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-trunc-packus.ll

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

4635 lines
184 KiB
LLVM
Raw Normal View History

[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=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-prefixes=AVX,AVX2,AVX2-SLOW
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX2-FAST
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512F
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512VL
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512BW
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512BWVL
;
; PACKUS saturation truncation to vXi32
;
define <4 x i32> @trunc_packus_v4i64_v4i32(<4 x i64> %a0) {
; SSE2-LABEL: trunc_packus_v4i64_v4i32:
; SSE2: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [4294967295,4294967295]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm3
; SSE2-NEXT: pxor %xmm2, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [2147483647,2147483647]
; SSE2-NEXT: movdqa %xmm5, %xmm6
; SSE2-NEXT: pcmpgtd %xmm3, %xmm6
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm5, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
; SSE2-NEXT: pand %xmm7, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
; SSE2-NEXT: por %xmm4, %xmm3
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pand %xmm3, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pandn %xmm8, %xmm3
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: por %xmm0, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: pxor %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm5, %xmm4
; SSE2-NEXT: pcmpgtd %xmm0, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm5, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm4
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pandn %xmm8, %xmm4
; SSE2-NEXT: por %xmm1, %xmm4
; SSE2-NEXT: movdqa %xmm4, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pxor %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pcmpgtd %xmm2, %xmm1
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pcmpeqd %xmm2, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pand %xmm5, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: movdqa %xmm3, %xmm0
; SSE2-NEXT: pxor %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm4
; SSE2-NEXT: pcmpgtd %xmm2, %xmm4
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm2, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pand %xmm5, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm2, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pand %xmm3, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v4i64_v4i32:
; SSSE3: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [4294967295,4294967295]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: pxor %xmm2, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm5 = [2147483647,2147483647]
; SSSE3-NEXT: movdqa %xmm5, %xmm6
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm6
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
; SSSE3-NEXT: pand %xmm7, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm3
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pand %xmm3, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pandn %xmm8, %xmm3
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: por %xmm0, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm5, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm4
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: pandn %xmm8, %xmm4
; SSSE3-NEXT: por %xmm1, %xmm4
; SSSE3-NEXT: movdqa %xmm4, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pand %xmm5, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: por %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pand %xmm5, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm2, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pand %xmm3, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v4i64_v4i32:
; SSE41: # %bb.0:
; SSE41-NEXT: movdqa %xmm0, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movapd {{.*#+}} xmm4 = [4294967295,4294967295]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
; SSE41-NEXT: pxor %xmm3, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm6 = [2147483647,2147483647]
; SSE41-NEXT: movdqa %xmm6, %xmm5
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm6, %xmm7
; SSE41-NEXT: pcmpgtd %xmm0, %xmm7
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm7, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm5
; SSE41-NEXT: movdqa %xmm1, %xmm0
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pxor %xmm3, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm6, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm2
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm4
; SSE41-NEXT: xorpd %xmm1, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm4, %xmm2
; SSE41-NEXT: xorpd %xmm3, %xmm2
; SSE41-NEXT: movapd %xmm2, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pcmpgtd %xmm3, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm2, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pxor %xmm2, %xmm2
; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm5, %xmm4
; SSE41-NEXT: xorpd %xmm3, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pcmpgtd %xmm3, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm1
; SSE41-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm2[0,2]
; SSE41-NEXT: movaps %xmm1, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v4i64_v4i32:
; AVX1: # %bb.0:
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [4294967295,4294967295]
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm2, %xmm3
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm2, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm2, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm5
; AVX1-NEXT: vblendvpd %xmm3, %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpand %xmm0, %xmm5, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-SLOW-LABEL: trunc_packus_v4i64_v4i32:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm1 = [4294967295,4294967295,4294967295,4294967295]
; AVX2-SLOW-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-SLOW-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-SLOW-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1
; AVX2-SLOW-NEXT: vpand %ymm0, %ymm1, %ymm0
[x86] split more v8f32/v8i32 shuffles in lowering Similar to D57867 - this is a small patch with lots of test diffs. With half-vector-width narrowing potential, using an extract + 128-bit vshufps is a win because it replaces a 256-bit shuffle with a 128-bit shufle. This seems like it should be a win even for targets with 'fast-variable-shuffle', but we are intentionally deferring that to an independent change to make sure that is true. Differential Revision: https://reviews.llvm.org/D58181 llvm-svn: 354279
2019-02-19 00:46:12 +08:00
; AVX2-SLOW-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-SLOW-NEXT: vzeroupper
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: trunc_packus_v4i64_v4i32:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm1 = [4294967295,4294967295,4294967295,4294967295]
; AVX2-FAST-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-FAST-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-FAST-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1
; AVX2-FAST-NEXT: vpand %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [0,2,4,6,4,6,6,7]
; AVX2-FAST-NEXT: vpermd %ymm0, %ymm1, %ymm0
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; AVX2-FAST-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-FAST-NEXT: vzeroupper
; AVX2-FAST-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v4i64_v4i32:
; AVX512F: # %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
[X86] Add broadcast load folding patterns to NoVLX VPMULLQ/VPMAXSQ/VPMAXUQ/VPMINSQ/VPMINUQ patterns. More fixes for PR36191. llvm-svn: 373560
2019-10-03 11:16:27 +08:00
; AVX512F-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512F-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpmovqd %zmm0, %ymm0
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v4i64_v4i32:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512VL-NEXT: vpmovusqd %ymm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v4i64_v4i32:
; AVX512BW: # %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
[X86] Add broadcast load folding patterns to NoVLX VPMULLQ/VPMAXSQ/VPMAXUQ/VPMINSQ/VPMINUQ patterns. More fixes for PR36191. llvm-svn: 373560
2019-10-03 11:16:27 +08:00
; AVX512BW-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BW-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: vpmovqd %zmm0, %ymm0
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v4i64_v4i32:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512BWVL-NEXT: vpmovusqd %ymm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
%1 = icmp slt <4 x i64> %a0, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
%2 = select <4 x i1> %1, <4 x i64> %a0, <4 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
%3 = icmp sgt <4 x i64> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i64> %2, <4 x i64> zeroinitializer
%5 = trunc <4 x i64> %4 to <4 x i32>
ret <4 x i32> %5
}
define <8 x i32> @trunc_packus_v8i64_v8i32(<8 x i64> %a0) {
; SSE2-LABEL: trunc_packus_v8i64_v8i32:
; SSE2: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [4294967295,4294967295]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm5
; SSE2-NEXT: pxor %xmm10, %xmm5
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147483647,2147483647]
; SSE2-NEXT: movdqa %xmm9, %xmm6
; SSE2-NEXT: pcmpgtd %xmm5, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
; SSE2-NEXT: pand %xmm7, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: por %xmm4, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm5, %xmm0
; SSE2-NEXT: pandn %xmm8, %xmm5
; SSE2-NEXT: por %xmm0, %xmm5
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
; SSE2-NEXT: movdqa %xmm9, %xmm4
; SSE2-NEXT: pcmpgtd %xmm0, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm0
; SSE2-NEXT: pand %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm8, %xmm0
; SSE2-NEXT: por %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm9, %xmm4
; SSE2-NEXT: pcmpgtd %xmm1, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pand %xmm6, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: por %xmm1, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm6, %xmm2
; SSE2-NEXT: pandn %xmm8, %xmm6
; SSE2-NEXT: por %xmm2, %xmm6
; SSE2-NEXT: movdqa %xmm3, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm9, %xmm2
; SSE2-NEXT: pcmpgtd %xmm1, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm4, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm2, %xmm3
; SSE2-NEXT: pandn %xmm8, %xmm2
; SSE2-NEXT: por %xmm3, %xmm2
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: movdqa %xmm1, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpgtd %xmm10, %xmm3
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm10, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pand %xmm4, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: por %xmm1, %xmm3
; SSE2-NEXT: pand %xmm2, %xmm3
; SSE2-NEXT: movdqa %xmm6, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm10, %xmm2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm10, %xmm1
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm7
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: por %xmm7, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm6, %xmm1
; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm3[0,2]
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pxor %xmm10, %xmm2
; SSE2-NEXT: movdqa %xmm2, %xmm3
; SSE2-NEXT: pcmpgtd %xmm10, %xmm3
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm10, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pand %xmm4, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSE2-NEXT: por %xmm2, %xmm3
; SSE2-NEXT: pand %xmm0, %xmm3
; SSE2-NEXT: movdqa %xmm5, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pcmpgtd %xmm10, %xmm2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm10, %xmm0
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: por %xmm6, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm5, %xmm0
; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v8i64_v8i32:
; SSSE3: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [4294967295,4294967295]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm5
; SSSE3-NEXT: pxor %xmm10, %xmm5
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147483647,2147483647]
; SSSE3-NEXT: movdqa %xmm9, %xmm6
; SSSE3-NEXT: pcmpgtd %xmm5, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
; SSSE3-NEXT: pand %xmm7, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: por %xmm4, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: pandn %xmm8, %xmm5
; SSSE3-NEXT: por %xmm0, %xmm5
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
; SSSE3-NEXT: movdqa %xmm9, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm0
; SSSE3-NEXT: pand %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm8, %xmm0
; SSSE3-NEXT: por %xmm1, %xmm0
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm9, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: por %xmm1, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm2
; SSSE3-NEXT: pandn %xmm8, %xmm6
; SSSE3-NEXT: por %xmm2, %xmm6
; SSSE3-NEXT: movdqa %xmm3, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm9, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pand %xmm4, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: pandn %xmm8, %xmm2
; SSSE3-NEXT: por %xmm3, %xmm2
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: movdqa %xmm1, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pand %xmm4, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: por %xmm1, %xmm3
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: movdqa %xmm6, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm7
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: por %xmm7, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm1
; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm3[0,2]
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: pxor %xmm10, %xmm2
; SSSE3-NEXT: movdqa %xmm2, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pand %xmm4, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSSE3-NEXT: por %xmm2, %xmm3
; SSSE3-NEXT: pand %xmm0, %xmm3
; SSSE3-NEXT: movdqa %xmm5, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: por %xmm6, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v8i64_v8i32:
; SSE41: # %bb.0:
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm0, %xmm9
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movapd {{.*#+}} xmm7 = [4294967295,4294967295]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pxor %xmm10, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm5 = [2147483647,2147483647]
; SSE41-NEXT: movdqa %xmm5, %xmm4
; SSE41-NEXT: pcmpeqd %xmm0, %xmm4
; SSE41-NEXT: movdqa %xmm5, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm4, %xmm0
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: por %xmm6, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm8
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm9, %xmm8
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm5, %xmm4
; SSE41-NEXT: pcmpeqd %xmm0, %xmm4
; SSE41-NEXT: movdqa %xmm5, %xmm6
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm4, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm9
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm9
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm5, %xmm1
; SSE41-NEXT: pcmpeqd %xmm0, %xmm1
; SSE41-NEXT: movdqa %xmm5, %xmm4
; SSE41-NEXT: pcmpgtd %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4
; SSE41-NEXT: movdqa %xmm3, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm5, %xmm1
; SSE41-NEXT: pcmpeqd %xmm0, %xmm1
; SSE41-NEXT: pcmpgtd %xmm0, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: por %xmm5, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm7
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: xorpd %xmm2, %xmm2
; SSE41-NEXT: movapd %xmm7, %xmm1
; SSE41-NEXT: xorpd %xmm10, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm3
; SSE41-NEXT: pcmpeqd %xmm10, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm3, %xmm0
; SSE41-NEXT: por %xmm1, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pxor %xmm3, %xmm3
; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm4, %xmm1
; SSE41-NEXT: xorpd %xmm10, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm1, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm1
; SSE41-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm3[0,2]
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm9, %xmm3
; SSE41-NEXT: xorpd %xmm10, %xmm3
; SSE41-NEXT: movapd %xmm3, %xmm4
; SSE41-NEXT: pcmpeqd %xmm10, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm4, %xmm0
; SSE41-NEXT: por %xmm3, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pxor %xmm3, %xmm3
