSE41-NEXT: pmovsxbq 2(%eax), %xmm1
; X32-SSE41-NEXT: pmovsxbq 4(%eax), %xmm2
; X32-SSE41-NEXT: pmovsxbq 6(%eax), %xmm3
; X32-SSE41-NEXT: retl
entry:
%X = load <8 x i8>, <8 x i8>* %ptr
%Y = sext <8 x i8> %X to <8 x i64>
ret <8 x i64> %Y
}
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
define <8 x i32> @load_sext_8i1_to_8i32(<8 x i1> *%ptr) {
; SSE2-LABEL: load_sext_8i1_to_8i32:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movzbl (%rdi), %eax
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $6, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $2, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm1
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $4, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $5, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $3, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: shrl $7, %eax
; SSE2-NEXT: movzwl %ax, %eax
; SSE2-NEXT: movd %eax, %xmm3
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: pslld $31, %xmm0
; SSE2-NEXT: psrad $31, %xmm0
[X86][SSE] Match zero/any extension shuffles that don't start from the first element This patch generalizes the lowering of shuffles as zero extensions to allow extensions that don't start from the first element. It now recognises extensions starting anywhere in the lower 128-bits or at the start of any higher 128-bit lane. The motivation was to reduce the number of high cost pshufb calls, but it also improves the SSE2 case as well. Differential Revision: http://reviews.llvm.org/D12561 llvm-svn: 248250
2015-09-22 16:16:08 +08:00
; SSE2-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: pslld $31, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_8i1_to_8i32:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movzbl (%rdi), %eax
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $6, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $2, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm1
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $4, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $5, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $3, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: shrl $7, %eax
; SSSE3-NEXT: movzwl %ax, %eax
; SSSE3-NEXT: movd %eax, %xmm3
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: pslld $31, %xmm0
; SSSE3-NEXT: psrad $31, %xmm0
[X86][SSE] Match zero/any extension shuffles that don't start from the first element This patch generalizes the lowering of shuffles as zero extensions to allow extensions that don't start from the first element. It now recognises extensions starting anywhere in the lower 128-bits or at the start of any higher 128-bit lane. The motivation was to reduce the number of high cost pshufb calls, but it also improves the SSE2 case as well. Differential Revision: http://reviews.llvm.org/D12561 llvm-svn: 248250
2015-09-22 16:16:08 +08:00
; SSSE3-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: pslld $31, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_8i1_to_8i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: movzbl (%rdi), %eax
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: movl %eax, %edx
; SSE41-NEXT: andl $1, %edx
; SSE41-NEXT: movd %edx, %xmm1
; SSE41-NEXT: pinsrw $1, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $2, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrw $2, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $3, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrw $3, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $4, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrw $4, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $5, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrw $5, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $6, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrw $6, %ecx, %xmm1
; SSE41-NEXT: shrl $7, %eax
; SSE41-NEXT: movzwl %ax, %eax
; SSE41-NEXT: pinsrw $7, %eax, %xmm1
; SSE41-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
; SSE41-NEXT: pslld $31, %xmm0
; SSE41-NEXT: psrad $31, %xmm0
[X86][SSE] Use lowerVectorShuffleWithUNPCK instead of custom matches. Most 128-bit and 256-bit shuffles were manually matching UNPCK patterns - use lowerVectorShuffleWithUNPCK to be more thorough. llvm-svn: 251211
2015-10-25 06:45:04 +08:00
; SSE41-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pslld $31, %xmm1
; SSE41-NEXT: psrad $31, %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: load_sext_8i1_to_8i32:
; AVX1: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: movsbq (%rdi), %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $58, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: movq %rax, %rdx
; AVX1-NEXT: shlq $59, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vmovd %edx, %xmm0
; AVX1-NEXT: vpinsrd $1, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $57, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrd $2, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: shrq $7, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: vpinsrd $3, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $62, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: movq %rax, %rdx
; AVX1-NEXT: shlq $63, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vmovd %edx, %xmm1
; AVX1-NEXT: vpinsrd $1, %ecx, %xmm1, %xmm1
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $61, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrd $2, %ecx, %xmm1, %xmm1
; AVX1-NEXT: shlq $60, %rax
; AVX1-NEXT: sarq $63, %rax
; AVX1-NEXT: vpinsrd $3, %eax, %xmm1, %xmm1
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_8i1_to_8i32:
; AVX2: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: movsbq (%rdi), %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $58, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: movq %rax, %rdx
; AVX2-NEXT: shlq $59, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vmovd %edx, %xmm0
; AVX2-NEXT: vpinsrd $1, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $57, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrd $2, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: shrq $7, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: vpinsrd $3, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $62, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: movq %rax, %rdx
; AVX2-NEXT: shlq $63, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vmovd %edx, %xmm1
; AVX2-NEXT: vpinsrd $1, %ecx, %xmm1, %xmm1
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $61, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrd $2, %ecx, %xmm1, %xmm1
; AVX2-NEXT: shlq $60, %rax
; AVX2-NEXT: sarq $63, %rax
; AVX2-NEXT: vpinsrd $3, %eax, %xmm1, %xmm1
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512BW] Add sext/zext AVX512BW 512-bit vector tests llvm-svn: 277957
2016-08-07 20:41:36 +08:00
; AVX512F-LABEL: load_sext_8i1_to_8i32:
; AVX512F: # BB#0: # %entry
; AVX512F-NEXT: movzbl (%rdi), %eax
; AVX512F-NEXT: kmovw %eax, %k1
; AVX512F-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512F-NEXT: vpmovqd %zmm0, %ymm0
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: load_sext_8i1_to_8i32:
; AVX512BW: # BB#0: # %entry
; AVX512BW-NEXT: movzbl (%rdi), %eax
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm0 {%k1} {z}
; AVX512BW-NEXT: vpmovqd %zmm0, %ymm0
; AVX512BW-NEXT: retq
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
;
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-LABEL: load_sext_8i1_to_8i32:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: movzbl (%eax), %eax
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: movl %eax, %edx
; X32-SSE41-NEXT: andl $1, %edx
; X32-SSE41-NEXT: movd %edx, %xmm1
; X32-SSE41-NEXT: pinsrw $1, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $2, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrw $2, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $3, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrw $3, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $4, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrw $4, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $5, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrw $5, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $6, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrw $6, %ecx, %xmm1
; X32-SSE41-NEXT: shrl $7, %eax
; X32-SSE41-NEXT: pinsrw $7, %eax, %xmm1
; X32-SSE41-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
; X32-SSE41-NEXT: pslld $31, %xmm0
; X32-SSE41-NEXT: psrad $31, %xmm0
[X86][SSE] Use lowerVectorShuffleWithUNPCK instead of custom matches. Most 128-bit and 256-bit shuffles were manually matching UNPCK patterns - use lowerVectorShuffleWithUNPCK to be more thorough. llvm-svn: 251211
2015-10-25 06:45:04 +08:00
; X32-SSE41-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pslld $31, %xmm1
; X32-SSE41-NEXT: psrad $31, %xmm1
; X32-SSE41-NEXT: retl
entry:
%X = load <8 x i1>, <8 x i1>* %ptr
%Y = sext <8 x i1> %X to <8 x i32>
ret <8 x i32> %Y
}
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
define <8 x i32> @load_sext_8i8_to_8i32(<8 x i8> *%ptr) {
; SSE2-LABEL: load_sext_8i8_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0: # %entry
[X86][SSE] Tidied up vector extend/truncation tests. NFCI. llvm-svn: 241995
2015-07-13 01:40:49 +08:00
; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; SSE2-NEXT: psrad $24, %xmm0
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE2-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
; SSE2-NEXT: psrad $24, %xmm1
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; SSE2-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-LABEL: load_sext_8i8_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0: # %entry
[X86][SSE] Tidied up vector extend/truncation tests. NFCI. llvm-svn: 241995
2015-07-13 01:40:49 +08:00
; SSSE3-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; SSSE3-NEXT: psrad $24, %xmm0
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: psrad $24, %xmm1
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; SSSE3-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-LABEL: load_sext_8i8_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0: # %entry
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: pmovsxbd (%rdi), %xmm0
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-NEXT: pmovsxbd 4(%rdi), %xmm1
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX1-LABEL: load_sext_8i8_to_8i32:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbw (%rdi), %xmm0
; AVX1-NEXT: vpmovsxwd %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxwd %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX2-LABEL: load_sext_8i8_to_8i32:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxbd (%rdi), %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_8i8_to_8i32:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxbd (%rdi), %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: load_sext_8i8_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0: # %entry
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxbd (%eax), %xmm0
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-NEXT: pmovsxbd 4(%eax), %xmm1
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: retl
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
entry:
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
%X = load <8 x i8>, <8 x i8>* %ptr
%Y = sext <8 x i8> %X to <8 x i32>
ret <8 x i32> %Y
Optimized load + SIGN_EXTEND patterns in the X86 backend. llvm-svn: 170506
2012-12-19 15:50:20 +08:00
}
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
define <16 x i8> @load_sext_16i1_to_16i8(<16 x i1> *%ptr) nounwind readnone {
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-LABEL: load_sext_16i1_to_16i8:
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: pushq %rbp
; SSE2-NEXT: pushq %r15
; SSE2-NEXT: pushq %r14
; SSE2-NEXT: pushq %r13
; SSE2-NEXT: pushq %r12
; SSE2-NEXT: pushq %rbx
; SSE2-NEXT: movswq (%rdi), %rax
; SSE2-NEXT: movq %rax, %r8
; SSE2-NEXT: movq %rax, %r9
; SSE2-NEXT: movq %rax, %r10
; SSE2-NEXT: movq %rax, %r11
; SSE2-NEXT: movq %rax, %r14
; SSE2-NEXT: movq %rax, %r15
; SSE2-NEXT: movq %rax, %r12
; SSE2-NEXT: movq %rax, %r13
; SSE2-NEXT: movq %rax, %rbx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: movq %rax, %rdx
; SSE2-NEXT: movq %rax, %rsi
; SSE2-NEXT: movq %rax, %rdi
; SSE2-NEXT: movq %rax, %rbp
; SSE2-NEXT: shlq $49, %rbp
; SSE2-NEXT: sarq $63, %rbp
; SSE2-NEXT: movd %ebp, %xmm0
; SSE2-NEXT: movq %rax, %rbp
; SSE2-NEXT: movsbq %al, %rax
; SSE2-NEXT: shlq $57, %r8
; SSE2-NEXT: sarq $63, %r8
; SSE2-NEXT: movd %r8d, %xmm1
; SSE2-NEXT: shlq $53, %r9
; SSE2-NEXT: sarq $63, %r9
; SSE2-NEXT: movd %r9d, %xmm2
; SSE2-NEXT: shlq $61, %r10
; SSE2-NEXT: sarq $63, %r10
; SSE2-NEXT: movd %r10d, %xmm3
; SSE2-NEXT: shlq $51, %r11
; SSE2-NEXT: sarq $63, %r11
; SSE2-NEXT: movd %r11d, %xmm4