; SSE41-NEXT: blendvpd %xmm0, %xmm9, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm8, %xmm4
; SSE41-NEXT: xorpd %xmm10, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm2
; SSE41-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,2],xmm3[0,2]
; SSE41-NEXT: movaps %xmm2, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v8i64_v8i32:
; AVX1: # %bb.0:
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [4294967295,4294967295]
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm8
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm3, %xmm5
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm6, %xmm3, %xmm7
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm3, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm3, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm9
; AVX1-NEXT: vblendvpd %xmm7, %xmm6, %xmm3, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm6, %xmm7
; AVX1-NEXT: vblendvpd %xmm5, %xmm1, %xmm3, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm5
; AVX1-NEXT: vblendvpd %xmm8, %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm2, %xmm3
; AVX1-NEXT: vpand %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm5, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,2],xmm2[0,2]
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vpand %xmm6, %xmm7, %xmm2
; AVX1-NEXT: vpand %xmm0, %xmm9, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,2]
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-SLOW-LABEL: trunc_packus_v8i64_v8i32:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm2 = [4294967295,4294967295,4294967295,4294967295]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; AVX2-SLOW-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3
; AVX2-SLOW-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1
[x86] split more v8f32/v8i32 shuffles in lowering Similar to D57867 - this is a small patch with lots of test diffs. With half-vector-width narrowing potential, using an extract + 128-bit vshufps is a win because it replaces a 256-bit shuffle with a 128-bit shufle. This seems like it should be a win even for targets with 'fast-variable-shuffle', but we are intentionally deferring that to an independent change to make sure that is true. Differential Revision: https://reviews.llvm.org/D58181 llvm-svn: 354279
2019-02-19 00:46:12 +08:00
; AVX2-SLOW-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3
; AVX2-SLOW-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-SLOW-NEXT: vpxor %xmm2, %xmm2, %xmm2
[x86] split more v8f32/v8i32 shuffles in lowering Similar to D57867 - this is a small patch with lots of test diffs. With half-vector-width narrowing potential, using an extract + 128-bit vshufps is a win because it replaces a 256-bit shuffle with a 128-bit shufle. This seems like it should be a win even for targets with 'fast-variable-shuffle', but we are intentionally deferring that to an independent change to make sure that is true. Differential Revision: https://reviews.llvm.org/D58181 llvm-svn: 354279
2019-02-19 00:46:12 +08:00
; AVX2-SLOW-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm3
; AVX2-SLOW-NEXT: vpand %ymm0, %ymm3, %ymm0
; AVX2-SLOW-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm2
; AVX2-SLOW-NEXT: vpand %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,2],xmm2[0,2]
; AVX2-SLOW-NEXT: vextracti128 $1, %ymm0, %xmm2
; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,2]
; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: trunc_packus_v8i64_v8i32:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm2 = [4294967295,4294967295,4294967295,4294967295]
; AVX2-FAST-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3
; AVX2-FAST-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; AVX2-FAST-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3
; AVX2-FAST-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-FAST-NEXT: vpxor %xmm2, %xmm2, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; AVX2-FAST-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm3
; AVX2-FAST-NEXT: vpand %ymm1, %ymm3, %ymm1
; AVX2-FAST-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm2
; AVX2-FAST-NEXT: vpand %ymm0, %ymm2, %ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [0,2,4,6,4,6,6,7]
; AVX2-FAST-NEXT: vpermd %ymm0, %ymm2, %ymm0
; AVX2-FAST-NEXT: vpermd %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-FAST-NEXT: retq
;
; AVX512-LABEL: trunc_packus_v8i64_v8i32:
; AVX512: # %bb.0:
; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512-NEXT: vpmovusqd %zmm0, %ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512-NEXT: retq
%1 = icmp slt <8 x i64> %a0, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
%2 = select <8 x i1> %1, <8 x i64> %a0, <8 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
%3 = icmp sgt <8 x i64> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer
%5 = trunc <8 x i64> %4 to <8 x i32>
ret <8 x i32> %5
}
;
; PACKUS saturation truncation to vXi16
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
define <4 x i16> @trunc_packus_v4i64_v4i16(<4 x i64> %a0) {
; SSE2-LABEL: trunc_packus_v4i64_v4i16:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: pxor %xmm2, %xmm3
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [2147549183,2147549183]
; SSE2-NEXT: movdqa %xmm5, %xmm6
; SSE2-NEXT: pcmpgtd %xmm3, %xmm6
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pcmpeqd %xmm5, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pand %xmm7, %xmm4
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
; SSE2-NEXT: por %xmm4, %xmm3
; SSE2-NEXT: pand %xmm3, %xmm1
; SSE2-NEXT: pandn %xmm8, %xmm3
; SSE2-NEXT: por %xmm1, %xmm3
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pxor %xmm2, %xmm1
; SSE2-NEXT: movdqa %xmm5, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpgtd %xmm1, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pcmpeqd %xmm5, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm4
; SSE2-NEXT: pand %xmm4, %xmm0
; SSE2-NEXT: pandn %xmm8, %xmm4
; SSE2-NEXT: por %xmm0, %xmm4
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: pxor %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pcmpgtd %xmm2, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm2, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm5, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: por %xmm0, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pand %xmm4, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm3, %xmm0
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pxor %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm4
; SSE2-NEXT: pcmpgtd %xmm2, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm2, %xmm0
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pand %xmm5, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSE2-NEXT: pand %xmm3, %xmm2
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm0[0,2,2,3,4,5,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-LABEL: trunc_packus_v4i64_v4i16:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pxor %xmm2, %xmm3
; SSSE3-NEXT: movdqa {{.*#+}} xmm5 = [2147549183,2147549183]
; SSSE3-NEXT: movdqa %xmm5, %xmm6
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm6
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pand %xmm7, %xmm4
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm1
; SSSE3-NEXT: pandn %xmm8, %xmm3
; SSSE3-NEXT: por %xmm1, %xmm3
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pxor %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm5, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm4
; SSSE3-NEXT: pand %xmm4, %xmm0
; SSSE3-NEXT: pandn %xmm8, %xmm4
; SSSE3-NEXT: por %xmm0, %xmm4
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: por %xmm0, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pand %xmm4, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm3, %xmm0
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pand %xmm3, %xmm2
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
; SSSE3-NEXT: pshuflw {{.*#+}} xmm2 = xmm0[0,2,2,3,4,5,6,7]
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
; SSSE3-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v4i64_v4i16:
; SSE41: # %bb.0:
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: movapd {{.*#+}} xmm4 = [65535,65535]
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm3, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm6 = [2147549183,2147549183]
; SSE41-NEXT: movdqa %xmm6, %xmm5
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm6, %xmm7
; SSE41-NEXT: pcmpgtd %xmm0, %xmm7
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm7, %xmm0
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm5
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm3, %xmm0
; SSE41-NEXT: movdqa %xmm6, %xmm1
; SSE41-NEXT: pcmpeqd %xmm0, %xmm1
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: movapd %xmm4, %xmm2
; SSE41-NEXT: xorpd %xmm3, %xmm2
; SSE41-NEXT: movapd %xmm2, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
; SSE41-NEXT: pcmpgtd %xmm3, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm2, %xmm2
; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm2
; SSE41-NEXT: movapd %xmm5, %xmm4
; SSE41-NEXT: xorpd %xmm3, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
; SSE41-NEXT: pcmpgtd %xmm3, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
; SSE41-NEXT: pshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
; SSE41-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; SSE41-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v4i64_v4i16:
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [65535,65535]
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm2, %xmm3
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm2, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm2, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm5
; AVX1-NEXT: vblendvpd %xmm3, %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; AVX1-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX1-NEXT: vpand %xmm0, %xmm5, %xmm0
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; AVX1-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-SLOW-LABEL: trunc_packus_v4i64_v4i16:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm1 = [65535,65535,65535,65535]
; AVX2-SLOW-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-SLOW-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-SLOW-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1
; AVX2-SLOW-NEXT: vpand %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX2-SLOW-NEXT: vzeroupper
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: trunc_packus_v4i64_v4i16:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm1 = [65535,65535,65535,65535]
; AVX2-FAST-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-FAST-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-FAST-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1
; AVX2-FAST-NEXT: vpand %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,8,9,8,9,10,11,8,9,10,11,12,13,14,15]
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FAST-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX2-FAST-NEXT: vzeroupper
; AVX2-FAST-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v4i64_v4i16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512F-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512F-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpmovqw %zmm0, %xmm0
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v4i64_v4i16:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
; AVX512VL-NEXT: vpmovusqw %ymm0, %xmm0
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v4i64_v4i16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512BW-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BW-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: vpmovqw %zmm0, %xmm0
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v4i64_v4i16:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
; AVX512BWVL-NEXT: vpmovusqw %ymm0, %xmm0
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
%1 = icmp slt <4 x i64> %a0, <i64 65535, i64 65535, i64 65535, i64 65535>
%2 = select <4 x i1> %1, <4 x i64> %a0, <4 x i64> <i64 65535, i64 65535, i64 65535, i64 65535>
%3 = icmp sgt <4 x i64> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i64> %2, <4 x i64> zeroinitializer
%5 = trunc <4 x i64> %4 to <4 x i16>
ret <4 x i16> %5
}
define void @trunc_packus_v4i64_v4i16_store(<4 x i64> %a0, <4 x i16> *%p1) {
; SSE2-LABEL: trunc_packus_v4i64_v4i16_store:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535]
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: pxor %xmm2, %xmm3
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [2147549183,2147549183]
; SSE2-NEXT: movdqa %xmm5, %xmm6
; SSE2-NEXT: pcmpgtd %xmm3, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm5, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
; SSE2-NEXT: pand %xmm7, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
; SSE2-NEXT: por %xmm4, %xmm3
; SSE2-NEXT: pand %xmm3, %xmm1
; SSE2-NEXT: pandn %xmm8, %xmm3
; SSE2-NEXT: por %xmm1, %xmm3
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pxor %xmm2, %xmm1
; SSE2-NEXT: movdqa %xmm5, %xmm4
; SSE2-NEXT: pcmpgtd %xmm1, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm5, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm4
; SSE2-NEXT: pand %xmm4, %xmm0
; SSE2-NEXT: pandn %xmm8, %xmm4
; SSE2-NEXT: por %xmm0, %xmm4
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: pxor %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pcmpgtd %xmm2, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm2, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm5, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: movdqa %xmm3, %xmm0
; SSE2-NEXT: pxor %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm4
; SSE2-NEXT: pcmpgtd %xmm2, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm2, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm5, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm2
; SSE2-NEXT: pand %xmm3, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE2-NEXT: movq %xmm1, (%rdi)
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v4i64_v4i16_store:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535]
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: movdqa %xmm1, %xmm3
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pxor %xmm2, %xmm3
; SSSE3-NEXT: movdqa {{.*#+}} xmm5 = [2147549183,2147549183]
; SSSE3-NEXT: movdqa %xmm5, %xmm6
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
; SSSE3-NEXT: pand %xmm7, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm1
; SSSE3-NEXT: pandn %xmm8, %xmm3
; SSSE3-NEXT: por %xmm1, %xmm3
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pxor %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm5, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm4
; SSSE3-NEXT: pand %xmm4, %xmm0
; SSSE3-NEXT: pandn %xmm8, %xmm4
; SSSE3-NEXT: por %xmm0, %xmm4
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm3, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
; SSSE3-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSSE3-NEXT: movq %xmm1, (%rdi)
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v4i64_v4i16_store:
; SSE41: # %bb.0:
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: movapd {{.*#+}} xmm4 = [65535,65535]
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm3, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm6 = [2147549183,2147549183]
; SSE41-NEXT: movdqa %xmm6, %xmm5
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm6, %xmm7
; SSE41-NEXT: pcmpgtd %xmm0, %xmm7
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm7, %xmm0
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm5
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm3, %xmm0
; SSE41-NEXT: movdqa %xmm6, %xmm1
; SSE41-NEXT: pcmpeqd %xmm0, %xmm1
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: movapd %xmm4, %xmm2
; SSE41-NEXT: xorpd %xmm3, %xmm2
; SSE41-NEXT: movapd %xmm2, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
; SSE41-NEXT: pcmpgtd %xmm3, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm2, %xmm2
; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm2
; SSE41-NEXT: movapd %xmm5, %xmm4
; SSE41-NEXT: xorpd %xmm3, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
; SSE41-NEXT: pcmpgtd %xmm3, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
; SSE41-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3]
; SSE41-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; SSE41-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE41-NEXT: movq %xmm1, (%rdi)
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v4i64_v4i16_store:
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [65535,65535]
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm2, %xmm3
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm2, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm2, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm5
; AVX1-NEXT: vblendvpd %xmm3, %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; AVX1-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX1-NEXT: vpand %xmm0, %xmm5, %xmm0
; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; AVX1-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX1-NEXT: vmovq %xmm0, (%rdi)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-SLOW-LABEL: trunc_packus_v4i64_v4i16_store:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm1 = [65535,65535,65535,65535]
; AVX2-SLOW-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-SLOW-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-SLOW-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1
; AVX2-SLOW-NEXT: vpand %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
; AVX2-SLOW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX2-SLOW-NEXT: vmovq %xmm0, (%rdi)
; AVX2-SLOW-NEXT: vzeroupper
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: trunc_packus_v4i64_v4i16_store:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm1 = [65535,65535,65535,65535]
; AVX2-FAST-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-FAST-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-FAST-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1
; AVX2-FAST-NEXT: vpand %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,8,9,8,9,10,11,8,9,10,11,12,13,14,15]
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FAST-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX2-FAST-NEXT: vmovq %xmm0, (%rdi)
; AVX2-FAST-NEXT: vzeroupper
; AVX2-FAST-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v4i64_v4i16_store:
; AVX512F: # %bb.0:
; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512F-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512F-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpmovqw %zmm0, %xmm0
; AVX512F-NEXT: vmovq %xmm0, (%rdi)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v4i64_v4i16_store:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
; AVX512VL-NEXT: vpmovusqw %ymm0, (%rdi)
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v4i64_v4i16_store:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512BW-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BW-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: vpmovqw %zmm0, %xmm0
; AVX512BW-NEXT: vmovq %xmm0, (%rdi)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v4i64_v4i16_store:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
; AVX512BWVL-NEXT: vpmovusqw %ymm0, (%rdi)
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
%1 = icmp slt <4 x i64> %a0, <i64 65535, i64 65535, i64 65535, i64 65535>
%2 = select <4 x i1> %1, <4 x i64> %a0, <4 x i64> <i64 65535, i64 65535, i64 65535, i64 65535>
%3 = icmp sgt <4 x i64> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i64> %2, <4 x i64> zeroinitializer
%5 = trunc <4 x i64> %4 to <4 x i16>
store <4 x i16> %5, <4 x i16> *%p1
ret void
}
define <8 x i16> @trunc_packus_v8i64_v8i16(<8 x i64> %a0) {
; SSE2-LABEL: trunc_packus_v8i64_v8i16:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535]
; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
; SSE2-NEXT: movdqa %xmm1, %xmm5
; SSE2-NEXT: pxor %xmm10, %xmm5
; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147549183,2147549183]
; SSE2-NEXT: movdqa %xmm9, %xmm6
; SSE2-NEXT: pcmpgtd %xmm5, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
; SSE2-NEXT: pand %xmm7, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3]
; SSE2-NEXT: por %xmm4, %xmm5
; SSE2-NEXT: pand %xmm5, %xmm1
; SSE2-NEXT: pandn %xmm8, %xmm5
; SSE2-NEXT: por %xmm1, %xmm5
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm9, %xmm4
; SSE2-NEXT: pcmpgtd %xmm1, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm1
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: pandn %xmm8, %xmm1
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: movdqa %xmm3, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
; SSE2-NEXT: movdqa %xmm9, %xmm4
; SSE2-NEXT: pcmpgtd %xmm0, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm6
; SSE2-NEXT: pand %xmm6, %xmm3
; SSE2-NEXT: pandn %xmm8, %xmm6
; SSE2-NEXT: por %xmm3, %xmm6
; SSE2-NEXT: movdqa %xmm2, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
; SSE2-NEXT: movdqa %xmm9, %xmm3
; SSE2-NEXT: pcmpgtd %xmm0, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm3
; SSE2-NEXT: pand %xmm3, %xmm2
; SSE2-NEXT: pandn %xmm8, %xmm3
; SSE2-NEXT: por %xmm2, %xmm3
; SSE2-NEXT: movdqa %xmm3, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pcmpgtd %xmm10, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm0
; SSE2-NEXT: pand %xmm3, %xmm0
; SSE2-NEXT: movdqa %xmm6, %xmm2
; SSE2-NEXT: pxor %xmm10, %xmm2
; SSE2-NEXT: movdqa %xmm2, %xmm3
; SSE2-NEXT: pcmpgtd %xmm10, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm2
; SSE2-NEXT: pand %xmm6, %xmm2
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: pxor %xmm10, %xmm3
; SSE2-NEXT: movdqa %xmm3, %xmm4
; SSE2-NEXT: pcmpgtd %xmm10, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm3, %xmm4
; SSE2-NEXT: pand %xmm1, %xmm4
; SSE2-NEXT: movdqa %xmm5, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: pcmpgtd %xmm10, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm3
; SSE2-NEXT: pand %xmm5, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
; SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,2,4,5,6,7]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v8i64_v8i16:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535]
; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
; SSSE3-NEXT: movdqa %xmm1, %xmm5
; SSSE3-NEXT: pxor %xmm10, %xmm5
; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147549183,2147549183]
; SSSE3-NEXT: movdqa %xmm9, %xmm6
; SSSE3-NEXT: pcmpgtd %xmm5, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm5
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
; SSSE3-NEXT: pand %xmm7, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm5
; SSSE3-NEXT: pand %xmm5, %xmm1
; SSSE3-NEXT: pandn %xmm8, %xmm5
; SSSE3-NEXT: por %xmm1, %xmm5
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm9, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm1
; SSSE3-NEXT: pand %xmm1, %xmm0
; SSSE3-NEXT: pandn %xmm8, %xmm1
; SSSE3-NEXT: por %xmm0, %xmm1
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
; SSSE3-NEXT: movdqa %xmm9, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm6
; SSSE3-NEXT: pand %xmm6, %xmm3
; SSSE3-NEXT: pandn %xmm8, %xmm6
; SSSE3-NEXT: por %xmm3, %xmm6
; SSSE3-NEXT: movdqa %xmm2, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
; SSSE3-NEXT: movdqa %xmm9, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm2
; SSSE3-NEXT: pandn %xmm8, %xmm3
; SSSE3-NEXT: por %xmm2, %xmm3
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm0
; SSSE3-NEXT: pand %xmm3, %xmm0
; SSSE3-NEXT: movdqa %xmm6, %xmm2
; SSSE3-NEXT: pxor %xmm10, %xmm2
; SSSE3-NEXT: movdqa %xmm2, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm2
; SSSE3-NEXT: pand %xmm6, %xmm2
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pxor %xmm10, %xmm3
; SSSE3-NEXT: movdqa %xmm3, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm3, %xmm4
; SSSE3-NEXT: pand %xmm1, %xmm4
; SSSE3-NEXT: movdqa %xmm5, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm3
; SSSE3-NEXT: pand %xmm5, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3]
; SSSE3-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3]
; SSSE3-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3]
; SSSE3-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7]
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSSE3-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,2,4,5,6,7]
; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSSE3-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v8i64_v8i16:
; SSE41: # %bb.0:
; SSE41-NEXT: movdqa %xmm0, %xmm9
; SSE41-NEXT: movapd {{.*#+}} xmm7 = [65535,65535]
; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm4 = [2147549183,2147549183]
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm4, %xmm6
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
; SSE41-NEXT: movapd %xmm7, %xmm8
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm8
; SSE41-NEXT: movdqa %xmm3, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm2
; SSE41-NEXT: pcmpeqd %xmm0, %xmm2
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpgtd %xmm0, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: por %xmm5, %xmm0
; SSE41-NEXT: movapd %xmm7, %xmm2
; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm2
; SSE41-NEXT: movdqa %xmm9, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm3
; SSE41-NEXT: pcmpeqd %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpgtd %xmm0, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
; SSE41-NEXT: pand %xmm3, %xmm0
; SSE41-NEXT: por %xmm5, %xmm0
; SSE41-NEXT: movapd %xmm7, %xmm6
; SSE41-NEXT: blendvpd %xmm0, %xmm9, %xmm6
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm3
; SSE41-NEXT: pcmpeqd %xmm0, %xmm3
; SSE41-NEXT: pcmpgtd %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm3, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm7
; SSE41-NEXT: pxor %xmm3, %xmm3
; SSE41-NEXT: movapd %xmm7, %xmm1
; SSE41-NEXT: xorpd %xmm10, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm4
; SSE41-NEXT: pcmpeqd %xmm10, %xmm4
; SSE41-NEXT: pcmpgtd %xmm10, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
; SSE41-NEXT: pand %xmm4, %xmm0
; SSE41-NEXT: por %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm4, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm4
; SSE41-NEXT: movapd %xmm6, %xmm1
; SSE41-NEXT: xorpd %xmm10, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
; SSE41-NEXT: pcmpgtd %xmm10, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: blendvpd %xmm0, %xmm6, %xmm1
; SSE41-NEXT: packusdw %xmm4, %xmm1
; SSE41-NEXT: movapd %xmm2, %xmm4
; SSE41-NEXT: xorpd %xmm10, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
; SSE41-NEXT: pcmpgtd %xmm10, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
; SSE41-NEXT: pxor %xmm4, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4
; SSE41-NEXT: movapd %xmm8, %xmm2
; SSE41-NEXT: xorpd %xmm10, %xmm2
; SSE41-NEXT: movapd %xmm2, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
; SSE41-NEXT: pcmpgtd %xmm10, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm2, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm3
; SSE41-NEXT: packusdw %xmm4, %xmm3
; SSE41-NEXT: packusdw %xmm3, %xmm1
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v8i64_v8i16:
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [65535,65535]
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm8
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm3, %xmm5
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm6, %xmm3, %xmm7
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm3, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm3, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm9
; AVX1-NEXT: vblendvpd %xmm7, %xmm6, %xmm3, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm6, %xmm7
; AVX1-NEXT: vblendvpd %xmm5, %xmm1, %xmm3, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm5
; AVX1-NEXT: vblendvpd %xmm8, %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm2, %xmm3
; AVX1-NEXT: vpand %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm5, %xmm1
; AVX1-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
; AVX1-NEXT: vpand %xmm6, %xmm7, %xmm2
; AVX1-NEXT: vpand %xmm0, %xmm9, %xmm0
; AVX1-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v8i64_v8i16:
; AVX2: # %bb.0:
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [65535,65535,65535,65535]
; AVX2-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3
; AVX2-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0
; AVX2-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3
; AVX2-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1
; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm3
; AVX2-NEXT: vpand %ymm1, %ymm3, %ymm1
; AVX2-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm2
; AVX2-NEXT: vpand %ymm0, %ymm2, %ymm0
; AVX2-NEXT: vpackusdw %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
; AVX512-LABEL: trunc_packus_v8i64_v8i16:
; AVX512: # %bb.0:
; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512-NEXT: vpmovusqw %zmm0, %xmm0
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
%1 = icmp slt <8 x i64> %a0, <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535>
%2 = select <8 x i1> %1, <8 x i64> %a0, <8 x i64> <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535>
%3 = icmp sgt <8 x i64> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer
%5 = trunc <8 x i64> %4 to <8 x i16>
ret <8 x i16> %5
}
define <4 x i16> @trunc_packus_v4i32_v4i16(<4 x i32> %a0) {
; SSE2-LABEL: trunc_packus_v4i32_v4i16:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm0, %xmm2
; SSE2-NEXT: pand %xmm2, %xmm0
; SSE2-NEXT: pandn %xmm1, %xmm2
; SSE2-NEXT: por %xmm0, %xmm2
; SSE2-NEXT: pxor %xmm0, %xmm0
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: pcmpgtd %xmm0, %xmm1
; SSE2-NEXT: pand %xmm2, %xmm1
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm1[0,2,2,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v4i32_v4i16:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pandn %xmm1, %xmm2
; SSSE3-NEXT: por %xmm0, %xmm2
; SSSE3-NEXT: pxor %xmm1, %xmm1
; SSSE3-NEXT: movdqa %xmm2, %xmm0
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm0
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v4i32_v4i16:
; SSE41: # %bb.0:
; SSE41-NEXT: packusdw %xmm0, %xmm0
; SSE41-NEXT: retq
;
; AVX-LABEL: trunc_packus_v4i32_v4i16:
; AVX: # %bb.0:
; AVX-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
; AVX-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v4i32_v4i16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v4i32_v4i16:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512VL-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v4i32_v4i16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v4i32_v4i16:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512BWVL-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
; AVX512BWVL-NEXT: retq
%1 = icmp slt <4 x i32> %a0, <i32 65535, i32 65535, i32 65535, i32 65535>
%2 = select <4 x i1> %1, <4 x i32> %a0, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>
%3 = icmp sgt <4 x i32> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i32> %2, <4 x i32> zeroinitializer
%5 = trunc <4 x i32> %4 to <4 x i16>
ret <4 x i16> %5
}
define void @trunc_packus_v4i32_v4i16_store(<4 x i32> %a0, <4 x i16> *%p1) {
; SSE2-LABEL: trunc_packus_v4i32_v4i16_store:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm0, %xmm2
; SSE2-NEXT: pand %xmm2, %xmm0
; SSE2-NEXT: pandn %xmm1, %xmm2
; SSE2-NEXT: por %xmm0, %xmm2
; SSE2-NEXT: pxor %xmm0, %xmm0
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: pcmpgtd %xmm0, %xmm1
; SSE2-NEXT: pand %xmm2, %xmm1
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm1[0,2,2,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE2-NEXT: movq %xmm0, (%rdi)
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v4i32_v4i16_store:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pandn %xmm1, %xmm2
; SSSE3-NEXT: por %xmm0, %xmm2
; SSSE3-NEXT: pxor %xmm0, %xmm0
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm2, %xmm1
; SSSE3-NEXT: pshufb {{.*#+}} xmm1 = xmm1[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
; SSSE3-NEXT: movq %xmm1, (%rdi)
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v4i32_v4i16_store:
; SSE41: # %bb.0:
; SSE41-NEXT: packusdw %xmm0, %xmm0
; SSE41-NEXT: movq %xmm0, (%rdi)
; SSE41-NEXT: retq
;
; AVX-LABEL: trunc_packus_v4i32_v4i16_store:
; AVX: # %bb.0:
; AVX-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
; AVX-NEXT: vmovq %xmm0, (%rdi)
; AVX-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v4i32_v4i16_store:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
; AVX512F-NEXT: vmovq %xmm0, (%rdi)
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v4i32_v4i16_store:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512VL-NEXT: vpmovusdw %xmm0, (%rdi)
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v4i32_v4i16_store:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpackusdw %xmm0, %xmm0, %xmm0
; AVX512BW-NEXT: vmovq %xmm0, (%rdi)
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v4i32_v4i16_store:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512BWVL-NEXT: vpmovusdw %xmm0, (%rdi)
; AVX512BWVL-NEXT: retq
%1 = icmp slt <4 x i32> %a0, <i32 65535, i32 65535, i32 65535, i32 65535>
%2 = select <4 x i1> %1, <4 x i32> %a0, <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>
%3 = icmp sgt <4 x i32> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i32> %2, <4 x i32> zeroinitializer
%5 = trunc <4 x i32> %4 to <4 x i16>
store <4 x i16> %5, <4 x i16> *%p1
ret void
}
define <8 x i16> @trunc_packus_v8i32_v8i16(<8 x i32> %a0) {