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-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]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: shlq $59, %r14
; SSE2-NEXT: sarq $63, %r14
; SSE2-NEXT: movd %r14d, %xmm5
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSE2-NEXT: shlq $55, %r15
; SSE2-NEXT: sarq $63, %r15
; SSE2-NEXT: movd %r15d, %xmm2
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; SSE2-NEXT: shlq $63, %r12
; SSE2-NEXT: sarq $63, %r12
; SSE2-NEXT: movd %r12d, %xmm0
; SSE2-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; SSE2-NEXT: shlq $50, %r13
; SSE2-NEXT: sarq $63, %r13
; SSE2-NEXT: movd %r13d, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: shlq $58, %rbx
; SSE2-NEXT: sarq $63, %rbx
; SSE2-NEXT: movd %ebx, %xmm2
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: shlq $54, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm4
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: shlq $62, %rdx
; SSE2-NEXT: sarq $63, %rdx
; SSE2-NEXT: movd %edx, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSE2-NEXT: shlq $52, %rsi
; SSE2-NEXT: sarq $63, %rsi
; SSE2-NEXT: movd %esi, %xmm1
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; SSE2-NEXT: shlq $60, %rdi
; SSE2-NEXT: sarq $63, %rdi
; SSE2-NEXT: movd %edi, %xmm4
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3],xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
; SSE2-NEXT: shrq $15, %rbp
; SSE2-NEXT: movd %ebp, %xmm1
; SSE2-NEXT: shrq $7, %rax
; SSE2-NEXT: movd %eax, %xmm2
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSE2-NEXT: popq %rbx
; SSE2-NEXT: popq %r12
; SSE2-NEXT: popq %r13
; SSE2-NEXT: popq %r14
; SSE2-NEXT: popq %r15
; SSE2-NEXT: popq %rbp
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE2-NEXT: retq
;
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-LABEL: load_sext_16i1_to_16i8:
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: pushq %rbp
; SSSE3-NEXT: pushq %r15
; SSSE3-NEXT: pushq %r14
; SSSE3-NEXT: pushq %r13
; SSSE3-NEXT: pushq %r12
; SSSE3-NEXT: pushq %rbx
; SSSE3-NEXT: movswq (%rdi), %rax
; SSSE3-NEXT: movq %rax, %r8
; SSSE3-NEXT: movq %rax, %r9
; SSSE3-NEXT: movq %rax, %r10
; SSSE3-NEXT: movq %rax, %r11
; SSSE3-NEXT: movq %rax, %r14
; SSSE3-NEXT: movq %rax, %r15
; SSSE3-NEXT: movq %rax, %r12
; SSSE3-NEXT: movq %rax, %r13
; SSSE3-NEXT: movq %rax, %rbx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: movq %rax, %rdx
; SSSE3-NEXT: movq %rax, %rsi
; SSSE3-NEXT: movq %rax, %rdi
; SSSE3-NEXT: movq %rax, %rbp
; SSSE3-NEXT: shlq $49, %rbp
; SSSE3-NEXT: sarq $63, %rbp
; SSSE3-NEXT: movd %ebp, %xmm0
; SSSE3-NEXT: movq %rax, %rbp
; SSSE3-NEXT: movsbq %al, %rax
; SSSE3-NEXT: shlq $57, %r8
; SSSE3-NEXT: sarq $63, %r8
; SSSE3-NEXT: movd %r8d, %xmm1
; SSSE3-NEXT: shlq $53, %r9
; SSSE3-NEXT: sarq $63, %r9
; SSSE3-NEXT: movd %r9d, %xmm2
; SSSE3-NEXT: shlq $61, %r10
; SSSE3-NEXT: sarq $63, %r10
; SSSE3-NEXT: movd %r10d, %xmm3
; SSSE3-NEXT: shlq $51, %r11
; SSSE3-NEXT: sarq $63, %r11
; SSSE3-NEXT: movd %r11d, %xmm4
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-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]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: shlq $59, %r14
; SSSE3-NEXT: sarq $63, %r14
; SSSE3-NEXT: movd %r14d, %xmm5
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSSE3-NEXT: shlq $55, %r15
; SSSE3-NEXT: sarq $63, %r15
; SSSE3-NEXT: movd %r15d, %xmm2
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; SSSE3-NEXT: shlq $63, %r12
; SSSE3-NEXT: sarq $63, %r12
; SSSE3-NEXT: movd %r12d, %xmm0
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; SSSE3-NEXT: shlq $50, %r13
; SSSE3-NEXT: sarq $63, %r13
; SSSE3-NEXT: movd %r13d, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: shlq $58, %rbx
; SSSE3-NEXT: sarq $63, %rbx
; SSSE3-NEXT: movd %ebx, %xmm2
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: shlq $54, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm4
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: shlq $62, %rdx
; SSSE3-NEXT: sarq $63, %rdx
; SSSE3-NEXT: movd %edx, %xmm3
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSSE3-NEXT: shlq $52, %rsi
; SSSE3-NEXT: sarq $63, %rsi
; SSSE3-NEXT: movd %esi, %xmm1
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; SSSE3-NEXT: shlq $60, %rdi
; SSSE3-NEXT: sarq $63, %rdi
; SSSE3-NEXT: movd %edi, %xmm4
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3],xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
; SSSE3-NEXT: shrq $15, %rbp
; SSSE3-NEXT: movd %ebp, %xmm1
; SSSE3-NEXT: shrq $7, %rax
; SSSE3-NEXT: movd %eax, %xmm2
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSSE3-NEXT: popq %rbx
; SSSE3-NEXT: popq %r12
; SSSE3-NEXT: popq %r13
; SSSE3-NEXT: popq %r14
; SSSE3-NEXT: popq %r15
; SSSE3-NEXT: popq %rbp
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSSE3-NEXT: retq
;
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-LABEL: load_sext_16i1_to_16i8:
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: movswq (%rdi), %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $62, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: movq %rax, %rdx
; SSE41-NEXT: shlq $63, %rdx
; SSE41-NEXT: sarq $63, %rdx
; SSE41-NEXT: movd %edx, %xmm0
; SSE41-NEXT: pinsrb $1, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $61, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $2, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $60, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $3, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $59, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $4, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $58, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $5, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $57, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $6, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: movsbq %al, %rcx
; SSE41-NEXT: shrq $7, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: pinsrb $7, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $55, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $8, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $54, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $9, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $53, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $10, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $52, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $11, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $51, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $12, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $50, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $13, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $49, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $14, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: shrq $15, %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: pinsrb $15, %eax, %xmm0
; SSE41-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX1-LABEL: load_sext_16i1_to_16i8:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: movswq (%rdi), %rax
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $62, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: movq %rax, %rdx
; AVX1-NEXT: shlq $63, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vmovd %edx, %xmm0
; AVX1-NEXT: vpinsrb $1, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $61, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $2, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $60, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $3, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $59, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $4, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $58, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $5, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $57, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $6, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movsbq %al, %rcx
; AVX1-NEXT: shrq $7, %rcx
; AVX1-NEXT: vpinsrb $7, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $55, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $8, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $54, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $9, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $53, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $10, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $52, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $11, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $51, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $12, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $50, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $13, %ecx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: shlq $49, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrb $14, %ecx, %xmm0, %xmm0
; AVX1-NEXT: shrq $15, %rax
; AVX1-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_16i1_to_16i8:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: movswq (%rdi), %rax
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $62, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: movq %rax, %rdx
; AVX2-NEXT: shlq $63, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vmovd %edx, %xmm0
; AVX2-NEXT: vpinsrb $1, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $61, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $2, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $60, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $3, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $59, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $4, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $58, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $5, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $57, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $6, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movsbq %al, %rcx
; AVX2-NEXT: shrq $7, %rcx
; AVX2-NEXT: vpinsrb $7, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $55, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $8, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $54, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $9, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $53, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $10, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $52, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $11, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $51, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $12, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $50, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $13, %ecx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: shlq $49, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrb $14, %ecx, %xmm0, %xmm0
; AVX2-NEXT: shrq $15, %rax
; AVX2-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; AVX2-NEXT: retq
;
; AVX512-LABEL: load_sext_16i1_to_16i8:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: kmovw (%rdi), %k1
[AVX512] Use vpternlog with an immediate of 0xff to create 512-bit all one vectors. llvm-svn: 275045
2016-07-11 13:36:48 +08:00
; AVX512-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm0 {%k1} {z}
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-NEXT: vpmovdb %zmm0, %xmm0
; AVX512-NEXT: retq
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
;
; X32-SSE41-LABEL: load_sext_16i1_to_16i8:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: movswl (%eax), %eax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $30, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: movl %eax, %edx
; X32-SSE41-NEXT: shll $31, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: movd %edx, %xmm0
; X32-SSE41-NEXT: pinsrb $1, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $29, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $2, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $28, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $3, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $27, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $4, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $26, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $5, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $25, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $6, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: movsbl %al, %ecx
; X32-SSE41-NEXT: shrl $7, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: pinsrb $7, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $23, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $8, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $22, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $9, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $21, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $10, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $20, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $11, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $19, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $12, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $18, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $13, %ecx, %xmm0