; SSE2-LABEL: trunc_packus_v8i32_v8i16:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,65535]
; SSE2-NEXT: movdqa %xmm2, %xmm3
; SSE2-NEXT: pcmpgtd %xmm1, %xmm3
; SSE2-NEXT: pand %xmm3, %xmm1
; SSE2-NEXT: pandn %xmm2, %xmm3
; SSE2-NEXT: por %xmm1, %xmm3
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: pcmpgtd %xmm0, %xmm1
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: pandn %xmm2, %xmm1
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: pxor %xmm2, %xmm2
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: pcmpgtd %xmm2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm3, %xmm1
; SSE2-NEXT: pcmpgtd %xmm2, %xmm1
; SSE2-NEXT: pand %xmm3, %xmm1
; SSE2-NEXT: pslld $16, %xmm1
; SSE2-NEXT: psrad $16, %xmm1
; SSE2-NEXT: pslld $16, %xmm0
; SSE2-NEXT: psrad $16, %xmm0
; SSE2-NEXT: packssdw %xmm1, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v8i32_v8i16:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,65535]
; SSSE3-NEXT: movdqa %xmm2, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm1
; SSSE3-NEXT: pandn %xmm2, %xmm3
; SSSE3-NEXT: por %xmm1, %xmm3
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm1, %xmm0
; SSSE3-NEXT: pandn %xmm2, %xmm1
; SSSE3-NEXT: por %xmm0, %xmm1
; SSSE3-NEXT: pxor %xmm2, %xmm2
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm0
; SSSE3-NEXT: pand %xmm1, %xmm0
; SSSE3-NEXT: movdqa %xmm3, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1
; SSSE3-NEXT: pand %xmm3, %xmm1
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
; SSSE3-NEXT: pshufb %xmm2, %xmm1
; SSSE3-NEXT: pshufb %xmm2, %xmm0
; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v8i32_v8i16:
; SSE41: # %bb.0:
; SSE41-NEXT: packusdw %xmm1, %xmm0
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v8i32_v8i16:
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v8i32_v8i16:
; AVX2: # %bb.0:
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v8i32_v8i16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v8i32_v8i16:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0
; AVX512VL-NEXT: vpmovusdw %ymm0, %xmm0
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v8i32_v8i16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512BW-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v8i32_v8i16:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0
; AVX512BWVL-NEXT: vpmovusdw %ymm0, %xmm0
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
%1 = icmp slt <8 x i32> %a0, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
%2 = select <8 x i1> %1, <8 x i32> %a0, <8 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
%3 = icmp sgt <8 x i32> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i32> %2, <8 x i32> zeroinitializer
%5 = trunc <8 x i32> %4 to <8 x i16>
ret <8 x i16> %5
}
define <16 x i16> @trunc_packus_v16i32_v16i16(<16 x i32> %a0) {
; SSE2-LABEL: trunc_packus_v16i32_v16i16:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,65535,65535]
; SSE2-NEXT: movdqa %xmm6, %xmm4
; SSE2-NEXT: pcmpgtd %xmm1, %xmm4
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pandn %xmm6, %xmm4
; SSE2-NEXT: por %xmm1, %xmm4
; SSE2-NEXT: movdqa %xmm6, %xmm5
; SSE2-NEXT: pcmpgtd %xmm0, %xmm5
; SSE2-NEXT: pand %xmm5, %xmm0
; SSE2-NEXT: pandn %xmm6, %xmm5
; SSE2-NEXT: por %xmm0, %xmm5
; SSE2-NEXT: movdqa %xmm6, %xmm0
; SSE2-NEXT: pcmpgtd %xmm3, %xmm0
; SSE2-NEXT: pand %xmm0, %xmm3
; SSE2-NEXT: pandn %xmm6, %xmm0
; SSE2-NEXT: por %xmm3, %xmm0
; SSE2-NEXT: movdqa %xmm6, %xmm3
; SSE2-NEXT: pcmpgtd %xmm2, %xmm3
; SSE2-NEXT: pand %xmm3, %xmm2
; SSE2-NEXT: pandn %xmm6, %xmm3
; SSE2-NEXT: por %xmm2, %xmm3
; SSE2-NEXT: pxor %xmm2, %xmm2
; SSE2-NEXT: movdqa %xmm3, %xmm1
; SSE2-NEXT: pcmpgtd %xmm2, %xmm1
; SSE2-NEXT: pand %xmm3, %xmm1
; SSE2-NEXT: movdqa %xmm0, %xmm3
; SSE2-NEXT: pcmpgtd %xmm2, %xmm3
; SSE2-NEXT: pand %xmm0, %xmm3
; SSE2-NEXT: movdqa %xmm5, %xmm0
; SSE2-NEXT: pcmpgtd %xmm2, %xmm0
; SSE2-NEXT: pand %xmm5, %xmm0
; SSE2-NEXT: movdqa %xmm4, %xmm5
; SSE2-NEXT: pcmpgtd %xmm2, %xmm5
; SSE2-NEXT: pand %xmm4, %xmm5
; SSE2-NEXT: pslld $16, %xmm5
; SSE2-NEXT: psrad $16, %xmm5
; SSE2-NEXT: pslld $16, %xmm0
; SSE2-NEXT: psrad $16, %xmm0
; SSE2-NEXT: packssdw %xmm5, %xmm0
; SSE2-NEXT: pslld $16, %xmm3
; SSE2-NEXT: psrad $16, %xmm3
; SSE2-NEXT: pslld $16, %xmm1
; SSE2-NEXT: psrad $16, %xmm1
; SSE2-NEXT: packssdw %xmm3, %xmm1
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v16i32_v16i16:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,65535,65535]
; SSSE3-NEXT: movdqa %xmm6, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: pandn %xmm6, %xmm4
; SSSE3-NEXT: por %xmm1, %xmm4
; SSSE3-NEXT: movdqa %xmm6, %xmm5
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm5
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: pandn %xmm6, %xmm5
; SSSE3-NEXT: por %xmm0, %xmm5
; SSSE3-NEXT: movdqa %xmm6, %xmm0
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm0
; SSSE3-NEXT: pand %xmm0, %xmm3
; SSSE3-NEXT: pandn %xmm6, %xmm0
; SSSE3-NEXT: por %xmm3, %xmm0
; SSSE3-NEXT: movdqa %xmm6, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm2
; SSSE3-NEXT: pandn %xmm6, %xmm3
; SSSE3-NEXT: por %xmm2, %xmm3
; SSSE3-NEXT: pxor %xmm2, %xmm2
; SSSE3-NEXT: movdqa %xmm3, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1
; SSSE3-NEXT: pand %xmm3, %xmm1
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm3
; SSSE3-NEXT: pand %xmm0, %xmm3
; SSSE3-NEXT: movdqa %xmm5, %xmm0
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm0
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: movdqa %xmm4, %xmm5
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm5
; SSSE3-NEXT: pand %xmm4, %xmm5
; SSSE3-NEXT: pslld $16, %xmm5
; SSSE3-NEXT: psrad $16, %xmm5
; SSSE3-NEXT: pslld $16, %xmm0
; SSSE3-NEXT: psrad $16, %xmm0
; SSSE3-NEXT: packssdw %xmm5, %xmm0
; SSSE3-NEXT: pslld $16, %xmm3
; SSSE3-NEXT: psrad $16, %xmm3
; SSSE3-NEXT: pslld $16, %xmm1
; SSSE3-NEXT: psrad $16, %xmm1
; SSSE3-NEXT: packssdw %xmm3, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE41-LABEL: trunc_packus_v16i32_v16i16:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE41-NEXT: packusdw %xmm1, %xmm0
; SSE41-NEXT: packusdw %xmm3, %xmm2
; SSE41-NEXT: movdqa %xmm2, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX1-LABEL: trunc_packus_v16i32_v16i16:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1: # %bb.0:
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; AVX1-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; AVX1-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX2-LABEL: trunc_packus_v16i32_v16i16:
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; AVX2: # %bb.0:
; AVX2-NEXT: vpackusdw %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512-LABEL: trunc_packus_v16i32_v16i16:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512: # %bb.0:
; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512-NEXT: vpmaxsd %zmm1, %zmm0, %zmm0
; AVX512-NEXT: vpmovusdw %zmm0, %ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512-NEXT: retq
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
%1 = icmp slt <16 x i32> %a0, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
%2 = select <16 x i1> %1, <16 x i32> %a0, <16 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
%3 = icmp sgt <16 x i32> %2, zeroinitializer
%4 = select <16 x i1> %3, <16 x i32> %2, <16 x i32> zeroinitializer
%5 = trunc <16 x i32> %4 to <16 x i16>
ret <16 x i16> %5
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
}
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
;
; PACKUS saturation truncation to vXi8
;
define <4 x i8> @trunc_packus_v4i64_v4i8(<4 x i64> %a0) {
; SSE2-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [255,255]
; SSE2-NEXT: movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
; SSE2-NEXT: movdqa %xmm1, %xmm4
; SSE2-NEXT: pxor %xmm3, %xmm4
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [2147483903,2147483903]
; SSE2-NEXT: movdqa %xmm5, %xmm6
; SSE2-NEXT: pcmpgtd %xmm4, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm5, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSE2-NEXT: pand %xmm7, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
; SSE2-NEXT: por %xmm2, %xmm4
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pandn %xmm8, %xmm4
; SSE2-NEXT: por %xmm1, %xmm4
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pxor %xmm3, %xmm1
; SSE2-NEXT: movdqa %xmm5, %xmm2
; SSE2-NEXT: pcmpgtd %xmm1, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm5, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm5, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm1, %xmm0
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pandn %xmm8, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm0
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pxor %xmm3, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pcmpgtd %xmm3, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm3, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm5, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm6, %xmm0
; SSE2-NEXT: movdqa %xmm4, %xmm2
; SSE2-NEXT: pxor %xmm3, %xmm2
; SSE2-NEXT: movdqa %xmm2, %xmm5
; SSE2-NEXT: pcmpgtd %xmm3, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm3, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
; SSE2-NEXT: por %xmm2, %xmm3
; SSE2-NEXT: pand %xmm8, %xmm3
; SSE2-NEXT: pand %xmm4, %xmm3
; SSE2-NEXT: pand %xmm8, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm1, %xmm0
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: packuswb %xmm3, %xmm0
; SSE2-NEXT: packuswb %xmm0, %xmm0
; SSE2-NEXT: packuswb %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [255,255]
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pxor %xmm2, %xmm3
; SSSE3-NEXT: movdqa {{.*#+}} xmm5 = [2147483903,2147483903]
; SSSE3-NEXT: movdqa %xmm5, %xmm6
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
; SSSE3-NEXT: pand %xmm7, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm3
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pand %xmm3, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pandn %xmm8, %xmm3
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: por %xmm1, %xmm3
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pxor %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm5, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm4
; SSSE3-NEXT: pand %xmm4, %xmm0
; SSSE3-NEXT: pandn %xmm8, %xmm4
; SSSE3-NEXT: por %xmm0, %xmm4
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm5, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
; SSSE3-NEXT: por %xmm6, %xmm0
; SSSE3-NEXT: pand %xmm4, %xmm0
; SSSE3-NEXT: movdqa %xmm3, %xmm1
; SSSE3-NEXT: pxor %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm5, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm2
; SSSE3-NEXT: pand %xmm3, %xmm2
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; SSSE3-NEXT: pshufb %xmm1, %xmm2
; SSSE3-NEXT: pshufb %xmm1, %xmm0
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE41-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: movapd {{.*#+}} xmm4 = [255,255]
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm3, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm6 = [2147483903,2147483903]
; SSE41-NEXT: movdqa %xmm6, %xmm5
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm6, %xmm7
; SSE41-NEXT: pcmpgtd %xmm0, %xmm7
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm7, %xmm0
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm5
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm3, %xmm0
; SSE41-NEXT: movdqa %xmm6, %xmm1
; SSE41-NEXT: pcmpeqd %xmm0, %xmm1
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4
; SSE41-NEXT: xorpd %xmm2, %xmm2
; SSE41-NEXT: movapd %xmm4, %xmm1
; SSE41-NEXT: xorpd %xmm3, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
; SSE41-NEXT: pcmpgtd %xmm3, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm1
; SSE41-NEXT: movapd %xmm5, %xmm4
; SSE41-NEXT: xorpd %xmm3, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
; SSE41-NEXT: pcmpgtd %xmm3, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm2
; SSE41-NEXT: movdqa {{.*#+}} xmm0 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; SSE41-NEXT: pshufb %xmm0, %xmm2
; SSE41-NEXT: pshufb %xmm0, %xmm1
; SSE41-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE41-NEXT: movdqa %xmm1, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX1-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [255,255]
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm2, %xmm3
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm2, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm2, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm5
; AVX1-NEXT: vblendvpd %xmm3, %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX1-NEXT: vpand %xmm0, %xmm5, %xmm0
; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX2-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm1 = [255,255,255,255]
; AVX2-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1
; AVX2-NEXT: vpand %ymm0, %ymm1, %ymm0
[X86][AVX] Add PACKSSDW/PACKUSDW support for truncation of clamped values SSE and shorter vector sizes will have to wait until we can add support for general SMIN/SMAX matching. llvm-svn: 324485
2018-02-07 23:48:44 +08:00
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512F-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512F: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512F-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512F-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpmovqb %zmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512VL-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512VL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
; AVX512VL-NEXT: vpmovusqb %ymm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512BW-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BW: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512BW-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BW-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: vpmovqb %zmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512BWVL-LABEL: trunc_packus_v4i64_v4i8:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512BWVL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
; AVX512BWVL-NEXT: vpmovusqb %ymm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
%1 = icmp slt <4 x i64> %a0, <i64 255, i64 255, i64 255, i64 255>
%2 = select <4 x i1> %1, <4 x i64> %a0, <4 x i64> <i64 255, i64 255, i64 255, i64 255>
%3 = icmp sgt <4 x i64> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i64> %2, <4 x i64> zeroinitializer
%5 = trunc <4 x i64> %4 to <4 x i8>
ret <4 x i8> %5
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
}
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
define void @trunc_packus_v4i64_v4i8_store(<4 x i64> %a0, <4 x i8> *%p1) {
; SSE2-LABEL: trunc_packus_v4i64_v4i8_store:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [255,255]
; SSE2-NEXT: movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
; SSE2-NEXT: movdqa %xmm1, %xmm4
; SSE2-NEXT: pxor %xmm3, %xmm4
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [2147483903,2147483903]
; SSE2-NEXT: movdqa %xmm5, %xmm6
; SSE2-NEXT: pcmpgtd %xmm4, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm5, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSE2-NEXT: pand %xmm7, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
; SSE2-NEXT: por %xmm2, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm4, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pandn %xmm8, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: por %xmm1, %xmm4
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pxor %xmm3, %xmm1
; SSE2-NEXT: movdqa %xmm5, %xmm2
; SSE2-NEXT: pcmpgtd %xmm1, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm5, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm5, %xmm1
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: pandn %xmm8, %xmm1
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: pxor %xmm3, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pcmpgtd %xmm3, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm3, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pand %xmm5, %xmm0
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm2
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: pxor %xmm3, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm5
; SSE2-NEXT: pcmpgtd %xmm3, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm3, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm3
; SSE2-NEXT: pand %xmm8, %xmm3
; SSE2-NEXT: pand %xmm4, %xmm3
; SSE2-NEXT: pand %xmm8, %xmm2
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: packuswb %xmm3, %xmm2
; SSE2-NEXT: packuswb %xmm0, %xmm2
; SSE2-NEXT: packuswb %xmm0, %xmm2
; SSE2-NEXT: movd %xmm2, (%rdi)
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-LABEL: trunc_packus_v4i64_v4i8_store:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [255,255]
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pxor %xmm2, %xmm3