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shll $17, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
; X32-SSE41-NEXT: pinsrb $14, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: shrl $15, %eax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: pinsrb $15, %eax, %xmm0
; X32-SSE41-NEXT: retl
entry:
%X = load <16 x i1>, <16 x i1>* %ptr
%Y = sext <16 x i1> %X to <16 x i8>
ret <16 x i8> %Y
}
define <16 x i16> @load_sext_16i1_to_16i16(<16 x i1> *%ptr) {
; SSE2-LABEL: load_sext_16i1_to_16i16:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movzwl (%rdi), %eax
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $14, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $6, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm1
; SSE2-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]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $10, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $2, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-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]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $12, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $4, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm1
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $8, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-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]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $13, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $5, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-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]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $9, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm3
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm0
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $11, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $3, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSE2-NEXT: movl %eax, %ecx
; SSE2-NEXT: shrl $7, %ecx
; SSE2-NEXT: andl $1, %ecx
; SSE2-NEXT: movd %ecx, %xmm2
; SSE2-NEXT: shrl $15, %eax
; SSE2-NEXT: movzwl %ax, %eax
; SSE2-NEXT: movd %eax, %xmm4
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSE2-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]
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: psllw $15, %xmm0
; SSE2-NEXT: psraw $15, %xmm0
; SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; SSE2-NEXT: psllw $15, %xmm1
; SSE2-NEXT: psraw $15, %xmm1
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_16i1_to_16i16:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movzwl (%rdi), %eax
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $14, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $6, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm1
; SSSE3-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]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $10, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $2, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-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]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $12, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $4, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm3
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm1
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $8, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-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]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $13, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $5, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-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]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $9, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm3
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm0
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $11, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $3, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm3
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSSE3-NEXT: movl %eax, %ecx
; SSSE3-NEXT: shrl $7, %ecx
; SSSE3-NEXT: andl $1, %ecx
; SSSE3-NEXT: movd %ecx, %xmm2
; SSSE3-NEXT: shrl $15, %eax
; SSSE3-NEXT: movzwl %ax, %eax
; SSSE3-NEXT: movd %eax, %xmm4
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSSE3-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]
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: psllw $15, %xmm0
; SSSE3-NEXT: psraw $15, %xmm0
; SSSE3-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; SSSE3-NEXT: psllw $15, %xmm1
; SSSE3-NEXT: psraw $15, %xmm1
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_16i1_to_16i16:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: movzwl (%rdi), %eax
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: movl %eax, %edx
; SSE41-NEXT: andl $1, %edx
; SSE41-NEXT: movd %edx, %xmm1
; SSE41-NEXT: pinsrb $1, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $2, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $2, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $3, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $3, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $4, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $4, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $5, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $5, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $6, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $6, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $7, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $7, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $8, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $9, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $9, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $10, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $10, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $11, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $11, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $12, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $12, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $13, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $13, %ecx, %xmm1
; SSE41-NEXT: movl %eax, %ecx
; SSE41-NEXT: shrl $14, %ecx
; SSE41-NEXT: andl $1, %ecx
; SSE41-NEXT: pinsrb $14, %ecx, %xmm1
; SSE41-NEXT: shrl $15, %eax
; SSE41-NEXT: movzwl %ax, %eax
; SSE41-NEXT: pinsrb $15, %eax, %xmm1
; SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
; SSE41-NEXT: psllw $15, %xmm0
; SSE41-NEXT: psraw $15, %xmm0
; SSE41-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE41-NEXT: psllw $15, %xmm1
; SSE41-NEXT: psraw $15, %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: load_sext_16i1_to_16i16:
; AVX1: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: pushq %rbp
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi0:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_def_cfa_offset 16
; AVX1-NEXT: pushq %r15
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi1:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_def_cfa_offset 24
; AVX1-NEXT: pushq %r14
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi2:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_def_cfa_offset 32
; AVX1-NEXT: pushq %r13
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi3:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_def_cfa_offset 40
; AVX1-NEXT: pushq %r12
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi4:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_def_cfa_offset 48
; AVX1-NEXT: pushq %rbx
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi5:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_def_cfa_offset 56
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi6:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_offset %rbx, -56
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi7:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_offset %r12, -48
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi8:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_offset %r13, -40
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi9:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_offset %r14, -32
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi10:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_offset %r15, -24
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX1-NEXT: .Lcfi11:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: .cfi_offset %rbp, -16
; AVX1-NEXT: movswq (%rdi), %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: shlq $55, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: sarq $63, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: vmovd %ecx, %xmm0
; AVX1-NEXT: movq %rax, %r8
; AVX1-NEXT: movq %rax, %r10
; AVX1-NEXT: movq %rax, %r11
; AVX1-NEXT: movq %rax, %r14
; AVX1-NEXT: movq %rax, %r15
; AVX1-NEXT: movq %rax, %r9
; AVX1-NEXT: movq %rax, %r12
; AVX1-NEXT: movq %rax, %r13
; AVX1-NEXT: movq %rax, %rbx
; AVX1-NEXT: movq %rax, %rdi
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: movq %rax, %rcx
; AVX1-NEXT: movq %rax, %rdx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: movq %rax, %rsi
; AVX1-NEXT: movsbq %al, %rbp
; AVX1-NEXT: shlq $54, %rax
; AVX1-NEXT: sarq $63, %rax
; AVX1-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0
; AVX1-NEXT: shlq $53, %r8
; AVX1-NEXT: sarq $63, %r8
; AVX1-NEXT: vpinsrw $2, %r8d, %xmm0, %xmm0
; AVX1-NEXT: shlq $52, %r10
; AVX1-NEXT: sarq $63, %r10
; AVX1-NEXT: vpinsrw $3, %r10d, %xmm0, %xmm0
; AVX1-NEXT: shlq $51, %r11
; AVX1-NEXT: sarq $63, %r11
; AVX1-NEXT: vpinsrw $4, %r11d, %xmm0, %xmm0
; AVX1-NEXT: shlq $50, %r14
; AVX1-NEXT: sarq $63, %r14
; AVX1-NEXT: vpinsrw $5, %r14d, %xmm0, %xmm0
; AVX1-NEXT: shlq $49, %r15
; AVX1-NEXT: sarq $63, %r15
; AVX1-NEXT: vpinsrw $6, %r15d, %xmm0, %xmm0
; AVX1-NEXT: shrq $15, %r9
; AVX1-NEXT: vpinsrw $7, %r9d, %xmm0, %xmm0
; AVX1-NEXT: shlq $63, %r13
; AVX1-NEXT: sarq $63, %r13
; AVX1-NEXT: vmovd %r13d, %xmm1
; AVX1-NEXT: shlq $62, %r12
; AVX1-NEXT: sarq $63, %r12
; AVX1-NEXT: vpinsrw $1, %r12d, %xmm1, %xmm1
; AVX1-NEXT: shlq $61, %rbx
; AVX1-NEXT: sarq $63, %rbx
; AVX1-NEXT: vpinsrw $2, %ebx, %xmm1, %xmm1
; AVX1-NEXT: shlq $60, %rdi
; AVX1-NEXT: sarq $63, %rdi
; AVX1-NEXT: vpinsrw $3, %edi, %xmm1, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: shlq $59, %rcx
; AVX1-NEXT: sarq $63, %rcx
; AVX1-NEXT: vpinsrw $4, %ecx, %xmm1, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: shlq $58, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vpinsrw $5, %edx, %xmm1, %xmm1
; AVX1-NEXT: shlq $57, %rsi
; AVX1-NEXT: sarq $63, %rsi
; AVX1-NEXT: vpinsrw $6, %esi, %xmm1, %xmm1
; AVX1-NEXT: shrq $7, %rbp
; AVX1-NEXT: vpinsrw $7, %ebp, %xmm1, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: popq %rbx
; AVX1-NEXT: popq %r12
; AVX1-NEXT: popq %r13
; AVX1-NEXT: popq %r14
; AVX1-NEXT: popq %r15
; AVX1-NEXT: popq %rbp
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_16i1_to_16i16:
; AVX2: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: pushq %rbp
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi0:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_def_cfa_offset 16
; AVX2-NEXT: pushq %r15
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi1:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_def_cfa_offset 24
; AVX2-NEXT: pushq %r14
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi2:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_def_cfa_offset 32
; AVX2-NEXT: pushq %r13
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi3:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_def_cfa_offset 40
; AVX2-NEXT: pushq %r12
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi4:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_def_cfa_offset 48
; AVX2-NEXT: pushq %rbx
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi5:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_def_cfa_offset 56
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi6:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_offset %rbx, -56
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi7:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_offset %r12, -48
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi8:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_offset %r13, -40
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi9:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_offset %r14, -32
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi10:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_offset %r15, -24
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; AVX2-NEXT: .Lcfi11:
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: .cfi_offset %rbp, -16
; AVX2-NEXT: movswq (%rdi), %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: shlq $55, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: sarq $63, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: vmovd %ecx, %xmm0
; AVX2-NEXT: movq %rax, %r8
; AVX2-NEXT: movq %rax, %r10
; AVX2-NEXT: movq %rax, %r11
; AVX2-NEXT: movq %rax, %r14
; AVX2-NEXT: movq %rax, %r15
; AVX2-NEXT: movq %rax, %r9
; AVX2-NEXT: movq %rax, %r12
; AVX2-NEXT: movq %rax, %r13
; AVX2-NEXT: movq %rax, %rbx
; AVX2-NEXT: movq %rax, %rdi
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: movq %rax, %rcx
; AVX2-NEXT: movq %rax, %rdx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: movq %rax, %rsi
; AVX2-NEXT: movsbq %al, %rbp
; AVX2-NEXT: shlq $54, %rax
; AVX2-NEXT: sarq $63, %rax
; AVX2-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0
; AVX2-NEXT: shlq $53, %r8
; AVX2-NEXT: sarq $63, %r8
; AVX2-NEXT: vpinsrw $2, %r8d, %xmm0, %xmm0
; AVX2-NEXT: shlq $52, %r10
; AVX2-NEXT: sarq $63, %r10
; AVX2-NEXT: vpinsrw $3, %r10d, %xmm0, %xmm0
; AVX2-NEXT: shlq $51, %r11
; AVX2-NEXT: sarq $63, %r11
; AVX2-NEXT: vpinsrw $4, %r11d, %xmm0, %xmm0
; AVX2-NEXT: shlq $50, %r14
; AVX2-NEXT: sarq $63, %r14
; AVX2-NEXT: vpinsrw $5, %r14d, %xmm0, %xmm0
; AVX2-NEXT: shlq $49, %r15
; AVX2-NEXT: sarq $63, %r15
; AVX2-NEXT: vpinsrw $6, %r15d, %xmm0, %xmm0
; AVX2-NEXT: shrq $15, %r9
; AVX2-NEXT: vpinsrw $7, %r9d, %xmm0, %xmm0
; AVX2-NEXT: shlq $63, %r13
; AVX2-NEXT: sarq $63, %r13
; AVX2-NEXT: vmovd %r13d, %xmm1
; AVX2-NEXT: shlq $62, %r12
; AVX2-NEXT: sarq $63, %r12
; AVX2-NEXT: vpinsrw $1, %r12d, %xmm1, %xmm1
; AVX2-NEXT: shlq $61, %rbx
; AVX2-NEXT: sarq $63, %rbx
; AVX2-NEXT: vpinsrw $2, %ebx, %xmm1, %xmm1
; AVX2-NEXT: shlq $60, %rdi
; AVX2-NEXT: sarq $63, %rdi
; AVX2-NEXT: vpinsrw $3, %edi, %xmm1, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: shlq $59, %rcx
; AVX2-NEXT: sarq $63, %rcx
; AVX2-NEXT: vpinsrw $4, %ecx, %xmm1, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: shlq $58, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vpinsrw $5, %edx, %xmm1, %xmm1
; AVX2-NEXT: shlq $57, %rsi
; AVX2-NEXT: sarq $63, %rsi
; AVX2-NEXT: vpinsrw $6, %esi, %xmm1, %xmm1
; AVX2-NEXT: shrq $7, %rbp
; AVX2-NEXT: vpinsrw $7, %ebp, %xmm1, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: popq %rbx
; AVX2-NEXT: popq %r12
; AVX2-NEXT: popq %r13
; AVX2-NEXT: popq %r14
; AVX2-NEXT: popq %r15
; AVX2-NEXT: popq %rbp
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_16i1_to_16i16:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: kmovw (%rdi), %k1
[AVX512] Use vpternlog with an immediate of 0xff to create 512-bit all one vectors. llvm-svn: 275045
2016-07-11 13:36:48 +08:00
; AVX512-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm0 {%k1} {z}
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-NEXT: vpmovdw %zmm0, %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-LABEL: load_sext_16i1_to_16i16:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: movzwl (%eax), %eax
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: movl %eax, %edx
; X32-SSE41-NEXT: andl $1, %edx
; X32-SSE41-NEXT: movd %edx, %xmm1
; X32-SSE41-NEXT: pinsrb $1, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $2, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $2, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $3, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $3, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $4, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $4, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $5, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $5, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $6, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $6, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $7, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $7, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $8, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $8, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $9, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $9, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $10, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $10, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $11, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $11, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $12, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $12, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $13, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $13, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
; X32-SSE41-NEXT: shrl $14, %ecx
; X32-SSE41-NEXT: andl $1, %ecx
; X32-SSE41-NEXT: pinsrb $14, %ecx, %xmm1
; X32-SSE41-NEXT: shrl $15, %eax
; X32-SSE41-NEXT: pinsrb $15, %eax, %xmm1
; X32-SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
; X32-SSE41-NEXT: psllw $15, %xmm0
; X32-SSE41-NEXT: psraw $15, %xmm0
; X32-SSE41-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; X32-SSE41-NEXT: psllw $15, %xmm1
; X32-SSE41-NEXT: psraw $15, %xmm1
; X32-SSE41-NEXT: retl
entry:
%X = load <16 x i1>, <16 x i1>* %ptr
%Y = sext <16 x i1> %X to <16 x i16>
ret <16 x i16> %Y
}
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
define <32 x i8> @load_sext_32i1_to_32i8(<32 x i1> *%ptr) nounwind readnone {
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-LABEL: load_sext_32i1_to_32i8:
; SSE2: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: pushq %rbp
; SSE2-NEXT: pushq %r15
; SSE2-NEXT: pushq %r14
; SSE2-NEXT: pushq %r13
; SSE2-NEXT: pushq %r12
; SSE2-NEXT: pushq %rbx
; SSE2-NEXT: movswq (%rdi), %rbx
; SSE2-NEXT: movq %rbx, %r10
; SSE2-NEXT: movq %rbx, %r8
; SSE2-NEXT: movq %rbx, %r9
; SSE2-NEXT: movq %rbx, %r11
; SSE2-NEXT: movq %rbx, %r14
; SSE2-NEXT: movq %rbx, %r15
; SSE2-NEXT: movq %rbx, %r12
; SSE2-NEXT: movq %rbx, %r13
; SSE2-NEXT: movq %rbx, %rdx
; SSE2-NEXT: movq %rbx, %rsi
; SSE2-NEXT: movq %rbx, %rcx
; SSE2-NEXT: movq %rbx, %rbp
; SSE2-NEXT: movq %rbx, %rax
; SSE2-NEXT: shlq $49, %rax
; SSE2-NEXT: sarq $63, %rax
; SSE2-NEXT: movd %eax, %xmm0
; SSE2-NEXT: movq %rbx, %rax
; SSE2-NEXT: shlq $57, %r10
; SSE2-NEXT: sarq $63, %r10
; SSE2-NEXT: movd %r10d, %xmm15
; SSE2-NEXT: movq %rbx, %r10
; SSE2-NEXT: movsbq %bl, %rbx
; SSE2-NEXT: punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3],xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
; SSE2-NEXT: shlq $53, %r8
; SSE2-NEXT: sarq $63, %r8
; SSE2-NEXT: movd %r8d, %xmm8
; SSE2-NEXT: shlq $61, %r9
; SSE2-NEXT: sarq $63, %r9
; SSE2-NEXT: movd %r9d, %xmm2
; SSE2-NEXT: shlq $51, %r11
; SSE2-NEXT: sarq $63, %r11
; SSE2-NEXT: movd %r11d, %xmm9
; SSE2-NEXT: shlq $59, %r14
; SSE2-NEXT: sarq $63, %r14
; SSE2-NEXT: movd %r14d, %xmm5
; SSE2-NEXT: shlq $55, %r15
; SSE2-NEXT: sarq $63, %r15
; SSE2-NEXT: movd %r15d, %xmm10
; SSE2-NEXT: shlq $63, %r12
; SSE2-NEXT: sarq $63, %r12
; SSE2-NEXT: movd %r12d, %xmm0
; SSE2-NEXT: shlq $50, %r13
; SSE2-NEXT: sarq $63, %r13
; SSE2-NEXT: movd %r13d, %xmm11
; SSE2-NEXT: shlq $58, %rdx
; SSE2-NEXT: sarq $63, %rdx
; SSE2-NEXT: movd %edx, %xmm4
; SSE2-NEXT: shlq $54, %rsi
; SSE2-NEXT: sarq $63, %rsi
; SSE2-NEXT: movd %esi, %xmm12
; SSE2-NEXT: shlq $62, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: sarq $63, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: movd %ecx, %xmm6
; SSE2-NEXT: shlq $52, %rbp
; SSE2-NEXT: sarq $63, %rbp
; SSE2-NEXT: movd %ebp, %xmm13
; SSE2-NEXT: shlq $60, %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: sarq $63, %rax
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: movd %eax, %xmm7
; SSE2-NEXT: shrq $15, %r10
; SSE2-NEXT: movd %r10d, %xmm14
; SSE2-NEXT: shrq $7, %rbx
; SSE2-NEXT: movd %ebx, %xmm3
; SSE2-NEXT: movswq 2(%rdi), %rdx
; SSE2-NEXT: movq %rdx, %r8
; SSE2-NEXT: movq %rdx, %r9
; SSE2-NEXT: movq %rdx, %r10
; SSE2-NEXT: movq %rdx, %r11
; SSE2-NEXT: movq %rdx, %r14
; SSE2-NEXT: movq %rdx, %r15
; SSE2-NEXT: movq %rdx, %r12
; SSE2-NEXT: movq %rdx, %r13
; SSE2-NEXT: movq %rdx, %rbx
; SSE2-NEXT: movq %rdx, %rax
; SSE2-NEXT: movq %rdx, %rcx
; SSE2-NEXT: movq %rdx, %rsi
; SSE2-NEXT: movq %rdx, %rdi
; SSE2-NEXT: movq %rdx, %rbp
; SSE2-NEXT: shlq $49, %rbp
; SSE2-NEXT: sarq $63, %rbp
; SSE2-NEXT: movd %ebp, %xmm1
; SSE2-NEXT: movq %rdx, %rbp
; SSE2-NEXT: movsbq %dl, %rdx
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3],xmm2[4],xmm8[4],xmm2[5],xmm8[5],xmm2[6],xmm8[6],xmm2[7],xmm8[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3],xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm9[0],xmm5[1],xmm9[1],xmm5[2],xmm9[2],xmm5[3],xmm9[3],xmm5[4],xmm9[4],xmm5[5],xmm9[5],xmm5[6],xmm9[6],xmm5[7],xmm9[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3],xmm0[4],xmm10[4],xmm0[5],xmm10[5],xmm0[6],xmm10[6],xmm0[7],xmm10[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3],xmm4[4],xmm11[4],xmm4[5],xmm11[5],xmm4[6],xmm11[6],xmm4[7],xmm11[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm12[0],xmm6[1],xmm12[1],xmm6[2],xmm12[2],xmm6[3],xmm12[3],xmm6[4],xmm12[4],xmm6[5],xmm12[5],xmm6[6],xmm12[6],xmm6[7],xmm12[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3],xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm13[0],xmm7[1],xmm13[1],xmm7[2],xmm13[2],xmm7[3],xmm13[3],xmm7[4],xmm13[4],xmm7[5],xmm13[5],xmm7[6],xmm13[6],xmm7[7],xmm13[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm14[0],xmm3[1],xmm14[1],xmm3[2],xmm14[2],xmm3[3],xmm14[3],xmm3[4],xmm14[4],xmm3[5],xmm14[5],xmm3[6],xmm14[6],xmm3[7],xmm14[7]
; SSE2-NEXT: shlq $57, %r8
; SSE2-NEXT: sarq $63, %r8
; SSE2-NEXT: movd %r8d, %xmm2
; SSE2-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3],xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
; SSE2-NEXT: shlq $53, %r9
; SSE2-NEXT: sarq $63, %r9
; SSE2-NEXT: movd %r9d, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
; SSE2-NEXT: shlq $61, %r10
; SSE2-NEXT: sarq $63, %r10
; SSE2-NEXT: movd %r10d, %xmm4
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
; SSE2-NEXT: shlq $51, %r11
; SSE2-NEXT: sarq $63, %r11
; SSE2-NEXT: movd %r11d, %xmm5
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: shlq $59, %r14
; SSE2-NEXT: sarq $63, %r14
; SSE2-NEXT: movd %r14d, %xmm6
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; SSE2-NEXT: shlq $55, %r15
; SSE2-NEXT: sarq $63, %r15
; SSE2-NEXT: movd %r15d, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
; SSE2-NEXT: shlq $63, %r12
; SSE2-NEXT: sarq $63, %r12
; SSE2-NEXT: movd %r12d, %xmm1
; SSE2-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; SSE2-NEXT: shlq $50, %r13
; SSE2-NEXT: sarq $63, %r13
; SSE2-NEXT: movd %r13d, %xmm2
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: shlq $58, %rbx
; SSE2-NEXT: sarq $63, %rbx
; SSE2-NEXT: movd %ebx, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSE2-NEXT: shlq $54, %rax
; SSE2-NEXT: sarq $63, %rax
; SSE2-NEXT: movd %eax, %xmm5
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: shlq $62, %rcx
; SSE2-NEXT: sarq $63, %rcx
; SSE2-NEXT: movd %ecx, %xmm4
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSE2-NEXT: shlq $52, %rsi
; SSE2-NEXT: sarq $63, %rsi
; SSE2-NEXT: movd %esi, %xmm2
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; SSE2-NEXT: shlq $60, %rdi
; SSE2-NEXT: sarq $63, %rdi
; SSE2-NEXT: movd %edi, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSE2-NEXT: shrq $15, %rbp
; SSE2-NEXT: movd %ebp, %xmm2
; SSE2-NEXT: shrq $7, %rdx
; SSE2-NEXT: movd %edx, %xmm5
; SSE2-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3],xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE2-NEXT: popq %rbx
; SSE2-NEXT: popq %r12
; SSE2-NEXT: popq %r13
; SSE2-NEXT: popq %r14
; SSE2-NEXT: popq %r15
; SSE2-NEXT: popq %rbp
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_32i1_to_32i8:
; SSSE3: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: pushq %rbp
; SSSE3-NEXT: pushq %r15
; SSSE3-NEXT: pushq %r14
; SSSE3-NEXT: pushq %r13
; SSSE3-NEXT: pushq %r12
; SSSE3-NEXT: pushq %rbx
; SSSE3-NEXT: movswq (%rdi), %rbx
; SSSE3-NEXT: movq %rbx, %r10
; SSSE3-NEXT: movq %rbx, %r8
; SSSE3-NEXT: movq %rbx, %r9
; SSSE3-NEXT: movq %rbx, %r11
; SSSE3-NEXT: movq %rbx, %r14
; SSSE3-NEXT: movq %rbx, %r15
; SSSE3-NEXT: movq %rbx, %r12
; SSSE3-NEXT: movq %rbx, %r13
; SSSE3-NEXT: movq %rbx, %rdx
; SSSE3-NEXT: movq %rbx, %rsi
; SSSE3-NEXT: movq %rbx, %rcx
; SSSE3-NEXT: movq %rbx, %rbp
; SSSE3-NEXT: movq %rbx, %rax
; SSSE3-NEXT: shlq $49, %rax
; SSSE3-NEXT: sarq $63, %rax
; SSSE3-NEXT: movd %eax, %xmm0
; SSSE3-NEXT: movq %rbx, %rax
; SSSE3-NEXT: shlq $57, %r10
; SSSE3-NEXT: sarq $63, %r10
; SSSE3-NEXT: movd %r10d, %xmm15
; SSSE3-NEXT: movq %rbx, %r10
; SSSE3-NEXT: movsbq %bl, %rbx
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3],xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