; SSSE3-NEXT: movdqa {{.*#+}} xmm5 = [2147483903,2147483903]
; SSSE3-NEXT: movdqa %xmm5, %xmm6
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
; SSSE3-NEXT: pand %xmm7, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm1
; SSSE3-NEXT: pandn %xmm8, %xmm3
; SSSE3-NEXT: por %xmm1, %xmm3
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pxor %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm5, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm4
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm5, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: por %xmm1, %xmm4
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pand %xmm4, %xmm0
; SSSE3-NEXT: pandn %xmm8, %xmm4
; SSSE3-NEXT: por %xmm0, %xmm4
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm1
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm5, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: pxor %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pand %xmm5, %xmm0
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm3, %xmm2
; SSSE3-NEXT: movdqa {{.*#+}} xmm0 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; SSSE3-NEXT: pshufb %xmm0, %xmm2
; SSSE3-NEXT: pshufb %xmm0, %xmm1
; SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSSE3-NEXT: movd %xmm1, (%rdi)
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE41-LABEL: trunc_packus_v4i64_v4i8_store:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: movapd {{.*#+}} xmm4 = [255,255]
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm3, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm6 = [2147483903,2147483903]
; SSE41-NEXT: movdqa %xmm6, %xmm5
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm6, %xmm7
; SSE41-NEXT: pcmpgtd %xmm0, %xmm7
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm7, %xmm0
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm5
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm3, %xmm0
; SSE41-NEXT: movdqa %xmm6, %xmm1
; SSE41-NEXT: pcmpeqd %xmm0, %xmm1
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: movapd %xmm4, %xmm2
; SSE41-NEXT: xorpd %xmm3, %xmm2
; SSE41-NEXT: movapd %xmm2, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
; SSE41-NEXT: pcmpgtd %xmm3, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm2, %xmm0
; SSE41-NEXT: pxor %xmm2, %xmm2
; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm2
; SSE41-NEXT: movapd %xmm5, %xmm4
; SSE41-NEXT: xorpd %xmm3, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm6
; SSE41-NEXT: pcmpeqd %xmm3, %xmm6
; SSE41-NEXT: pcmpgtd %xmm3, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm1
; SSE41-NEXT: movdqa {{.*#+}} xmm0 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; SSE41-NEXT: pshufb %xmm0, %xmm1
; SSE41-NEXT: pshufb %xmm0, %xmm2
; SSE41-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE41-NEXT: movd %xmm2, (%rdi)
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX1-LABEL: trunc_packus_v4i64_v4i8_store:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [255,255]
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm2, %xmm3
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm2, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm2, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm5
; AVX1-NEXT: vblendvpd %xmm3, %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX1-NEXT: vpand %xmm0, %xmm5, %xmm0
; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX1-NEXT: vmovd %xmm0, (%rdi)
; AVX1-NEXT: vzeroupper
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX2-LABEL: trunc_packus_v4i64_v4i8_store:
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2: # %bb.0:
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm1 = [255,255,255,255]
; AVX2-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-NEXT: vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm1
; AVX2-NEXT: vpand %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX2-NEXT: vmovd %xmm0, (%rdi)
; AVX2-NEXT: vzeroupper
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-NEXT: retq
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512F-LABEL: trunc_packus_v4i64_v4i8_store:
; AVX512F: # %bb.0:
; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512F-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512F-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpmovqb %zmm0, %xmm0
; AVX512F-NEXT: vmovd %xmm0, (%rdi)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v4i64_v4i8_store:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
; AVX512VL-NEXT: vpmovusqb %ymm0, (%rdi)
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512BW-LABEL: trunc_packus_v4i64_v4i8_store:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512BW-NEXT: vpminsq {{.*}}(%rip){1to8}, %zmm0, %zmm0
; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BW-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: vpmovqb %zmm0, %xmm0
; AVX512BW-NEXT: vmovd %xmm0, (%rdi)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
; AVX512BWVL-LABEL: trunc_packus_v4i64_v4i8_store:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsq %ymm1, %ymm0, %ymm0
; AVX512BWVL-NEXT: vpmovusqb %ymm0, (%rdi)
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
%1 = icmp slt <4 x i64> %a0, <i64 255, i64 255, i64 255, i64 255>
%2 = select <4 x i1> %1, <4 x i64> %a0, <4 x i64> <i64 255, i64 255, i64 255, i64 255>
%3 = icmp sgt <4 x i64> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i64> %2, <4 x i64> zeroinitializer
%5 = trunc <4 x i64> %4 to <4 x i8>
store <4 x i8> %5, <4 x i8> *%p1
ret void
}
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
define <8 x i8> @trunc_packus_v8i64_v8i8(<8 x i64> %a0) {
; SSE2-LABEL: trunc_packus_v8i64_v8i8:
; SSE2: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [255,255]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: movdqa %xmm2, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pxor %xmm10, %xmm5
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147483903,2147483903]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: movdqa %xmm9, %xmm7
; SSE2-NEXT: pcmpgtd %xmm5, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm7[0,0,2,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: pcmpeqd %xmm9, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pand %xmm6, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: por %xmm4, %xmm5
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pand %xmm5, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pandn %xmm8, %xmm5
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: por %xmm2, %xmm5
; SSE2-NEXT: movdqa %xmm3, %xmm2
; SSE2-NEXT: pxor %xmm10, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm9, %xmm4
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pcmpgtd %xmm2, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pcmpeqd %xmm9, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm6, %xmm7
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm2
; SSE2-NEXT: pand %xmm2, %xmm3
; SSE2-NEXT: pandn %xmm8, %xmm2
; SSE2-NEXT: por %xmm3, %xmm2
; SSE2-NEXT: movdqa %xmm0, %xmm3
; SSE2-NEXT: pxor %xmm10, %xmm3
; SSE2-NEXT: movdqa %xmm9, %xmm4
; SSE2-NEXT: pcmpgtd %xmm3, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm3
; SSE2-NEXT: pand %xmm3, %xmm0
; SSE2-NEXT: pandn %xmm8, %xmm3
; SSE2-NEXT: por %xmm0, %xmm3
; SSE2-NEXT: movdqa %xmm1, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pxor %xmm10, %xmm0
; SSE2-NEXT: movdqa %xmm9, %xmm4
; SSE2-NEXT: pcmpgtd %xmm0, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm4
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pandn %xmm8, %xmm4
; SSE2-NEXT: por %xmm1, %xmm4
; SSE2-NEXT: movdqa %xmm4, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pxor %xmm10, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pcmpgtd %xmm10, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pand %xmm6, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: pand %xmm4, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm3, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm4
; SSE2-NEXT: pcmpgtd %xmm10, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm10, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pand %xmm6, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: por %xmm7, %xmm0
; SSE2-NEXT: pand %xmm3, %xmm0
[X86][SSE] Consistently prefer lowering to PACKUS over PACKSS We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS. PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines. The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets. llvm-svn: 334276
2018-06-08 18:29:00 +08:00
; SSE2-NEXT: packuswb %xmm1, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: movdqa %xmm2, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: pcmpgtd %xmm10, %xmm3
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm10, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pand %xmm4, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm3
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE2-NEXT: pand %xmm2, %xmm3
; SSE2-NEXT: movdqa %xmm5, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm10, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm2
; SSE2-NEXT: pand %xmm5, %xmm2
[X86][SSE] Consistently prefer lowering to PACKUS over PACKSS We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS. PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines. The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets. llvm-svn: 334276
2018-06-08 18:29:00 +08:00
; SSE2-NEXT: packuswb %xmm3, %xmm2
; SSE2-NEXT: packuswb %xmm2, %xmm0
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; SSE2-NEXT: packuswb %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v8i64_v8i8:
; SSSE3: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [255,255]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: movdqa %xmm2, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pxor %xmm10, %xmm5
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147483903,2147483903]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: movdqa %xmm9, %xmm7
; SSSE3-NEXT: pcmpgtd %xmm5, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm7[0,0,2,2]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm5
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: por %xmm4, %xmm5
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pand %xmm5, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pandn %xmm8, %xmm5
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: por %xmm2, %xmm5
; SSSE3-NEXT: movdqa %xmm3, %xmm2
; SSSE3-NEXT: pxor %xmm10, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm9, %xmm4
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm7
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm2
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: pandn %xmm8, %xmm2
; SSSE3-NEXT: por %xmm3, %xmm2
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: pxor %xmm10, %xmm3
; SSSE3-NEXT: movdqa %xmm9, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm0
; SSSE3-NEXT: pandn %xmm8, %xmm3
; SSSE3-NEXT: por %xmm0, %xmm3
; SSSE3-NEXT: movdqa %xmm1, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pxor %xmm10, %xmm0
; SSSE3-NEXT: movdqa %xmm9, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm4
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: pandn %xmm8, %xmm4
; SSSE3-NEXT: por %xmm1, %xmm4
; SSSE3-NEXT: movdqa %xmm4, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pxor %xmm10, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm4, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: por %xmm7, %xmm0
; SSSE3-NEXT: pand %xmm3, %xmm0
[X86][SSE] Consistently prefer lowering to PACKUS over PACKSS We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS. PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines. The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets. llvm-svn: 334276
2018-06-08 18:29:00 +08:00
; SSSE3-NEXT: packuswb %xmm1, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: movdqa %xmm2, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pand %xmm4, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: por %xmm1, %xmm3
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: movdqa %xmm5, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm2
; SSSE3-NEXT: pand %xmm5, %xmm2
[X86][SSE] Consistently prefer lowering to PACKUS over PACKSS We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS. PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines. The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets. llvm-svn: 334276
2018-06-08 18:29:00 +08:00
; SSSE3-NEXT: packuswb %xmm3, %xmm2
; SSSE3-NEXT: packuswb %xmm2, %xmm0
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; SSSE3-NEXT: packuswb %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v8i64_v8i8:
; SSE41: # %bb.0:
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: movdqa %xmm0, %xmm9
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movapd {{.*#+}} xmm7 = [255,255]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm2, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: pxor %xmm10, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm4 = [2147483903,2147483903]
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm4, %xmm6
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: por %xmm6, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm8
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm8
; SSE41-NEXT: movdqa %xmm3, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm4, %xmm2
; SSE41-NEXT: pcmpeqd %xmm0, %xmm2
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpgtd %xmm0, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: por %xmm5, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm2
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: movdqa %xmm9, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm4, %xmm3
; SSE41-NEXT: pcmpeqd %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpgtd %xmm0, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
; SSE41-NEXT: pand %xmm3, %xmm0
; SSE41-NEXT: por %xmm5, %xmm0
; SSE41-NEXT: movapd %xmm7, %xmm6
; SSE41-NEXT: blendvpd %xmm0, %xmm9, %xmm6
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm4, %xmm3
; SSE41-NEXT: pcmpeqd %xmm0, %xmm3
; SSE41-NEXT: pcmpgtd %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm3, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm7
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pxor %xmm3, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm1
; SSE41-NEXT: xorpd %xmm10, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm4
; SSE41-NEXT: pcmpeqd %xmm10, %xmm4
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
; SSE41-NEXT: pand %xmm4, %xmm0
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: por %xmm1, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pxor %xmm4, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm4
; SSE41-NEXT: movapd %xmm6, %xmm1
; SSE41-NEXT: xorpd %xmm10, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: por %xmm1, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: pxor %xmm1, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm6, %xmm1
; SSE41-NEXT: packusdw %xmm4, %xmm1
; SSE41-NEXT: movapd %xmm2, %xmm4
; SSE41-NEXT: xorpd %xmm10, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: pxor %xmm4, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm8, %xmm2
; SSE41-NEXT: xorpd %xmm10, %xmm2
; SSE41-NEXT: movapd %xmm2, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm2, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm3
[X86][SSE] Consistently prefer lowering to PACKUS over PACKSS We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS. PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines. The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets. llvm-svn: 334276
2018-06-08 18:29:00 +08:00
; SSE41-NEXT: packusdw %xmm4, %xmm3
; SSE41-NEXT: packusdw %xmm3, %xmm1
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; SSE41-NEXT: packuswb %xmm1, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm1, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v8i64_v8i8:
; AVX1: # %bb.0:
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [255,255]
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm8
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm3, %xmm5
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm6, %xmm3, %xmm7
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm3, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm3, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm9
; AVX1-NEXT: vblendvpd %xmm7, %xmm6, %xmm3, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm6, %xmm7
; AVX1-NEXT: vblendvpd %xmm5, %xmm1, %xmm3, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm5
; AVX1-NEXT: vblendvpd %xmm8, %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm2, %xmm3
; AVX1-NEXT: vpand %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm5, %xmm1
[X86][SSE] Consistently prefer lowering to PACKUS over PACKSS We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS. PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines. The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets. llvm-svn: 334276