; SSSE3-NEXT: shlq $53, %r8
; SSSE3-NEXT: sarq $63, %r8
; SSSE3-NEXT: movd %r8d, %xmm8
; SSSE3-NEXT: shlq $61, %r9
; SSSE3-NEXT: sarq $63, %r9
; SSSE3-NEXT: movd %r9d, %xmm2
; SSSE3-NEXT: shlq $51, %r11
; SSSE3-NEXT: sarq $63, %r11
; SSSE3-NEXT: movd %r11d, %xmm9
; SSSE3-NEXT: shlq $59, %r14
; SSSE3-NEXT: sarq $63, %r14
; SSSE3-NEXT: movd %r14d, %xmm5
; SSSE3-NEXT: shlq $55, %r15
; SSSE3-NEXT: sarq $63, %r15
; SSSE3-NEXT: movd %r15d, %xmm10
; SSSE3-NEXT: shlq $63, %r12
; SSSE3-NEXT: sarq $63, %r12
; SSSE3-NEXT: movd %r12d, %xmm0
; SSSE3-NEXT: shlq $50, %r13
; SSSE3-NEXT: sarq $63, %r13
; SSSE3-NEXT: movd %r13d, %xmm11
; SSSE3-NEXT: shlq $58, %rdx
; SSSE3-NEXT: sarq $63, %rdx
; SSSE3-NEXT: movd %edx, %xmm4
; SSSE3-NEXT: shlq $54, %rsi
; SSSE3-NEXT: sarq $63, %rsi
; SSSE3-NEXT: movd %esi, %xmm12
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: shlq $62, %rcx
; SSSE3-NEXT: sarq $63, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: movd %ecx, %xmm6
; SSSE3-NEXT: shlq $52, %rbp
; SSSE3-NEXT: sarq $63, %rbp
; SSSE3-NEXT: movd %ebp, %xmm13
; SSSE3-NEXT: shlq $60, %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: sarq $63, %rax
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: movd %eax, %xmm7
; SSSE3-NEXT: shrq $15, %r10
; SSSE3-NEXT: movd %r10d, %xmm14
; SSSE3-NEXT: shrq $7, %rbx
; SSSE3-NEXT: movd %ebx, %xmm3
; SSSE3-NEXT: movswq 2(%rdi), %rdx
; SSSE3-NEXT: movq %rdx, %r8
; SSSE3-NEXT: movq %rdx, %r9
; SSSE3-NEXT: movq %rdx, %r10
; SSSE3-NEXT: movq %rdx, %r11
; SSSE3-NEXT: movq %rdx, %r14
; SSSE3-NEXT: movq %rdx, %r15
; SSSE3-NEXT: movq %rdx, %r12
; SSSE3-NEXT: movq %rdx, %r13
; SSSE3-NEXT: movq %rdx, %rbx
; SSSE3-NEXT: movq %rdx, %rax
; SSSE3-NEXT: movq %rdx, %rcx
; SSSE3-NEXT: movq %rdx, %rsi
; SSSE3-NEXT: movq %rdx, %rdi
; SSSE3-NEXT: movq %rdx, %rbp
; SSSE3-NEXT: shlq $49, %rbp
; SSSE3-NEXT: sarq $63, %rbp
; SSSE3-NEXT: movd %ebp, %xmm1
; SSSE3-NEXT: movq %rdx, %rbp
; SSSE3-NEXT: movsbq %dl, %rdx
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3],xmm2[4],xmm8[4],xmm2[5],xmm8[5],xmm2[6],xmm8[6],xmm2[7],xmm8[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3],xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm9[0],xmm5[1],xmm9[1],xmm5[2],xmm9[2],xmm5[3],xmm9[3],xmm5[4],xmm9[4],xmm5[5],xmm9[5],xmm5[6],xmm9[6],xmm5[7],xmm9[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3],xmm0[4],xmm10[4],xmm0[5],xmm10[5],xmm0[6],xmm10[6],xmm0[7],xmm10[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3],xmm4[4],xmm11[4],xmm4[5],xmm11[5],xmm4[6],xmm11[6],xmm4[7],xmm11[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm12[0],xmm6[1],xmm12[1],xmm6[2],xmm12[2],xmm6[3],xmm12[3],xmm6[4],xmm12[4],xmm6[5],xmm12[5],xmm6[6],xmm12[6],xmm6[7],xmm12[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3],xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm13[0],xmm7[1],xmm13[1],xmm7[2],xmm13[2],xmm7[3],xmm13[3],xmm7[4],xmm13[4],xmm7[5],xmm13[5],xmm7[6],xmm13[6],xmm7[7],xmm13[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm14[0],xmm3[1],xmm14[1],xmm3[2],xmm14[2],xmm3[3],xmm14[3],xmm3[4],xmm14[4],xmm3[5],xmm14[5],xmm3[6],xmm14[6],xmm3[7],xmm14[7]
; SSSE3-NEXT: shlq $57, %r8
; SSSE3-NEXT: sarq $63, %r8
; SSSE3-NEXT: movd %r8d, %xmm2
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3],xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
; SSSE3-NEXT: shlq $53, %r9
; SSSE3-NEXT: sarq $63, %r9
; SSSE3-NEXT: movd %r9d, %xmm3
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
; SSSE3-NEXT: shlq $61, %r10
; SSSE3-NEXT: sarq $63, %r10
; SSSE3-NEXT: movd %r10d, %xmm4
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
; SSSE3-NEXT: shlq $51, %r11
; SSSE3-NEXT: sarq $63, %r11
; SSSE3-NEXT: movd %r11d, %xmm5
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: shlq $59, %r14
; SSSE3-NEXT: sarq $63, %r14
; SSSE3-NEXT: movd %r14d, %xmm6
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; SSSE3-NEXT: shlq $55, %r15
; SSSE3-NEXT: sarq $63, %r15
; SSSE3-NEXT: movd %r15d, %xmm3
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
; SSSE3-NEXT: shlq $63, %r12
; SSSE3-NEXT: sarq $63, %r12
; SSSE3-NEXT: movd %r12d, %xmm1
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; SSSE3-NEXT: shlq $50, %r13
; SSSE3-NEXT: sarq $63, %r13
; SSSE3-NEXT: movd %r13d, %xmm2
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: shlq $58, %rbx
; SSSE3-NEXT: sarq $63, %rbx
; SSSE3-NEXT: movd %ebx, %xmm3
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSSE3-NEXT: shlq $54, %rax
; SSSE3-NEXT: sarq $63, %rax
; SSSE3-NEXT: movd %eax, %xmm5
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: shlq $62, %rcx
; SSSE3-NEXT: sarq $63, %rcx
; SSSE3-NEXT: movd %ecx, %xmm4
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSSE3-NEXT: shlq $52, %rsi
; SSSE3-NEXT: sarq $63, %rsi
; SSSE3-NEXT: movd %esi, %xmm2
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; SSSE3-NEXT: shlq $60, %rdi
; SSSE3-NEXT: sarq $63, %rdi
; SSSE3-NEXT: movd %edi, %xmm3
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; SSSE3-NEXT: shrq $15, %rbp
; SSSE3-NEXT: movd %ebp, %xmm2
; SSSE3-NEXT: shrq $7, %rdx
; SSSE3-NEXT: movd %edx, %xmm5
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3],xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSSE3-NEXT: popq %rbx
; SSSE3-NEXT: popq %r12
; SSSE3-NEXT: popq %r13
; SSSE3-NEXT: popq %r14
; SSSE3-NEXT: popq %r15
; SSSE3-NEXT: popq %rbp
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_32i1_to_32i8:
; SSE41: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: movswq (%rdi), %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $62, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: movq %rax, %rdx
; SSE41-NEXT: shlq $63, %rdx
; SSE41-NEXT: sarq $63, %rdx
; SSE41-NEXT: movd %edx, %xmm0
; SSE41-NEXT: pinsrb $1, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $61, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $2, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $60, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $3, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $59, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $4, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $58, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $5, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $57, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $6, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: movsbq %al, %rcx
; SSE41-NEXT: shrq $7, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: pinsrb $7, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $55, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $8, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $54, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $9, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $53, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $10, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $52, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $11, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $51, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $12, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $50, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $13, %ecx, %xmm0
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $49, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $14, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: shrq $15, %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: pinsrb $15, %eax, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: movswq 2(%rdi), %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $62, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: movq %rax, %rdx
; SSE41-NEXT: shlq $63, %rdx
; SSE41-NEXT: sarq $63, %rdx
; SSE41-NEXT: movd %edx, %xmm1
; SSE41-NEXT: pinsrb $1, %ecx, %xmm1
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $61, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $2, %ecx, %xmm1
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $60, %rcx
; SSE41-NEXT: sarq $63, %rcx
; SSE41-NEXT: pinsrb $3, %ecx, %xmm1
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $59, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $4, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $58, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $5, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $57, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $6, %ecx, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: movsbq %al, %rcx
; SSE41-NEXT: shrq $7, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $7, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $55, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $8, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $54, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $9, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $53, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $10, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $52, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $11, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $51, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $12, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $50, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $13, %ecx, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; SSE41-NEXT: movq %rax, %rcx
; SSE41-NEXT: shlq $49, %rcx
; SSE41-NEXT: sarq $63, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $14, %ecx, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; SSE41-NEXT: shrq $15, %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; SSE41-NEXT: pinsrb $15, %eax, %xmm1
; SSE41-NEXT: retq
;
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-LABEL: load_sext_32i1_to_32i8:
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: pushq %rbp
; AVX1-NEXT: pushq %r15
; AVX1-NEXT: pushq %r14
; AVX1-NEXT: pushq %r13
; AVX1-NEXT: pushq %r12
; AVX1-NEXT: pushq %rbx
; AVX1-NEXT: movslq (%rdi), %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: shlq $47, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: sarq $63, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: vmovd %ecx, %xmm0
; AVX1-NEXT: movq %rax, %r8
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: movq %rax, %rdx
; AVX1-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: movq %rax, %rdi
; AVX1-NEXT: movq %rax, %r13
; AVX1-NEXT: movq %rax, %rsi
; AVX1-NEXT: movq %rax, %r10
; AVX1-NEXT: movq %rax, %r11
; AVX1-NEXT: movq %rax, %r9
; AVX1-NEXT: movq %rax, %rbx
; AVX1-NEXT: movq %rax, %r14
; AVX1-NEXT: movq %rax, %r15
; AVX1-NEXT: movq %rax, %r12
; AVX1-NEXT: movq %rax, %rbp
; AVX1-NEXT: shlq $46, %rbp
; AVX1-NEXT: sarq $63, %rbp
; AVX1-NEXT: vpinsrb $1, %ebp, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rbp
; AVX1-NEXT: shlq $45, %r8
; AVX1-NEXT: sarq $63, %r8
; AVX1-NEXT: vpinsrb $2, %r8d, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %r8
; AVX1-NEXT: shlq $44, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vpinsrb $3, %edx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rdx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: shlq $43, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: sarq $63, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: vpinsrb $4, %ecx, %xmm0, %xmm0
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: shlq $42, %rdi
; AVX1-NEXT: sarq $63, %rdi
; AVX1-NEXT: vpinsrb $5, %edi, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rdi
; AVX1-NEXT: shlq $41, %r13
; AVX1-NEXT: sarq $63, %r13
; AVX1-NEXT: vpinsrb $6, %r13d, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %r13
; AVX1-NEXT: shlq $40, %rsi
; AVX1-NEXT: sarq $63, %rsi
; AVX1-NEXT: vpinsrb $7, %esi, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rsi
; AVX1-NEXT: shlq $39, %r10
; AVX1-NEXT: sarq $63, %r10
; AVX1-NEXT: vpinsrb $8, %r10d, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %r10
; AVX1-NEXT: shlq $38, %r11
; AVX1-NEXT: sarq $63, %r11
; AVX1-NEXT: vpinsrb $9, %r11d, %xmm0, %xmm0
; AVX1-NEXT: movsbq %al, %r11
; AVX1-NEXT: shlq $37, %r9
; AVX1-NEXT: sarq $63, %r9
; AVX1-NEXT: vpinsrb $10, %r9d, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %r9
; AVX1-NEXT: shlq $36, %rbx
; AVX1-NEXT: sarq $63, %rbx
; AVX1-NEXT: vpinsrb $11, %ebx, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rbx
; AVX1-NEXT: shlq $35, %r14
; AVX1-NEXT: sarq $63, %r14
; AVX1-NEXT: vpinsrb $12, %r14d, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %r14
; AVX1-NEXT: shlq $34, %r15
; AVX1-NEXT: sarq $63, %r15
; AVX1-NEXT: vpinsrb $13, %r15d, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %r15
; AVX1-NEXT: shlq $33, %r12
; AVX1-NEXT: sarq $63, %r12
; AVX1-NEXT: vpinsrb $14, %r12d, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %r12
; AVX1-NEXT: shrq $31, %rbp
; AVX1-NEXT: vpinsrb $15, %ebp, %xmm0, %xmm0
; AVX1-NEXT: movq %rax, %rbp