2018-06-08 18:29:00 +08:00
; AVX1-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vpand %xmm6, %xmm7, %xmm2
; AVX1-NEXT: vpand %xmm0, %xmm9, %xmm0
[X86][SSE] Consistently prefer lowering to PACKUS over PACKSS We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS. PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines. The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets. llvm-svn: 334276
2018-06-08 18:29:00 +08:00
; AVX1-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; AVX1-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-LABEL: trunc_packus_v8i64_v8i8:
; AVX2: # %bb.0:
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [255,255,255,255]
Revert "[X86] Enable -x86-experimental-vector-widening-legalization by default." This reverts commit 3de33245d2c992c9e0af60372043540b60f3a810. This commit broke the MSan buildbots. See https://reviews.llvm.org/rL367901 for more information. llvm-svn: 368107
2019-08-07 07:00:43 +08:00
; AVX2-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3
; AVX2-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; AVX2-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3
; AVX2-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; AVX2-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm3
; AVX2-NEXT: vpand %ymm0, %ymm3, %ymm0
; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm2
; AVX2-NEXT: vpand %ymm1, %ymm2, %ymm1
; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = <u,u,0,8,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX2-NEXT: vpshufb %xmm3, %xmm2, %xmm2
; AVX2-NEXT: vpshufb %xmm3, %xmm1, %xmm1
; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX2-NEXT: vpshufb %xmm3, %xmm2, %xmm2
; AVX2-NEXT: vpshufb %xmm3, %xmm0, %xmm0
; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
; AVX512-LABEL: trunc_packus_v8i64_v8i8:
; AVX512: # %bb.0:
; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
[X86] Add DAG combine to form saturating VTRUNCUS/VTRUNCS from VTRUNC We already do this for ISD::TRUNCATE, but we can do the same for X86ISD::VTRUNC Differential Revision: https://reviews.llvm.org/D68432 llvm-svn: 373765
2019-10-05 01:53:18 +08:00
; AVX512-NEXT: vpmovusqb %zmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
%1 = icmp slt <8 x i64> %a0, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255>
%2 = select <8 x i1> %1, <8 x i64> %a0, <8 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255>
%3 = icmp sgt <8 x i64> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer
%5 = trunc <8 x i64> %4 to <8 x i8>
ret <8 x i8> %5
}
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
define void @trunc_packus_v8i64_v8i8_store(<8 x i64> %a0, <8 x i8> *%p1) {
; SSE2-LABEL: trunc_packus_v8i64_v8i8_store:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [255,255]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: movdqa %xmm2, %xmm5
; SSE2-NEXT: pxor %xmm10, %xmm5
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm9 = [2147483903,2147483903]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: movdqa %xmm9, %xmm7
; SSE2-NEXT: pcmpgtd %xmm5, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm7[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
; SSE2-NEXT: por %xmm4, %xmm5
; SSE2-NEXT: pand %xmm5, %xmm2
; SSE2-NEXT: pandn %xmm8, %xmm5
; SSE2-NEXT: por %xmm2, %xmm5
; SSE2-NEXT: movdqa %xmm3, %xmm2
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: pxor %xmm10, %xmm2
; SSE2-NEXT: movdqa %xmm9, %xmm4
; SSE2-NEXT: pcmpgtd %xmm2, %xmm4
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: pcmpeqd %xmm9, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: pand %xmm6, %xmm7
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm2
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: pand %xmm2, %xmm3
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: pandn %xmm8, %xmm2
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: por %xmm3, %xmm2
; SSE2-NEXT: movdqa %xmm0, %xmm3
; SSE2-NEXT: pxor %xmm10, %xmm3
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: movdqa %xmm9, %xmm4
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: pcmpgtd %xmm3, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm3
; SSE2-NEXT: pand %xmm3, %xmm0
; SSE2-NEXT: pandn %xmm8, %xmm3
; SSE2-NEXT: por %xmm0, %xmm3
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: movdqa %xmm9, %xmm4
; SSE2-NEXT: pcmpgtd %xmm0, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm9, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm4
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pandn %xmm8, %xmm4
; SSE2-NEXT: por %xmm1, %xmm4
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: pcmpgtd %xmm10, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: movdqa %xmm3, %xmm0
; SSE2-NEXT: pxor %xmm10, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm4
; SSE2-NEXT: pcmpgtd %xmm10, %xmm4
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: pcmpeqd %xmm10, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: pand %xmm6, %xmm7
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE2-NEXT: pand %xmm3, %xmm0
; SSE2-NEXT: packuswb %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: pcmpgtd %xmm10, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm3
; SSE2-NEXT: pand %xmm2, %xmm3
; SSE2-NEXT: movdqa %xmm5, %xmm1
; SSE2-NEXT: pxor %xmm10, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm10, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm10, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm2
; SSE2-NEXT: pand %xmm5, %xmm2
; SSE2-NEXT: packuswb %xmm3, %xmm2
; SSE2-NEXT: packuswb %xmm2, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE2-NEXT: packuswb %xmm0, %xmm0
; SSE2-NEXT: movq %xmm0, (%rdi)
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v8i64_v8i8_store:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [255,255]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: movdqa %xmm2, %xmm5
; SSSE3-NEXT: pxor %xmm10, %xmm5
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm9 = [2147483903,2147483903]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: movdqa %xmm9, %xmm7
; SSSE3-NEXT: pcmpgtd %xmm5, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm7[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm5
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm5
; SSSE3-NEXT: pand %xmm5, %xmm2
; SSSE3-NEXT: pandn %xmm8, %xmm5
; SSSE3-NEXT: por %xmm2, %xmm5
; SSSE3-NEXT: movdqa %xmm3, %xmm2
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: pxor %xmm10, %xmm2
; SSSE3-NEXT: movdqa %xmm9, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm7
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm2
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: pand %xmm2, %xmm3
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: pandn %xmm8, %xmm2
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: por %xmm3, %xmm2
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: pxor %xmm10, %xmm3
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: movdqa %xmm9, %xmm4
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm0
; SSSE3-NEXT: pandn %xmm8, %xmm3
; SSSE3-NEXT: por %xmm0, %xmm3
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: movdqa %xmm9, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm9, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm4
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: pandn %xmm8, %xmm4
; SSSE3-NEXT: por %xmm1, %xmm4
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: pxor %xmm10, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm4
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: pand %xmm6, %xmm7
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSSE3-NEXT: pand %xmm3, %xmm0
; SSSE3-NEXT: packuswb %xmm1, %xmm0
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm3
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: movdqa %xmm5, %xmm1
; SSSE3-NEXT: pxor %xmm10, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm10, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm10, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm2
; SSSE3-NEXT: pand %xmm5, %xmm2
; SSSE3-NEXT: packuswb %xmm3, %xmm2
; SSSE3-NEXT: packuswb %xmm2, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSSE3-NEXT: packuswb %xmm0, %xmm0
; SSSE3-NEXT: movq %xmm0, (%rdi)
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v8i64_v8i8_store:
; SSE41: # %bb.0:
; SSE41-NEXT: movdqa %xmm0, %xmm9
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: movapd {{.*#+}} xmm7 = [255,255]
; SSE41-NEXT: movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE41-NEXT: movdqa %xmm2, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pxor %xmm10, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm4 = [2147483903,2147483903]
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: movdqa %xmm4, %xmm6
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
; SSE41-NEXT: movapd %xmm7, %xmm8
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm8
; SSE41-NEXT: movdqa %xmm3, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm2
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpgtd %xmm0, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: por %xmm5, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm2
; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm2
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE41-NEXT: movdqa %xmm9, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pxor %xmm10, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm4, %xmm5
; SSE41-NEXT: pcmpgtd %xmm0, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
; SSE41-NEXT: pand %xmm3, %xmm0
; SSE41-NEXT: por %xmm5, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm3
; SSE41-NEXT: blendvpd %xmm0, %xmm9, %xmm3
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm10, %xmm0
; SSE41-NEXT: movdqa %xmm4, %xmm5
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: pcmpgtd %xmm0, %xmm4
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm7
; SSE41-NEXT: xorpd %xmm1, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm7, %xmm4
; SSE41-NEXT: xorpd %xmm10, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE41-NEXT: pxor %xmm5, %xmm5
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm5
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm3, %xmm4
; SSE41-NEXT: xorpd %xmm10, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm6
; SSE41-NEXT: pcmpeqd %xmm10, %xmm6
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pand %xmm6, %xmm0
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: por %xmm4, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE41-NEXT: pxor %xmm4, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm4
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: packusdw %xmm5, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm2, %xmm3
; SSE41-NEXT: xorpd %xmm10, %xmm3
; SSE41-NEXT: movapd %xmm3, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: por %xmm3, %xmm0
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pxor %xmm3, %xmm3
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm8, %xmm2
; SSE41-NEXT: xorpd %xmm10, %xmm2
; SSE41-NEXT: movapd %xmm2, %xmm5
; SSE41-NEXT: pcmpeqd %xmm10, %xmm5
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pcmpgtd %xmm10, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm2, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm1
; SSE41-NEXT: packusdw %xmm3, %xmm1
; SSE41-NEXT: packusdw %xmm1, %xmm4
[X86][SSE] Call SimplifyMultipleUseDemandedBits on PACKSS/PACKUS arguments. This mainly helps to replace unused arguments with UNDEF in the case where they have multiple users. llvm-svn: 368026
2019-08-06 21:10:42 +08:00
; SSE41-NEXT: packuswb %xmm0, %xmm4
[X86] combineVectorTruncationWithPACKUS - remove split/concatenation of mask combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together. This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356). This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit. My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great. Differential Revision: https://reviews.llvm.org/D60375#inline-539623 llvm-svn: 358692
2019-04-19 01:23:09 +08:00
; SSE41-NEXT: movq %xmm4, (%rdi)
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v8i64_v8i8_store:
; AVX1: # %bb.0:
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [255,255]
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm8
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm3, %xmm5
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm6, %xmm3, %xmm7
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm3, %xmm4
; AVX1-NEXT: vblendvpd %xmm4, %xmm0, %xmm3, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm0, %xmm9
; AVX1-NEXT: vblendvpd %xmm7, %xmm6, %xmm3, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm6, %xmm7
; AVX1-NEXT: vblendvpd %xmm5, %xmm1, %xmm3, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm1, %xmm5
; AVX1-NEXT: vblendvpd %xmm8, %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm2, %xmm3
; AVX1-NEXT: vpand %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpand %xmm1, %xmm5, %xmm1
[X86][SSE] Simplify combineVectorTruncationWithPACKUS to reduce code duplication Simplify combineVectorTruncationWithPACKUS to mask the upper bits followed by calling truncateVectorWithPACK instead of duplicating with similar code. This results in the codegen using (V)PACKUSDW on SSE41+ targets for vXi64/vXi32 inputs where before it always used PACKUSWB (along with a lot more bitcasting). I've raised PR37749 as until we avoid unnecessary concats back to 256-bit for bitwise ops, we can't avoid splitting the input value into 128-bit subvectors for masking. llvm-svn: 334289
2018-06-08 21:59:11 +08:00
; AVX1-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vpand %xmm6, %xmm7, %xmm2
; AVX1-NEXT: vpand %xmm0, %xmm9, %xmm0
[X86][SSE] Simplify combineVectorTruncationWithPACKUS to reduce code duplication Simplify combineVectorTruncationWithPACKUS to mask the upper bits followed by calling truncateVectorWithPACK instead of duplicating with similar code. This results in the codegen using (V)PACKUSDW on SSE41+ targets for vXi64/vXi32 inputs where before it always used PACKUSWB (along with a lot more bitcasting). I've raised PR37749 as until we avoid unnecessary concats back to 256-bit for bitwise ops, we can't avoid splitting the input value into 128-bit subvectors for masking. llvm-svn: 334289
2018-06-08 21:59:11 +08:00
; AVX1-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; AVX1-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX1-NEXT: vmovq %xmm0, (%rdi)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-LABEL: trunc_packus_v8i64_v8i8_store:
; AVX2: # %bb.0:
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [255,255,255,255]
; AVX2-NEXT: vpcmpgtq %ymm1, %ymm2, %ymm3
; AVX2-NEXT: vblendvpd %ymm3, %ymm1, %ymm2, %ymm1
; AVX2-NEXT: vpcmpgtq %ymm0, %ymm2, %ymm3
; AVX2-NEXT: vblendvpd %ymm3, %ymm0, %ymm2, %ymm0
; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
; AVX2-NEXT: vpcmpgtq %ymm2, %ymm0, %ymm3
; AVX2-NEXT: vpand %ymm0, %ymm3, %ymm0
; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm2
; AVX2-NEXT: vpand %ymm1, %ymm2, %ymm1
; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = <u,u,0,8,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX2-NEXT: vpshufb %xmm3, %xmm2, %xmm2
; AVX2-NEXT: vpshufb %xmm3, %xmm1, %xmm1
; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX2-NEXT: vpshufb %xmm3, %xmm2, %xmm2
; AVX2-NEXT: vpshufb %xmm3, %xmm0, %xmm0
; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-NEXT: vmovq %xmm0, (%rdi)
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
;
; AVX512-LABEL: trunc_packus_v8i64_v8i8_store:
; AVX512: # %bb.0:
; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512-NEXT: vpmovusqb %zmm0, (%rdi)
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
%1 = icmp slt <8 x i64> %a0, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255>
%2 = select <8 x i1> %1, <8 x i64> %a0, <8 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255>
%3 = icmp sgt <8 x i64> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i64> %2, <8 x i64> zeroinitializer
%5 = trunc <8 x i64> %4 to <8 x i8>
store <8 x i8> %5, <8 x i8> *%p1
ret void
}
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
define <16 x i8> @trunc_packus_v16i64_v16i8(<16 x i64> %a0) {
; SSE2-LABEL: trunc_packus_v16i64_v16i8:
; SSE2: # %bb.0:
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm10 = [255,255]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [2147483648,2147483648]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: movdqa %xmm6, %xmm9
; SSE2-NEXT: pxor %xmm8, %xmm9
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm11 = [2147483903,2147483903]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: movdqa %xmm11, %xmm12
; SSE2-NEXT: pcmpgtd %xmm9, %xmm12