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: shlq $63, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vmovd %edx, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: movq %rax, %rdx
; AVX1-NEXT: movswq %ax, %rax
; AVX1-NEXT: shlq $62, %r8
; AVX1-NEXT: sarq $63, %r8
; AVX1-NEXT: vpinsrb $1, %r8d, %xmm1, %xmm1
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: shlq $61, %rcx
; AVX1-NEXT: sarq $63, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: vpinsrb $2, %ecx, %xmm1, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: shlq $60, %rdi
; AVX1-NEXT: sarq $63, %rdi
; AVX1-NEXT: vpinsrb $3, %edi, %xmm1, %xmm1
; AVX1-NEXT: shlq $59, %r13
; AVX1-NEXT: sarq $63, %r13
; AVX1-NEXT: vpinsrb $4, %r13d, %xmm1, %xmm1
; AVX1-NEXT: shlq $58, %rsi
; AVX1-NEXT: sarq $63, %rsi
; AVX1-NEXT: vpinsrb $5, %esi, %xmm1, %xmm1
; AVX1-NEXT: shlq $57, %r10
; AVX1-NEXT: sarq $63, %r10
; AVX1-NEXT: vpinsrb $6, %r10d, %xmm1, %xmm1
; AVX1-NEXT: shrq $7, %r11
; AVX1-NEXT: vpinsrb $7, %r11d, %xmm1, %xmm1
; AVX1-NEXT: shlq $55, %r9
; AVX1-NEXT: sarq $63, %r9
; AVX1-NEXT: vpinsrb $8, %r9d, %xmm1, %xmm1
; AVX1-NEXT: shlq $54, %rbx
; AVX1-NEXT: sarq $63, %rbx
; AVX1-NEXT: vpinsrb $9, %ebx, %xmm1, %xmm1
; AVX1-NEXT: shlq $53, %r14
; AVX1-NEXT: sarq $63, %r14
; AVX1-NEXT: vpinsrb $10, %r14d, %xmm1, %xmm1
; AVX1-NEXT: shlq $52, %r15
; AVX1-NEXT: sarq $63, %r15
; AVX1-NEXT: vpinsrb $11, %r15d, %xmm1, %xmm1
; AVX1-NEXT: shlq $51, %r12
; AVX1-NEXT: sarq $63, %r12
; AVX1-NEXT: vpinsrb $12, %r12d, %xmm1, %xmm1
; AVX1-NEXT: shlq $50, %rbp
; AVX1-NEXT: sarq $63, %rbp
; AVX1-NEXT: vpinsrb $13, %ebp, %xmm1, %xmm1
; AVX1-NEXT: shlq $49, %rdx
; AVX1-NEXT: sarq $63, %rdx
; AVX1-NEXT: vpinsrb $14, %edx, %xmm1, %xmm1
; AVX1-NEXT: shrq $15, %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX1-NEXT: vpinsrb $15, %eax, %xmm1, %xmm1
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX1-NEXT: popq %rbx
; AVX1-NEXT: popq %r12
; AVX1-NEXT: popq %r13
; AVX1-NEXT: popq %r14
; AVX1-NEXT: popq %r15
; AVX1-NEXT: popq %rbp
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX1-NEXT: retq
;
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-LABEL: load_sext_32i1_to_32i8:
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2: # BB#0: # %entry
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: pushq %rbp
; AVX2-NEXT: pushq %r15
; AVX2-NEXT: pushq %r14
; AVX2-NEXT: pushq %r13
; AVX2-NEXT: pushq %r12
; AVX2-NEXT: pushq %rbx
; AVX2-NEXT: movslq (%rdi), %rax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: shlq $47, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: sarq $63, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: vmovd %ecx, %xmm0
; AVX2-NEXT: movq %rax, %r8
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: movq %rax, %rdx
; AVX2-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: movq %rax, %rdi
; AVX2-NEXT: movq %rax, %r13
; AVX2-NEXT: movq %rax, %rsi
; AVX2-NEXT: movq %rax, %r10
; AVX2-NEXT: movq %rax, %r11
; AVX2-NEXT: movq %rax, %r9
; AVX2-NEXT: movq %rax, %rbx
; AVX2-NEXT: movq %rax, %r14
; AVX2-NEXT: movq %rax, %r15
; AVX2-NEXT: movq %rax, %r12
; AVX2-NEXT: movq %rax, %rbp
; AVX2-NEXT: shlq $46, %rbp
; AVX2-NEXT: sarq $63, %rbp
; AVX2-NEXT: vpinsrb $1, %ebp, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rbp
; AVX2-NEXT: shlq $45, %r8
; AVX2-NEXT: sarq $63, %r8
; AVX2-NEXT: vpinsrb $2, %r8d, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %r8
; AVX2-NEXT: shlq $44, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vpinsrb $3, %edx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rdx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: shlq $43, %rcx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: sarq $63, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: vpinsrb $4, %ecx, %xmm0, %xmm0
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: movq %rax, %rcx
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: shlq $42, %rdi
; AVX2-NEXT: sarq $63, %rdi
; AVX2-NEXT: vpinsrb $5, %edi, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rdi
; AVX2-NEXT: shlq $41, %r13
; AVX2-NEXT: sarq $63, %r13
; AVX2-NEXT: vpinsrb $6, %r13d, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %r13
; AVX2-NEXT: shlq $40, %rsi
; AVX2-NEXT: sarq $63, %rsi
; AVX2-NEXT: vpinsrb $7, %esi, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rsi
; AVX2-NEXT: shlq $39, %r10
; AVX2-NEXT: sarq $63, %r10
; AVX2-NEXT: vpinsrb $8, %r10d, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %r10
; AVX2-NEXT: shlq $38, %r11
; AVX2-NEXT: sarq $63, %r11
; AVX2-NEXT: vpinsrb $9, %r11d, %xmm0, %xmm0
; AVX2-NEXT: movsbq %al, %r11
; AVX2-NEXT: shlq $37, %r9
; AVX2-NEXT: sarq $63, %r9
; AVX2-NEXT: vpinsrb $10, %r9d, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %r9
; AVX2-NEXT: shlq $36, %rbx
; AVX2-NEXT: sarq $63, %rbx
; AVX2-NEXT: vpinsrb $11, %ebx, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rbx
; AVX2-NEXT: shlq $35, %r14
; AVX2-NEXT: sarq $63, %r14
; AVX2-NEXT: vpinsrb $12, %r14d, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %r14
; AVX2-NEXT: shlq $34, %r15
; AVX2-NEXT: sarq $63, %r15
; AVX2-NEXT: vpinsrb $13, %r15d, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %r15
; AVX2-NEXT: shlq $33, %r12
; AVX2-NEXT: sarq $63, %r12
; AVX2-NEXT: vpinsrb $14, %r12d, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %r12
; AVX2-NEXT: shrq $31, %rbp
; AVX2-NEXT: vpinsrb $15, %ebp, %xmm0, %xmm0
; AVX2-NEXT: movq %rax, %rbp
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: shlq $63, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vmovd %edx, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: movq %rax, %rdx
; AVX2-NEXT: movswq %ax, %rax
; AVX2-NEXT: shlq $62, %r8
; AVX2-NEXT: sarq $63, %r8
; AVX2-NEXT: vpinsrb $1, %r8d, %xmm1, %xmm1
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: shlq $61, %rcx
; AVX2-NEXT: sarq $63, %rcx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: vpinsrb $2, %ecx, %xmm1, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: shlq $60, %rdi
; AVX2-NEXT: sarq $63, %rdi
; AVX2-NEXT: vpinsrb $3, %edi, %xmm1, %xmm1
; AVX2-NEXT: shlq $59, %r13
; AVX2-NEXT: sarq $63, %r13
; AVX2-NEXT: vpinsrb $4, %r13d, %xmm1, %xmm1
; AVX2-NEXT: shlq $58, %rsi
; AVX2-NEXT: sarq $63, %rsi
; AVX2-NEXT: vpinsrb $5, %esi, %xmm1, %xmm1
; AVX2-NEXT: shlq $57, %r10
; AVX2-NEXT: sarq $63, %r10
; AVX2-NEXT: vpinsrb $6, %r10d, %xmm1, %xmm1
; AVX2-NEXT: shrq $7, %r11
; AVX2-NEXT: vpinsrb $7, %r11d, %xmm1, %xmm1
; AVX2-NEXT: shlq $55, %r9
; AVX2-NEXT: sarq $63, %r9
; AVX2-NEXT: vpinsrb $8, %r9d, %xmm1, %xmm1
; AVX2-NEXT: shlq $54, %rbx
; AVX2-NEXT: sarq $63, %rbx
; AVX2-NEXT: vpinsrb $9, %ebx, %xmm1, %xmm1
; AVX2-NEXT: shlq $53, %r14
; AVX2-NEXT: sarq $63, %r14
; AVX2-NEXT: vpinsrb $10, %r14d, %xmm1, %xmm1
; AVX2-NEXT: shlq $52, %r15
; AVX2-NEXT: sarq $63, %r15
; AVX2-NEXT: vpinsrb $11, %r15d, %xmm1, %xmm1
; AVX2-NEXT: shlq $51, %r12
; AVX2-NEXT: sarq $63, %r12
; AVX2-NEXT: vpinsrb $12, %r12d, %xmm1, %xmm1
; AVX2-NEXT: shlq $50, %rbp
; AVX2-NEXT: sarq $63, %rbp
; AVX2-NEXT: vpinsrb $13, %ebp, %xmm1, %xmm1
; AVX2-NEXT: shlq $49, %rdx
; AVX2-NEXT: sarq $63, %rdx
; AVX2-NEXT: vpinsrb $14, %edx, %xmm1, %xmm1
; AVX2-NEXT: shrq $15, %rax
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; AVX2-NEXT: vpinsrb $15, %eax, %xmm1, %xmm1
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; AVX2-NEXT: popq %rbx
; AVX2-NEXT: popq %r12
; AVX2-NEXT: popq %r13
; AVX2-NEXT: popq %r14
; AVX2-NEXT: popq %r15
; AVX2-NEXT: popq %rbp
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_32i1_to_32i8:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: kmovw (%rdi), %k1
[X86] Add AVX512 load opcodes and a couple AVX load opcodes to X86InstrInfo::areLoadsFromSameBasePtr. llvm-svn: 275765
2016-07-18 14:14:43 +08:00
; AVX512-NEXT: kmovw 2(%rdi), %k2
[AVX512] Use vpternlog with an immediate of 0xff to create 512-bit all one vectors. llvm-svn: 275045
2016-07-11 13:36:48 +08:00
; AVX512-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} {z}
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-NEXT: vpmovdb %zmm1, %xmm1
[X86] Add AVX512 load opcodes and a couple AVX load opcodes to X86InstrInfo::areLoadsFromSameBasePtr. llvm-svn: 275765
2016-07-18 14:14:43 +08:00
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm0 {%k2} {z}
[AVX512] Use vpternlog with an immediate of 0xff to create 512-bit all one vectors. llvm-svn: 275045
2016-07-11 13:36:48 +08:00
; AVX512-NEXT: vpmovdb %zmm0, %xmm0
[AVX512] Add ExeDomain to vector extend and truncate instructions. llvm-svn: 276394
2016-07-22 13:46:44 +08:00
; AVX512-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-NEXT: retq
;
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-LABEL: load_sext_32i1_to_32i8:
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41: # BB#0: # %entry
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: pushl %esi
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: movswl (%eax), %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $30, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: movl %ecx, %esi
; X32-SSE41-NEXT: shll $31, %esi
; X32-SSE41-NEXT: sarl $31, %esi
; X32-SSE41-NEXT: movd %esi, %xmm0
; X32-SSE41-NEXT: pinsrb $1, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $29, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $2, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $28, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $3, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $27, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $4, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $26, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $5, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $25, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $6, %edx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: movsbl %cl, %edx
; X32-SSE41-NEXT: shrl $7, %edx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: pinsrb $7, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $23, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $8, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $22, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $9, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $21, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $10, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $20, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $11, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $19, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $12, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $18, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $13, %edx, %xmm0
; X32-SSE41-NEXT: movl %ecx, %edx
; X32-SSE41-NEXT: shll $17, %edx
; X32-SSE41-NEXT: sarl $31, %edx
; X32-SSE41-NEXT: pinsrb $14, %edx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: shrl $15, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: pinsrb $15, %ecx, %xmm0
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: movswl 2(%eax), %eax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $30, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: movl %eax, %edx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $31, %edx
; X32-SSE41-NEXT: sarl $31, %edx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: movd %edx, %xmm1
; X32-SSE41-NEXT: pinsrb $1, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $29, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $2, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $28, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $3, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $27, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $4, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $26, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $5, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $25, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $6, %ecx, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: movsbl %al, %ecx
; X32-SSE41-NEXT: shrl $7, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $7, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $23, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $8, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $22, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $9, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $21, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $10, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $20, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $11, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $19, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $12, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $18, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $13, %ecx, %xmm1
; X32-SSE41-NEXT: movl %eax, %ecx
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: shll $17, %ecx
; X32-SSE41-NEXT: sarl $31, %ecx
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $14, %ecx, %xmm1