; SSE2-NEXT: pshufd {{.*#+}} xmm13 = xmm12[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm11, %xmm9
; SSE2-NEXT: pshufd {{.*#+}} xmm14 = xmm9[1,1,3,3]
; SSE2-NEXT: pand %xmm13, %xmm14
; SSE2-NEXT: pshufd {{.*#+}} xmm9 = xmm12[1,1,3,3]
; SSE2-NEXT: por %xmm14, %xmm9
; SSE2-NEXT: pand %xmm9, %xmm6
; SSE2-NEXT: pandn %xmm10, %xmm9
; SSE2-NEXT: por %xmm6, %xmm9
; SSE2-NEXT: movdqa %xmm7, %xmm6
; SSE2-NEXT: pxor %xmm8, %xmm6
; SSE2-NEXT: movdqa %xmm11, %xmm12
; SSE2-NEXT: pcmpgtd %xmm6, %xmm12
; SSE2-NEXT: pshufd {{.*#+}} xmm13 = xmm12[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm11, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
; SSE2-NEXT: pand %xmm13, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm12 = xmm12[1,1,3,3]
; SSE2-NEXT: por %xmm6, %xmm12
; SSE2-NEXT: pand %xmm12, %xmm7
; SSE2-NEXT: pandn %xmm10, %xmm12
; SSE2-NEXT: por %xmm7, %xmm12
; SSE2-NEXT: movdqa %xmm4, %xmm6
; SSE2-NEXT: pxor %xmm8, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm11, %xmm7
; SSE2-NEXT: pcmpgtd %xmm6, %xmm7
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm13 = xmm7[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm11, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pand %xmm13, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm13 = xmm7[1,1,3,3]
; SSE2-NEXT: por %xmm6, %xmm13
; SSE2-NEXT: pand %xmm13, %xmm4
; SSE2-NEXT: pandn %xmm10, %xmm13
; SSE2-NEXT: por %xmm4, %xmm13
; SSE2-NEXT: movdqa %xmm5, %xmm4
; SSE2-NEXT: pxor %xmm8, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm11, %xmm6
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pcmpgtd %xmm4, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pcmpeqd %xmm11, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: pand %xmm7, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm14 = xmm6[1,1,3,3]
; SSE2-NEXT: por %xmm4, %xmm14
; SSE2-NEXT: pand %xmm14, %xmm5
; SSE2-NEXT: pandn %xmm10, %xmm14
; SSE2-NEXT: por %xmm5, %xmm14
; SSE2-NEXT: movdqa %xmm2, %xmm4
; SSE2-NEXT: pxor %xmm8, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm11, %xmm5
; SSE2-NEXT: pcmpgtd %xmm4, %xmm5
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm11, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pand %xmm7, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
; SSE2-NEXT: por %xmm4, %xmm5
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pand %xmm5, %xmm2
; SSE2-NEXT: pandn %xmm10, %xmm5
; SSE2-NEXT: por %xmm2, %xmm5
; SSE2-NEXT: movdqa %xmm3, %xmm2
; SSE2-NEXT: pxor %xmm8, %xmm2
; SSE2-NEXT: movdqa %xmm11, %xmm4
; SSE2-NEXT: pcmpgtd %xmm2, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm11, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm2[1,1,3,3]
; SSE2-NEXT: pand %xmm7, %xmm6
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm6, %xmm2
; SSE2-NEXT: pand %xmm2, %xmm3
; SSE2-NEXT: pandn %xmm10, %xmm2
; SSE2-NEXT: por %xmm3, %xmm2
; SSE2-NEXT: movdqa %xmm0, %xmm3
; SSE2-NEXT: pxor %xmm8, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm11, %xmm4
; SSE2-NEXT: pcmpgtd %xmm3, %xmm4
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pcmpeqd %xmm11, %xmm3
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm3
; SSE2-NEXT: pand %xmm3, %xmm0
; SSE2-NEXT: pandn %xmm10, %xmm3
; SSE2-NEXT: por %xmm0, %xmm3
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: pxor %xmm8, %xmm0
; SSE2-NEXT: movdqa %xmm11, %xmm4
; SSE2-NEXT: pcmpgtd %xmm0, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm11, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pand %xmm4, %xmm1
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pandn %xmm10, %xmm4
; SSE2-NEXT: por %xmm1, %xmm4
; SSE2-NEXT: movdqa %xmm4, %xmm0
; SSE2-NEXT: pxor %xmm8, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pcmpgtd %xmm8, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm8, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pand %xmm6, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: movdqa %xmm3, %xmm0
; SSE2-NEXT: pxor %xmm8, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm4
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pcmpgtd %xmm8, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm8, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
; SSE2-NEXT: pand %xmm6, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm7, %xmm0
; SSE2-NEXT: pand %xmm3, %xmm0
; SSE2-NEXT: packuswb %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: pxor %xmm8, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: pcmpgtd %xmm8, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm8, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE2-NEXT: por %xmm1, %xmm3
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE2-NEXT: pand %xmm2, %xmm3
; SSE2-NEXT: movdqa %xmm5, %xmm1
; SSE2-NEXT: pxor %xmm8, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm8, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm8, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm2
; SSE2-NEXT: pand %xmm5, %xmm2
; SSE2-NEXT: packuswb %xmm3, %xmm2
; SSE2-NEXT: packuswb %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm14, %xmm1
; SSE2-NEXT: pxor %xmm8, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm8, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm8, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm3, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: por %xmm1, %xmm2
; SSE2-NEXT: pand %xmm14, %xmm2
; SSE2-NEXT: movdqa %xmm13, %xmm1
; SSE2-NEXT: pxor %xmm8, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: pcmpgtd %xmm8, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm8, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
; SSE2-NEXT: por %xmm5, %xmm1
; SSE2-NEXT: pand %xmm13, %xmm1
; SSE2-NEXT: packuswb %xmm2, %xmm1
; SSE2-NEXT: movdqa %xmm12, %xmm2
; SSE2-NEXT: pxor %xmm8, %xmm2
; SSE2-NEXT: movdqa %xmm2, %xmm3
; SSE2-NEXT: pcmpgtd %xmm8, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm8, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: pand %xmm4, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSE2-NEXT: por %xmm2, %xmm3
; SSE2-NEXT: pand %xmm12, %xmm3
; SSE2-NEXT: movdqa %xmm9, %xmm2
; SSE2-NEXT: pxor %xmm8, %xmm2
; SSE2-NEXT: movdqa %xmm2, %xmm4
; SSE2-NEXT: pcmpgtd %xmm8, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
; SSE2-NEXT: pcmpeqd %xmm8, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE2-NEXT: pand %xmm5, %xmm2
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSE2-NEXT: por %xmm2, %xmm4
; SSE2-NEXT: pand %xmm9, %xmm4
; SSE2-NEXT: packuswb %xmm3, %xmm4
; SSE2-NEXT: packuswb %xmm4, %xmm1
; SSE2-NEXT: packuswb %xmm1, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v16i64_v16i8:
; SSSE3: # %bb.0:
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm10 = [255,255]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm8 = [2147483648,2147483648]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: movdqa %xmm6, %xmm9
; SSSE3-NEXT: pxor %xmm8, %xmm9
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm11 = [2147483903,2147483903]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: movdqa %xmm11, %xmm12
; SSSE3-NEXT: pcmpgtd %xmm9, %xmm12
; SSSE3-NEXT: pshufd {{.*#+}} xmm13 = xmm12[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm11, %xmm9
; SSSE3-NEXT: pshufd {{.*#+}} xmm14 = xmm9[1,1,3,3]
; SSSE3-NEXT: pand %xmm13, %xmm14
; SSSE3-NEXT: pshufd {{.*#+}} xmm9 = xmm12[1,1,3,3]
; SSSE3-NEXT: por %xmm14, %xmm9
; SSSE3-NEXT: pand %xmm9, %xmm6
; SSSE3-NEXT: pandn %xmm10, %xmm9
; SSSE3-NEXT: por %xmm6, %xmm9
; SSSE3-NEXT: movdqa %xmm7, %xmm6
; SSSE3-NEXT: pxor %xmm8, %xmm6
; SSSE3-NEXT: movdqa %xmm11, %xmm12
; SSSE3-NEXT: pcmpgtd %xmm6, %xmm12
; SSSE3-NEXT: pshufd {{.*#+}} xmm13 = xmm12[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm11, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
; SSSE3-NEXT: pand %xmm13, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm12 = xmm12[1,1,3,3]
; SSSE3-NEXT: por %xmm6, %xmm12
; SSSE3-NEXT: pand %xmm12, %xmm7
; SSSE3-NEXT: pandn %xmm10, %xmm12
; SSSE3-NEXT: por %xmm7, %xmm12
; SSSE3-NEXT: movdqa %xmm4, %xmm6
; SSSE3-NEXT: pxor %xmm8, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm11, %xmm7
; SSSE3-NEXT: pcmpgtd %xmm6, %xmm7
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm13 = xmm7[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm11, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pand %xmm13, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm13 = xmm7[1,1,3,3]
; SSSE3-NEXT: por %xmm6, %xmm13
; SSSE3-NEXT: pand %xmm13, %xmm4
; SSSE3-NEXT: pandn %xmm10, %xmm13
; SSSE3-NEXT: por %xmm4, %xmm13
; SSSE3-NEXT: movdqa %xmm5, %xmm4
; SSSE3-NEXT: pxor %xmm8, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm11, %xmm6
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pcmpgtd %xmm4, %xmm6
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pcmpeqd %xmm11, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: pand %xmm7, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm14 = xmm6[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm14
; SSSE3-NEXT: pand %xmm14, %xmm5
; SSSE3-NEXT: pandn %xmm10, %xmm14
; SSSE3-NEXT: por %xmm5, %xmm14
; SSSE3-NEXT: movdqa %xmm2, %xmm4
; SSSE3-NEXT: pxor %xmm8, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm11, %xmm5
; SSSE3-NEXT: pcmpgtd %xmm4, %xmm5
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm11, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pand %xmm7, %xmm4
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
; SSSE3-NEXT: por %xmm4, %xmm5
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pand %xmm5, %xmm2
; SSSE3-NEXT: pandn %xmm10, %xmm5
; SSSE3-NEXT: por %xmm2, %xmm5
; SSSE3-NEXT: movdqa %xmm3, %xmm2
; SSSE3-NEXT: pxor %xmm8, %xmm2
; SSSE3-NEXT: movdqa %xmm11, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm11, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm2[1,1,3,3]
; SSSE3-NEXT: pand %xmm7, %xmm6
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm6, %xmm2
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: pandn %xmm10, %xmm2
; SSSE3-NEXT: por %xmm3, %xmm2
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: pxor %xmm8, %xmm3
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm11, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm3, %xmm4
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pcmpeqd %xmm11, %xmm3
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm3
; SSSE3-NEXT: pand %xmm3, %xmm0
; SSSE3-NEXT: pandn %xmm10, %xmm3
; SSSE3-NEXT: por %xmm0, %xmm3
; SSSE3-NEXT: movdqa %xmm1, %xmm0
; SSSE3-NEXT: pxor %xmm8, %xmm0
; SSSE3-NEXT: movdqa %xmm11, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm11, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm4
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: pandn %xmm10, %xmm4
; SSSE3-NEXT: por %xmm1, %xmm4
; SSSE3-NEXT: movdqa %xmm4, %xmm0
; SSSE3-NEXT: pxor %xmm8, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm8, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm8, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: por %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm4, %xmm1
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: pxor %xmm8, %xmm0
; SSSE3-NEXT: movdqa %xmm0, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm8, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm8, %xmm0
; SSSE3-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
; SSSE3-NEXT: pand %xmm6, %xmm7
; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm7, %xmm0
; SSSE3-NEXT: pand %xmm3, %xmm0
; SSSE3-NEXT: packuswb %xmm1, %xmm0
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: pxor %xmm8, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm8, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm8, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: por %xmm1, %xmm3
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: movdqa %xmm5, %xmm1
; SSSE3-NEXT: pxor %xmm8, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: movdqa %xmm1, %xmm2
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pcmpgtd %xmm8, %xmm2
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pcmpeqd %xmm8, %xmm1
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pand %xmm4, %xmm1
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm2
; SSSE3-NEXT: pand %xmm5, %xmm2
; SSSE3-NEXT: packuswb %xmm3, %xmm2
; SSSE3-NEXT: packuswb %xmm2, %xmm0
; SSSE3-NEXT: movdqa %xmm14, %xmm1
; SSSE3-NEXT: pxor %xmm8, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm8, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm8, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm3, %xmm1
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSSE3-NEXT: por %xmm1, %xmm2
; SSSE3-NEXT: pand %xmm14, %xmm2
; SSSE3-NEXT: movdqa %xmm13, %xmm1
; SSSE3-NEXT: pxor %xmm8, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm8, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm8, %xmm1
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm5
; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
; SSSE3-NEXT: por %xmm5, %xmm1
; SSSE3-NEXT: pand %xmm13, %xmm1
; SSSE3-NEXT: packuswb %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm12, %xmm2
; SSSE3-NEXT: pxor %xmm8, %xmm2
; SSSE3-NEXT: movdqa %xmm2, %xmm3
; SSSE3-NEXT: pcmpgtd %xmm8, %xmm3
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm8, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSSE3-NEXT: pand %xmm4, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSSE3-NEXT: por %xmm2, %xmm3
; SSSE3-NEXT: pand %xmm12, %xmm3
; SSSE3-NEXT: movdqa %xmm9, %xmm2
; SSSE3-NEXT: pxor %xmm8, %xmm2
; SSSE3-NEXT: movdqa %xmm2, %xmm4
; SSSE3-NEXT: pcmpgtd %xmm8, %xmm4
; SSSE3-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
; SSSE3-NEXT: pcmpeqd %xmm8, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSSE3-NEXT: pand %xmm5, %xmm2
; SSSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
; SSSE3-NEXT: por %xmm2, %xmm4
; SSSE3-NEXT: pand %xmm9, %xmm4
; SSSE3-NEXT: packuswb %xmm3, %xmm4
; SSSE3-NEXT: packuswb %xmm4, %xmm1
; SSSE3-NEXT: packuswb %xmm1, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v16i64_v16i8:
; SSE41: # %bb.0:
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm0, %xmm8
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movapd {{.*#+}} xmm11 = [255,255]
[X86] When lowering unsigned v2i64 setcc without SSE42, flip the sign bits in the v2i64 type then bitcast to v4i32. This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step. Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast. There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it. This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues. llvm-svn: 344071
2018-10-10 03:05:50 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm9 = [2147483648,2147483648]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movdqa %xmm6, %xmm0
; SSE41-NEXT: pxor %xmm9, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm12 = [2147483903,2147483903]
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movdqa %xmm12, %xmm10
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm10
; SSE41-NEXT: movdqa %xmm12, %xmm13
; SSE41-NEXT: pcmpgtd %xmm0, %xmm13
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm13[0,0,2,2]
; SSE41-NEXT: pand %xmm10, %xmm0
; SSE41-NEXT: por %xmm13, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movapd %xmm11, %xmm10
; SSE41-NEXT: blendvpd %xmm0, %xmm6, %xmm10
; SSE41-NEXT: movdqa %xmm7, %xmm0
; SSE41-NEXT: pxor %xmm9, %xmm0
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movdqa %xmm12, %xmm13
; SSE41-NEXT: pcmpeqd %xmm0, %xmm13
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movdqa %xmm12, %xmm6
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm13, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movapd %xmm11, %xmm13
; SSE41-NEXT: blendvpd %xmm0, %xmm7, %xmm13
; SSE41-NEXT: movdqa %xmm4, %xmm0
; SSE41-NEXT: pxor %xmm9, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm12, %xmm6
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm6
; SSE41-NEXT: movdqa %xmm12, %xmm7
; SSE41-NEXT: pcmpgtd %xmm0, %xmm7
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm7, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movapd %xmm11, %xmm14
; SSE41-NEXT: blendvpd %xmm0, %xmm4, %xmm14
; SSE41-NEXT: movdqa %xmm5, %xmm0
; SSE41-NEXT: pxor %xmm9, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm12, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm4
; SSE41-NEXT: movdqa %xmm12, %xmm6
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm4, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movapd %xmm11, %xmm15
; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm15
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm2, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: pxor %xmm9, %xmm0
; SSE41-NEXT: movdqa %xmm12, %xmm5
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm12, %xmm6
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movapd %xmm11, %xmm5
; SSE41-NEXT: blendvpd %xmm0, %xmm2, %xmm5
; SSE41-NEXT: movdqa %xmm3, %xmm0