[X86] Improve shift combining This folds (ashr (shl a, [56,48,32,24,16]), SarConst) into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or into (lshr, (sext (a), SarConst - [56,48,32,24,16])) depending on sign of (SarConst - [56,48,32,24,16]) sexts in X86 are MOVs. The MOVs have the same code size as above SHIFTs (only SHIFT by 1 has lower code size). However the MOVs have 2 advantages to SHIFTs on x86: 1. MOVs can write to a register that differs from source. 2. MOVs accept memory operands. This fixes PR24373. Patch by: evgeny.v.stupachenko@intel.com Differential Revision: http://reviews.llvm.org/D13161 llvm-svn: 255761
2015-12-16 19:22:37 +08:00
; X32-SSE41-NEXT: shrl $15, %eax
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: pinsrb $15, %eax, %xmm1
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
; X32-SSE41-NEXT: popl %esi
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
; X32-SSE41-NEXT: retl
entry:
[X86][SSE] Added load+sext tests for 16i1->16i8 and 32i1->32i8 llvm-svn: 251661
2015-10-30 06:19:21 +08:00
%X = load <32 x i1>, <32 x i1>* %ptr
%Y = sext <32 x i1> %X to <32 x i8>
ret <32 x i8> %Y
[X86] Added i1 vector sextload tests llvm-svn: 247509
2015-09-12 23:36:41 +08:00
}
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
define <16 x i16> @load_sext_16i8_to_16i16(<16 x i8> *%ptr) {
; SSE2-LABEL: load_sext_16i8_to_16i16:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE2-NEXT: psraw $8, %xmm0
; SSE2-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: psraw $8, %xmm1
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-LABEL: load_sext_16i8_to_16i16:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-NEXT: psraw $8, %xmm0
; SSSE3-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: psraw $8, %xmm1
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-LABEL: load_sext_16i8_to_16i16:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-NEXT: pmovsxbw (%rdi), %xmm0
; SSE41-NEXT: pmovsxbw 8(%rdi), %xmm1
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX1-LABEL: load_sext_16i8_to_16i16:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxbw (%rdi), %xmm0
; AVX1-NEXT: vpmovsxbw 8(%rdi), %xmm1
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX2-LABEL: load_sext_16i8_to_16i16:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxbw (%rdi), %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_16i8_to_16i16:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxbw (%rdi), %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: load_sext_16i8_to_16i16:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0: # %entry
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-NEXT: pmovsxbw (%eax), %xmm0
; X32-SSE41-NEXT: pmovsxbw 8(%eax), %xmm1
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: retl
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
entry:
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
%X = load <16 x i8>, <16 x i8>* %ptr
%Y = sext <16 x i8> %X to <16 x i16>
ret <16 x i16> %Y
Optimized load + SIGN_EXTEND patterns in the X86 backend. llvm-svn: 170506
2012-12-19 15:50:20 +08:00
}
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
define <2 x i64> @load_sext_2i16_to_2i64(<2 x i16> *%ptr) {
; SSE2-LABEL: load_sext_2i16_to_2i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0: # %entry
[X86][SSE] Tidied up vector extend/truncation tests. NFCI. llvm-svn: 241995
2015-07-13 01:40:49 +08:00
; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: psrad $16, %xmm0
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSSE3-LABEL: load_sext_2i16_to_2i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0: # %entry
[X86][SSE] Tidied up vector extend/truncation tests. NFCI. llvm-svn: 241995
2015-07-13 01:40:49 +08:00
; SSSE3-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: psrad $16, %xmm0
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSE41-LABEL: load_sext_2i16_to_2i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0: # %entry
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: pmovsxwq (%rdi), %xmm0
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX-LABEL: load_sext_2i16_to_2i64:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxwq (%rdi), %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: load_sext_2i16_to_2i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxwq (%eax), %xmm0
; X32-SSE41-NEXT: retl
entry:
%X = load <2 x i16>, <2 x i16>* %ptr
%Y = sext <2 x i16> %X to <2 x i64>
ret <2 x i64> %Y
}
define <4 x i32> @load_sext_4i16_to_4i32(<4 x i16> *%ptr) {
; SSE2-LABEL: load_sext_4i16_to_4i32:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: psrad $16, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_4i16_to_4i32:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: psrad $16, %xmm0
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_4i16_to_4i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxwd (%rdi), %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: load_sext_4i16_to_4i32:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpmovsxwd (%rdi), %xmm0
; AVX-NEXT: retq
;
; X32-SSE41-LABEL: load_sext_4i16_to_4i32:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxwd (%eax), %xmm0
; X32-SSE41-NEXT: retl
entry:
%X = load <4 x i16>, <4 x i16>* %ptr
%Y = sext <4 x i16> %X to <4 x i32>
ret <4 x i32> %Y
}
define <4 x i64> @load_sext_4i16_to_4i64(<4 x i16> *%ptr) {
; SSE2-LABEL: load_sext_4i16_to_4i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movswq 2(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm1
; SSE2-NEXT: movswq (%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm0
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE2-NEXT: movswq 6(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm2
; SSE2-NEXT: movswq 4(%rdi), %rax
; SSE2-NEXT: movd %rax, %xmm1
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_4i16_to_4i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movswq 2(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm1
; SSSE3-NEXT: movswq (%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm0
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSSE3-NEXT: movswq 6(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm2
; SSSE3-NEXT: movswq 4(%rdi), %rax
; SSSE3-NEXT: movd %rax, %xmm1
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_4i16_to_4i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxwq (%rdi), %xmm0
; SSE41-NEXT: pmovsxwq 4(%rdi), %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: load_sext_4i16_to_4i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxwd (%rdi), %xmm0
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_4i16_to_4i64:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxwq (%rdi), %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_4i16_to_4i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxwq (%rdi), %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: load_sext_4i16_to_4i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxwq (%eax), %xmm0
; X32-SSE41-NEXT: pmovsxwq 4(%eax), %xmm1
; X32-SSE41-NEXT: retl
entry:
%X = load <4 x i16>, <4 x i16>* %ptr
%Y = sext <4 x i16> %X to <4 x i64>
ret <4 x i64> %Y
}
define <8 x i32> @load_sext_8i16_to_8i32(<8 x i16> *%ptr) {
; SSE2-LABEL: load_sext_8i16_to_8i32:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSE2-NEXT: psrad $16, %xmm0
; SSE2-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
; SSE2-NEXT: psrad $16, %xmm1
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_8i16_to_8i32:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: psrad $16, %xmm0
; SSSE3-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
; SSSE3-NEXT: psrad $16, %xmm1
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_8i16_to_8i32:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxwd (%rdi), %xmm0
; SSE41-NEXT: pmovsxwd 8(%rdi), %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: load_sext_8i16_to_8i32:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxwd (%rdi), %xmm0
; AVX1-NEXT: vpmovsxwd 8(%rdi), %xmm1
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_8i16_to_8i32:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxwd (%rdi), %ymm0
; AVX2-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_8i16_to_8i32:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxwd (%rdi), %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: load_sext_8i16_to_8i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0: # %entry
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-NEXT: pmovsxwd (%eax), %xmm0
; X32-SSE41-NEXT: pmovsxwd 8(%eax), %xmm1
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: retl
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
entry:
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
%X = load <8 x i16>, <8 x i16>* %ptr
%Y = sext <8 x i16> %X to <8 x i32>
ret <8 x i32> %Y
Optimized load + SIGN_EXTEND patterns in the X86 backend. llvm-svn: 170506
2012-12-19 15:50:20 +08:00
}
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
define <2 x i64> @load_sext_2i32_to_2i64(<2 x i32> *%ptr) {
; SSE2-LABEL: load_sext_2i32_to_2i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0: # %entry
[X86][SSE] Tidied up vector extend/truncation tests. NFCI. llvm-svn: 241995
2015-07-13 01:40:49 +08:00
; SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrad $31, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSSE3-LABEL: load_sext_2i32_to_2i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0: # %entry
[X86][SSE] Tidied up vector extend/truncation tests. NFCI. llvm-svn: 241995
2015-07-13 01:40:49 +08:00
; SSSE3-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: psrad $31, %xmm1
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; SSE41-LABEL: load_sext_2i32_to_2i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0: # %entry
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: pmovsxdq (%rdi), %xmm0
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; AVX-LABEL: load_sext_2i32_to_2i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX: # BB#0: # %entry
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX-NEXT: vpmovsxdq (%rdi), %xmm0
; AVX-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][SSE] Added additional vector sign/zero extension tests. llvm-svn: 243212
2015-07-25 19:17:35 +08:00
; X32-SSE41-LABEL: load_sext_2i32_to_2i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0: # %entry
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxdq (%eax), %xmm0
; X32-SSE41-NEXT: retl
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
entry:
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%X = load <2 x i32>, <2 x i32>* %ptr
Optimized load + SIGN_EXTEND patterns in the X86 backend. llvm-svn: 170506
2012-12-19 15:50:20 +08:00
%Y = sext <2 x i32> %X to <2 x i64>
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
ret <2 x i64> %Y
Optimized load + SIGN_EXTEND patterns in the X86 backend. llvm-svn: 170506
2012-12-19 15:50:20 +08:00
}
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
define <4 x i64> @load_sext_4i32_to_4i64(<4 x i32> *%ptr) {
; SSE2-LABEL: load_sext_4i32_to_4i64:
; SSE2: # BB#0: # %entry
; SSE2-NEXT: movdqa (%rdi), %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: load_sext_4i32_to_4i64:
; SSSE3: # BB#0: # %entry
; SSSE3-NEXT: movdqa (%rdi), %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: load_sext_4i32_to_4i64:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: pmovsxdq (%rdi), %xmm0
; SSE41-NEXT: pmovsxdq 8(%rdi), %xmm1
; SSE41-NEXT: retq
;
; AVX1-LABEL: load_sext_4i32_to_4i64:
; AVX1: # BB#0: # %entry
; AVX1-NEXT: vpmovsxdq (%rdi), %xmm0
; AVX1-NEXT: vpmovsxdq 8(%rdi), %xmm1
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: load_sext_4i32_to_4i64:
; AVX2: # BB#0: # %entry
; AVX2-NEXT: vpmovsxdq (%rdi), %ymm0
; AVX2-NEXT: retq
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: load_sext_4i32_to_4i64:
; AVX512: # BB#0: # %entry
; AVX512-NEXT: vpmovsxdq (%rdi), %ymm0
; AVX512-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: load_sext_4i32_to_4i64:
; X32-SSE41: # BB#0: # %entry
; X32-SSE41-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-SSE41-NEXT: pmovsxdq (%eax), %xmm0
; X32-SSE41-NEXT: pmovsxdq 8(%eax), %xmm1
; X32-SSE41-NEXT: retl
entry:
%X = load <4 x i32>, <4 x i32>* %ptr
%Y = sext <4 x i32> %X to <4 x i64>
ret <4 x i64> %Y
}
define i32 @sext_2i8_to_i32(<16 x i8> %A) nounwind uwtable readnone ssp {
; SSE2-LABEL: sext_2i8_to_i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0: # %entry
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: psraw $8, %xmm0
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE2-NEXT: movd %xmm0, %eax
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-LABEL: sext_2i8_to_i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0: # %entry
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSSE3-NEXT: psraw $8, %xmm0
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSSE3-NEXT: movd %xmm0, %eax
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-LABEL: sext_2i8_to_i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
; SSE41-NEXT: movd %xmm0, %eax
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX-LABEL: sext_2i8_to_i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; AVX-NEXT: vpmovsxbw %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-LABEL: sext_2i8_to_i32:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0: # %entry
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-NEXT: pushl %eax
MCStreamer: Use "cfi" for CFI related temp labels. Choosing a "cfi" name makes the intend a bit clearer in an assembly dump and more importantly the assembly dumps are slightly more stable as the numbers don't move around anymore when unrelated code calls createTempSymbol() more or less often. As they are temp labels the name doesn't influence the generated object code. Differential Revision: https://reviews.llvm.org/D27244 llvm-svn: 288290
2016-12-01 07:48:26 +08:00
; X32-SSE41-NEXT: .Lcfi0:
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
; X32-SSE41-NEXT: .cfi_def_cfa_offset 8
; X32-SSE41-NEXT: pmovsxbw %xmm0, %xmm0
; X32-SSE41-NEXT: movd %xmm0, %eax
findDeadCallerSavedReg needs to pay attention to calling convention Caller saved regs differ between SysV and Win64. Use the tail call available set to scavenge from. Refactor register info to create new helper to get at tail call GPRs. Added a new test case for windows. Fixed up a number of X64 tests since now RCX is preferred over RDX on SysV. Differential Revision: http://reviews.llvm.org/D14878 llvm-svn: 253927
2015-11-24 06:17:44 +08:00
; X32-SSE41-NEXT: popl %ecx
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: retl
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
entry:
[X86][SSE] Added additional vector sign/zero load extension tests. llvm-svn: 243216
2015-07-25 22:07:20 +08:00
%Shuf = shufflevector <16 x i8> %A, <16 x i8> undef, <2 x i32> <i32 0, i32 1>
%Ex = sext <2 x i8> %Shuf to <2 x i16>
%Bc = bitcast <2 x i16> %Ex to i32
ret i32 %Bc
Optimized load + SIGN_EXTEND patterns in the X86 backend. llvm-svn: 170506
2012-12-19 15:50:20 +08:00
}
I optimized the following patterns: sext <4 x i1> to <4 x i64> sext <4 x i8> to <4 x i64> sext <4 x i16> to <4 x i64> I'm running Combine on SIGN_EXTEND_IN_REG and revert SEXT patterns: (sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) -> (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT))) The sext_in_reg (v4i32 x) may be lowered to shl+sar operations. The "sar" does not exist on 64-bit operation, so lowering sext_in_reg (v4i64 x) has no vector solution. I also added a cost of this operations to the AVX costs table. llvm-svn: 175619
2013-02-20 20:42:54 +08:00
define <4 x i64> @sext_4i1_to_4i64(<4 x i1> %mask) {
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-LABEL: sext_4i1_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0:
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: pslld $31, %xmm0
; SSE2-NEXT: psrad $31, %xmm0
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_4i1_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0:
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: pslld $31, %xmm0
; SSSE3-NEXT: psrad $31, %xmm0
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_4i1_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0:
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: pslld $31, %xmm0
; SSE41-NEXT: psrad $31, %xmm0
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE41-NEXT: pmovsxdq %xmm0, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxdq %xmm0, %xmm1
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
; AVX1-LABEL: sext_4i1_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX1: # BB#0:
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX1-NEXT: vpslld $31, %xmm0, %xmm0
; AVX1-NEXT: vpsrad $31, %xmm0, %xmm0
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX2-LABEL: sext_4i1_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX2: # BB#0:
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX2-NEXT: vpslld $31, %xmm0, %xmm0
; AVX2-NEXT: vpsrad $31, %xmm0, %xmm0
; AVX2-NEXT: vpmovsxdq %xmm0, %ymm0
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; AVX2-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_4i1_to_4i64:
; AVX512: # BB#0:
; AVX512-NEXT: vpslld $31, %xmm0, %xmm0
; AVX512-NEXT: vpsrad $31, %xmm0, %xmm0
; AVX512-NEXT: vpmovsxdq %xmm0, %ymm0
; AVX512-NEXT: retq
;
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-LABEL: sext_4i1_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0:
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: pslld $31, %xmm0
; X32-SSE41-NEXT: psrad $31, %xmm0
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm1
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: movdqa %xmm2, %xmm0
; X32-SSE41-NEXT: retl
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
%extmask = sext <4 x i1> %mask to <4 x i64>
ret <4 x i64> %extmask
}
define <4 x i64> @sext_4i8_to_4i64(<4 x i8> %mask) {
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-LABEL: sext_4i8_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE2: # BB#0:
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: pslld $24, %xmm0
; SSE2-NEXT: psrad $24, %xmm0
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: psrad $31, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: sext_4i8_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSSE3: # BB#0:
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: pslld $24, %xmm0
; SSSE3-NEXT: psrad $24, %xmm0
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: psrad $31, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: sext_4i8_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; SSE41: # BB#0:
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: pslld $24, %xmm0
; SSE41-NEXT: psrad $24, %xmm0
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; SSE41-NEXT: pmovsxdq %xmm0, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE41-NEXT: pmovsxdq %xmm0, %xmm1
[x86] Teach both sext and zext vector tests to cover a nice wide range of architectures: SSE2, SSSE3, SSE4.1, AVX, and AVX2. Unfortunately, this exposses the absolute horror of the code we generate for many of these patterns. Anyone wanting to familiarize themselves with the x86 backend and improve performance could do a lot of good sitting down and making these test cases not look so terrible. While the new vector shuffle code I'm working on well help some, it won't fix all of the crimes here. llvm-svn: 218807
2014-10-02 04:41:36 +08:00
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: retq
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
;
; AVX1-LABEL: sext_4i8_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX1: # BB#0:
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX1-NEXT: vpslld $24, %xmm0, %xmm0
; AVX1-NEXT: vpsrad $24, %xmm0, %xmm0
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX2-LABEL: sext_4i8_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; AVX2: # BB#0:
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
; AVX2-NEXT: vpslld $24, %xmm0, %xmm0
; AVX2-NEXT: vpsrad $24, %xmm0, %xmm0
; AVX2-NEXT: vpmovsxdq %xmm0, %ymm0
[x86] Clean up and generate detailed FileCheck assertions for avx-sext.ll using my new script. Also add an AVX2 mode to this test. Part of cleaning up the test suite before enabling the new vector shuffle lowering. This also highlights some of the abysmal failures of the old shuffle lowering. Check out those 'pinsrw' and 'pextrw' sequences! llvm-svn: 218794
2014-10-02 04:19:32 +08:00
; AVX2-NEXT: retq
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
;
[X86][AVX512] Added AVX512F vector sign extend tests Now that Elena has confirmed that PR26474 has been fixed llvm-svn: 273560
2016-06-23 22:01:45 +08:00
; AVX512-LABEL: sext_4i8_to_4i64:
; AVX512: # BB#0:
; AVX512-NEXT: vpslld $24, %xmm0, %xmm0
; AVX512-NEXT: vpsrad $24, %xmm0, %xmm0
; AVX512-NEXT: vpmovsxdq %xmm0, %ymm0
; AVX512-NEXT: retq
;
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-LABEL: sext_4i8_to_4i64:
[x86] Switch some of the new consolidated vector tests to use a bare-metal triple and have nice BB labels, etc. No significant change here, just tidying up to have a consistent set of OS-agnostic vector functionality here. llvm-svn: 218854
2014-10-02 14:52:19 +08:00
; X32-SSE41: # BB#0:
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: pslld $24, %xmm0
; X32-SSE41-NEXT: psrad $24, %xmm0
[X86][SSE] Improve support for 128-bit vector sign extension This patch improves support for sign extension of the lower lanes of vectors of integers by making use of the SSE41 pmovsx* sign extension instructions where possible, and optimizing the sign extension by shifts on pre-SSE41 targets (avoiding the use of i64 arithmetic shifts which require scalarization). It converts SIGN_EXTEND nodes to SIGN_EXTEND_VECTOR_INREG where necessary, that more closely matches the pmovsx* instruction than the default approach of using SIGN_EXTEND_INREG which splits the operation (into an ANY_EXTEND lowered to a shuffle followed by shifts) making instruction matching difficult during lowering. Necessary support for SIGN_EXTEND_VECTOR_INREG has been added to the DAGCombiner. Differential Revision: http://reviews.llvm.org/D9848 llvm-svn: 237885
2015-05-21 18:05:03 +08:00
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm2
; X32-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; X32-SSE41-NEXT: pmovsxdq %xmm0, %xmm1
[x86] Add a 32-bit run to the sext test, and remove a sad vec_sext.ll test file. This old test had a bunch of functions that were never even checked. =/ The only thing it really did was to make sure that we did something reasonable in 32-bit mode with SSE4.1. Adding another run line to the main vector-sext.ll test seems a better way to do that. llvm-svn: 218810
2014-10-02 04:49:54 +08:00
; X32-SSE41-NEXT: movdqa %xmm2, %xmm0
; X32-SSE41-NEXT: retl
[x86] Sort the ISA-specific RUN lines for vector-sext.ll to go from oldest to newest. This makes more sense to me and is more consistent with other tests. llvm-svn: 218802
2014-10-02 04:32:44 +08:00
%extmask = sext <4 x i8> %mask to <4 x i64>
ret <4 x i64> %extmask
}
[X86][AVX512] Fix sext v32i1 -> v32i8 lowering. Fix PR30600. Differential Revision: https://reviews.llvm.org/D25554 llvm-svn: 284134
2016-10-14 01:20:38 +08:00
define <32 x i8> @sext_32xi1_to_32xi8(<32 x i16> %c1, <32 x i16> %c2)nounwind {
; SSE-LABEL: sext_32xi1_to_32xi8:
; SSE: # BB#0:
; SSE-NEXT: pcmpeqw %xmm5, %xmm1
; SSE-NEXT: pcmpeqw %xmm4, %xmm0
; SSE-NEXT: packsswb %xmm1, %xmm0
; SSE-NEXT: pcmpeqw %xmm7, %xmm3
; SSE-NEXT: pcmpeqw %xmm6, %xmm2
; SSE-NEXT: packsswb %xmm3, %xmm2
; SSE-NEXT: movdqa %xmm2, %xmm1
; SSE-NEXT: retq
;
; AVX1-LABEL: sext_32xi1_to_32xi8:
; AVX1: # BB#0:
; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm4
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm5
; AVX1-NEXT: vpcmpeqw %xmm4, %xmm5, %xmm4
; AVX1-NEXT: vpcmpeqw %xmm3, %xmm1, %xmm1
; AVX1-NEXT: vpacksswb %xmm4, %xmm1, %xmm1
; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4
; AVX1-NEXT: vpcmpeqw %xmm3, %xmm4, %xmm3
; AVX1-NEXT: vpcmpeqw %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpacksswb %xmm3, %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: sext_32xi1_to_32xi8:
; AVX2: # BB#0:
; AVX2-NEXT: vpcmpeqw %ymm3, %ymm1, %ymm1
; AVX2-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
; AVX2-NEXT: vpacksswb %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-NEXT: retq
;
; AVX512F-LABEL: sext_32xi1_to_32xi8:
; AVX512F: # BB#0:
; AVX512F-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0
; AVX512F-NEXT: vpmovdb %zmm0, %xmm0
; AVX512F-NEXT: vpcmpeqw %ymm3, %ymm1, %ymm1
; AVX512F-NEXT: vpmovsxwd %ymm1, %zmm1
; AVX512F-NEXT: vpmovdb %zmm1, %xmm1
; AVX512F-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: sext_32xi1_to_32xi8:
; AVX512BW: # BB#0:
; AVX512BW-NEXT: vpcmpeqw %zmm1, %zmm0, %k1
; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqu16 %zmm0, %zmm0 {%k1} {z}
; AVX512BW-NEXT: vpmovwb %zmm0, %ymm0
; AVX512BW-NEXT: retq
;
; X32-SSE41-LABEL: sext_32xi1_to_32xi8:
; X32-SSE41: # BB#0:
; X32-SSE41-NEXT: pushl %ebp
; X32-SSE41-NEXT: movl %esp, %ebp
; X32-SSE41-NEXT: andl $-16, %esp
; X32-SSE41-NEXT: subl $16, %esp
; X32-SSE41-NEXT: movdqa 8(%ebp), %xmm3
; X32-SSE41-NEXT: pcmpeqw 40(%ebp), %xmm1
; X32-SSE41-NEXT: pcmpeqw 24(%ebp), %xmm0
; X32-SSE41-NEXT: packsswb %xmm1, %xmm0
; X32-SSE41-NEXT: pcmpeqw 72(%ebp), %xmm3
; X32-SSE41-NEXT: pcmpeqw 56(%ebp), %xmm2
; X32-SSE41-NEXT: packsswb %xmm3, %xmm2
; X32-SSE41-NEXT: movdqa %xmm2, %xmm1
; X32-SSE41-NEXT: movl %ebp, %esp
; X32-SSE41-NEXT: popl %ebp
; X32-SSE41-NEXT: retl
%a = icmp eq <32 x i16> %c1, %c2
%b = sext <32 x i1> %a to <32 x i8>
ret <32 x i8> %b
}