; SSE41-NEXT: pxor %xmm9, %xmm0
; SSE41-NEXT: movdqa %xmm12, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm2
; SSE41-NEXT: movdqa %xmm12, %xmm6
; SSE41-NEXT: pcmpgtd %xmm0, %xmm6
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: por %xmm6, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movapd %xmm11, %xmm6
; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm6
; SSE41-NEXT: movdqa %xmm8, %xmm0
; SSE41-NEXT: pxor %xmm9, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm12, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm2
; SSE41-NEXT: movdqa %xmm12, %xmm3
; SSE41-NEXT: pcmpgtd %xmm0, %xmm3
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: por %xmm3, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: movapd %xmm11, %xmm3
; SSE41-NEXT: blendvpd %xmm0, %xmm8, %xmm3
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm9, %xmm0
; SSE41-NEXT: movdqa %xmm12, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pcmpeqd %xmm0, %xmm2
; SSE41-NEXT: pcmpgtd %xmm0, %xmm12
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm12[0,0,2,2]
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: por %xmm12, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm1, %xmm11
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pxor %xmm2, %xmm2
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm11, %xmm1
; SSE41-NEXT: xorpd %xmm9, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm7
; SSE41-NEXT: pcmpeqd %xmm9, %xmm7
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: pcmpgtd %xmm9, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm7, %xmm0
; SSE41-NEXT: por %xmm1, %xmm0
; SSE41-NEXT: pxor %xmm7, %xmm7
; SSE41-NEXT: blendvpd %xmm0, %xmm11, %xmm7
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm3, %xmm1
; SSE41-NEXT: xorpd %xmm9, %xmm1
; SSE41-NEXT: movapd %xmm1, %xmm4
; SSE41-NEXT: pcmpeqd %xmm9, %xmm4
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: pcmpgtd %xmm9, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm4, %xmm0
; SSE41-NEXT: por %xmm1, %xmm0
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: pxor %xmm1, %xmm1
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm3, %xmm1
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: packusdw %xmm7, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm6, %xmm3
; SSE41-NEXT: xorpd %xmm9, %xmm3
; SSE41-NEXT: movapd %xmm3, %xmm4
; SSE41-NEXT: pcmpeqd %xmm9, %xmm4
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: pcmpgtd %xmm9, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm4, %xmm0
; SSE41-NEXT: por %xmm3, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: pxor %xmm3, %xmm3
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: blendvpd %xmm0, %xmm6, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm5, %xmm4
; SSE41-NEXT: xorpd %xmm9, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm6
; SSE41-NEXT: pcmpeqd %xmm9, %xmm6
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pcmpgtd %xmm9, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm6, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: por %xmm4, %xmm0
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pxor %xmm4, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm5, %xmm4
; SSE41-NEXT: packusdw %xmm3, %xmm4
; SSE41-NEXT: packusdw %xmm4, %xmm1
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm15, %xmm3
; SSE41-NEXT: xorpd %xmm9, %xmm3
; SSE41-NEXT: movapd %xmm3, %xmm4
; SSE41-NEXT: pcmpeqd %xmm9, %xmm4
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: pcmpgtd %xmm9, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm4, %xmm0
; SSE41-NEXT: por %xmm3, %xmm0
; SSE41-NEXT: pxor %xmm4, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm15, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm14, %xmm3
; SSE41-NEXT: xorpd %xmm9, %xmm3
; SSE41-NEXT: movapd %xmm3, %xmm5
; SSE41-NEXT: pcmpeqd %xmm9, %xmm5
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pcmpgtd %xmm9, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm3, %xmm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: pxor %xmm3, %xmm3
; SSE41-NEXT: blendvpd %xmm0, %xmm14, %xmm3
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: packusdw %xmm4, %xmm3
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm13, %xmm4
; SSE41-NEXT: xorpd %xmm9, %xmm4
; SSE41-NEXT: movapd %xmm4, %xmm5
; SSE41-NEXT: pcmpeqd %xmm9, %xmm5
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: pcmpgtd %xmm9, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm5, %xmm0
; SSE41-NEXT: por %xmm4, %xmm0
; SSE41-NEXT: pxor %xmm4, %xmm4
; SSE41-NEXT: blendvpd %xmm0, %xmm13, %xmm4
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: movapd %xmm10, %xmm5
; SSE41-NEXT: xorpd %xmm9, %xmm5
; SSE41-NEXT: movapd %xmm5, %xmm6
; SSE41-NEXT: pcmpeqd %xmm9, %xmm6
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pcmpgtd %xmm9, %xmm5
[TargetLowering] SimplifyMultipleUseDemandedBits - add VECTOR_SHUFFLE support. If all the demanded elts are from one operand and are inline, then we can use the operand directly. The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements. llvm-svn: 366817
2019-07-23 23:35:55 +08:00
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
[TargetLowering] Add ISD::OR + ISD::XOR handling to SimplifyDemandedVectorElts Differential Revision: https://reviews.llvm.org/D55600 llvm-svn: 349264
2018-12-15 19:36:36 +08:00
; SSE41-NEXT: pand %xmm6, %xmm0
; SSE41-NEXT: por %xmm5, %xmm0
; SSE41-NEXT: blendvpd %xmm0, %xmm10, %xmm2
; SSE41-NEXT: packusdw %xmm4, %xmm2
; SSE41-NEXT: packusdw %xmm2, %xmm3
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; SSE41-NEXT: packuswb %xmm3, %xmm1
[X86][SSE] Enable SMIN/SMAX/UMIN/UMAX custom lowering for all legal types This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT. Differential Revision: https://reviews.llvm.org/D43014 llvm-svn: 324837
2018-02-11 18:52:37 +08:00
; SSE41-NEXT: movdqa %xmm1, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v16i64_v16i8:
; AVX1: # %bb.0:
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm8
; AVX1-NEXT: vmovdqa {{.*#+}} xmm5 = [255,255]
; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm9
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm7
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm5, %xmm6
; AVX1-NEXT: vblendvpd %xmm6, %xmm0, %xmm5, %xmm10
; AVX1-NEXT: vpcmpgtq %xmm4, %xmm5, %xmm6
; AVX1-NEXT: vblendvpd %xmm6, %xmm4, %xmm5, %xmm11
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm5, %xmm6
; AVX1-NEXT: vblendvpd %xmm6, %xmm1, %xmm5, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm7, %xmm5, %xmm6
; AVX1-NEXT: vblendvpd %xmm6, %xmm7, %xmm5, %xmm6
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm5, %xmm7
; AVX1-NEXT: vblendvpd %xmm7, %xmm2, %xmm5, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm9, %xmm5, %xmm7
; AVX1-NEXT: vblendvpd %xmm7, %xmm9, %xmm5, %xmm7
; AVX1-NEXT: vpcmpgtq %xmm3, %xmm5, %xmm0
; AVX1-NEXT: vblendvpd %xmm0, %xmm3, %xmm5, %xmm0
; AVX1-NEXT: vpcmpgtq %xmm8, %xmm5, %xmm3
; AVX1-NEXT: vblendvpd %xmm3, %xmm8, %xmm5, %xmm3
; AVX1-NEXT: vpxor %xmm5, %xmm5, %xmm5
; AVX1-NEXT: vpcmpgtq %xmm5, %xmm3, %xmm4
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; AVX1-NEXT: vpand %xmm3, %xmm4, %xmm3
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vpcmpgtq %xmm5, %xmm0, %xmm4
; AVX1-NEXT: vpand %xmm0, %xmm4, %xmm0
[X86][SSE] Consistently prefer lowering to PACKUS over PACKSS We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS. PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines. The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets. llvm-svn: 334276
2018-06-08 18:29:00 +08:00
; AVX1-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
[x86] narrow extract subvector of vector select This is a potentially large perf win for AVX1 targets because of the way we auto-vectorize to 256-bit but then expect the backend to legalize/optimize for the half-implemented AVX1 ISA. On the motivating example from PR37428 (even though this patch doesn't solve the vector shift issue): https://bugs.llvm.org/show_bug.cgi?id=37428 ...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell) because we eliminate the remaining 256-bit vblendv ops. I added comments on a couple of tests that require further work. If we have 256-bit logic ops separating the vselect and extract, we should probably narrow everything to 128-bit, but that requires a larger pattern match. Differential Revision: https://reviews.llvm.org/D62969 llvm-svn: 362797
2019-06-07 21:17:46 +08:00
; AVX1-NEXT: vpcmpgtq %xmm5, %xmm7, %xmm3
; AVX1-NEXT: vpand %xmm7, %xmm3, %xmm3
; AVX1-NEXT: vpcmpgtq %xmm5, %xmm2, %xmm4
; AVX1-NEXT: vpand %xmm2, %xmm4, %xmm2
; AVX1-NEXT: vpackusdw %xmm3, %xmm2, %xmm2
; AVX1-NEXT: vpackusdw %xmm0, %xmm2, %xmm0
; AVX1-NEXT: vpcmpgtq %xmm5, %xmm6, %xmm2
; AVX1-NEXT: vpand %xmm6, %xmm2, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm5, %xmm1, %xmm3
; AVX1-NEXT: vpand %xmm1, %xmm3, %xmm1
; AVX1-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm5, %xmm11, %xmm2
; AVX1-NEXT: vpand %xmm11, %xmm2, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm5, %xmm10, %xmm3
; AVX1-NEXT: vpand %xmm10, %xmm3, %xmm3
; AVX1-NEXT: vpackusdw %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpackusdw %xmm1, %xmm2, %xmm1
; AVX1-NEXT: vpackuswb %xmm0, %xmm1, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-LABEL: trunc_packus_v16i64_v16i8:
; AVX2: # %bb.0:
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm4 = [255,255,255,255]
; AVX2-NEXT: vpcmpgtq %ymm2, %ymm4, %ymm5
; AVX2-NEXT: vblendvpd %ymm5, %ymm2, %ymm4, %ymm2
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; AVX2-NEXT: vpcmpgtq %ymm3, %ymm4, %ymm5
; AVX2-NEXT: vblendvpd %ymm5, %ymm3, %ymm4, %ymm3
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-NEXT: vpcmpgtq %ymm0, %ymm4, %ymm5
; AVX2-NEXT: vblendvpd %ymm5, %ymm0, %ymm4, %ymm0
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; AVX2-NEXT: vpcmpgtq %ymm1, %ymm4, %ymm5
; AVX2-NEXT: vblendvpd %ymm5, %ymm1, %ymm4, %ymm1
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX2-NEXT: vpcmpgtq %ymm4, %ymm1, %ymm5
; AVX2-NEXT: vpand %ymm1, %ymm5, %ymm1
[SelectionDAG] ComputeKnownBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits (zeros/ones) will be at least the minimum of the LO and HI constants. ComputeKnownBits equivalent of D43338. Differential Revision: https://reviews.llvm.org/D43463 llvm-svn: 325521
2018-02-20 02:08:16 +08:00
; AVX2-NEXT: vpcmpgtq %ymm4, %ymm0, %ymm5
; AVX2-NEXT: vpand %ymm0, %ymm5, %ymm0
; AVX2-NEXT: vpackusdw %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpcmpgtq %ymm4, %ymm3, %ymm1
; AVX2-NEXT: vpand %ymm3, %ymm1, %ymm1
; AVX2-NEXT: vpcmpgtq %ymm4, %ymm2, %ymm3
; AVX2-NEXT: vpand %ymm2, %ymm3, %ymm2
; AVX2-NEXT: vpackusdw %ymm1, %ymm2, %ymm1
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,2,1,3]
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-NEXT: vpackusdw %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
[X86] Add custom type legalization for v16i64->v16i8 truncate and v8i64->v8i8 truncate when v8i64 isn't legal Summary: The default legalization for v16i64->v16i8 tries to create a multiple stage truncate concatenating after each stage and truncating again. But avx512 implements truncates with multiple uops. So it should be better to truncate all the way to the desired element size and then concatenate the pieces using unpckl instructions. This minimizes the number of 2 uop truncates. The unpcks are all single uop instructions. I tried to handle this by just custom splitting the v16i64->v16i8 shuffle. And hoped that the DAG combiner would leave the two halves in the state needed to make D68374 do the job for each half. This worked for the first half, but the second half got messed up. So I've implemented custom handling for v8i64->v8i8 when v8i64 needs to be split to produce the VTRUNCs directly. Reviewers: RKSimon, spatel Reviewed By: RKSimon Subscribers: hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68428 llvm-svn: 373864
2019-10-07 02:43:08 +08:00
; AVX512F-LABEL: trunc_packus_v16i64_v16i8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastq {{.*#+}} zmm2 = [255,255,255,255,255,255,255,255]
; AVX512F-NEXT: vpminsq %zmm2, %zmm0, %zmm0
; AVX512F-NEXT: vpminsq %zmm2, %zmm1, %zmm1
; AVX512F-NEXT: vpxor %xmm2, %xmm2, %xmm2
; AVX512F-NEXT: vpmaxsq %zmm2, %zmm1, %zmm1
; AVX512F-NEXT: vpmaxsq %zmm2, %zmm0, %zmm0
; AVX512F-NEXT: vpmovqd %zmm0, %ymm0
; AVX512F-NEXT: vpmovqd %zmm1, %ymm1
; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT: vpmovdb %zmm0, %xmm0
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v16i64_v16i8:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm2, %xmm2, %xmm2
; AVX512VL-NEXT: vpmaxsq %zmm2, %zmm1, %zmm1
; AVX512VL-NEXT: vpmovusqb %zmm1, %xmm1
; AVX512VL-NEXT: vpmaxsq %zmm2, %zmm0, %zmm0
; AVX512VL-NEXT: vpmovusqb %zmm0, %xmm0
; AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v16i64_v16i8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} zmm2 = [255,255,255,255,255,255,255,255]
; AVX512BW-NEXT: vpminsq %zmm2, %zmm0, %zmm0
; AVX512BW-NEXT: vpminsq %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT: vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT: vpmaxsq %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT: vpmaxsq %zmm2, %zmm0, %zmm0
; AVX512BW-NEXT: vpmovqd %zmm0, %ymm0
; AVX512BW-NEXT: vpmovqd %zmm1, %ymm1
; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512BW-NEXT: vpmovdb %zmm0, %xmm0
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v16i64_v16i8:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm2, %xmm2, %xmm2
; AVX512BWVL-NEXT: vpmaxsq %zmm2, %zmm1, %zmm1
; AVX512BWVL-NEXT: vpmovusqb %zmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsq %zmm2, %zmm0, %zmm0
; AVX512BWVL-NEXT: vpmovusqb %zmm0, %xmm0
; AVX512BWVL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
%1 = icmp slt <16 x i64> %a0, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255>
%2 = select <16 x i1> %1, <16 x i64> %a0, <16 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255>
%3 = icmp sgt <16 x i64> %2, zeroinitializer
%4 = select <16 x i1> %3, <16 x i64> %2, <16 x i64> zeroinitializer
%5 = trunc <16 x i64> %4 to <16 x i8>
ret <16 x i8> %5
}
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
define <4 x i8> @trunc_packus_v4i32_v4i8(<4 x i32> %a0) {
; SSE2-LABEL: trunc_packus_v4i32_v4i8:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [255,255,255,255]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm0, %xmm2
; SSE2-NEXT: pand %xmm2, %xmm0
; SSE2-NEXT: pandn %xmm1, %xmm2
; SSE2-NEXT: por %xmm0, %xmm2
; SSE2-NEXT: pxor %xmm1, %xmm1
; SSE2-NEXT: movdqa %xmm2, %xmm0
; SSE2-NEXT: pcmpgtd %xmm1, %xmm0
; SSE2-NEXT: pand %xmm2, %xmm0
; SSE2-NEXT: pand {{.*}}(%rip), %xmm0
; SSE2-NEXT: packuswb %xmm0, %xmm0
; SSE2-NEXT: packuswb %xmm0, %xmm0
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v4i32_v4i8:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [255,255,255,255]
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pandn %xmm1, %xmm2
; SSSE3-NEXT: por %xmm0, %xmm2
; SSSE3-NEXT: pxor %xmm1, %xmm1
; SSSE3-NEXT: movdqa %xmm2, %xmm0
; SSSE3-NEXT: pcmpgtd %xmm1, %xmm0
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v4i32_v4i8:
; SSE41: # %bb.0:
; SSE41-NEXT: pminsd {{.*}}(%rip), %xmm0
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: pmaxsd %xmm1, %xmm0
; SSE41-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v4i32_v4i8:
; AVX1: # %bb.0:
; AVX1-NEXT: vpminsd {{.*}}(%rip), %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v4i32_v4i8:
; AVX2: # %bb.0:
; AVX2-NEXT: vpbroadcastd {{.*#+}} xmm1 = [255,255,255,255]
; AVX2-NEXT: vpminsd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v4i32_v4i8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastd {{.*#+}} xmm1 = [255,255,255,255]
; AVX512F-NEXT: vpminsd %xmm1, %xmm0, %xmm0
; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512F-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512F-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v4i32_v4i8:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512VL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v4i32_v4i8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastd {{.*#+}} xmm1 = [255,255,255,255]
; AVX512BW-NEXT: vpminsd %xmm1, %xmm0, %xmm0
; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BW-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v4i32_v4i8:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512BWVL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BWVL-NEXT: retq
%1 = icmp slt <4 x i32> %a0, <i32 255, i32 255, i32 255, i32 255>
%2 = select <4 x i1> %1, <4 x i32> %a0, <4 x i32> <i32 255, i32 255, i32 255, i32 255>
%3 = icmp sgt <4 x i32> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i32> %2, <4 x i32> zeroinitializer
%5 = trunc <4 x i32> %4 to <4 x i8>
ret <4 x i8> %5
}
define void @trunc_packus_v4i32_v4i8_store(<4 x i32> %a0, <4 x i8> *%p1) {
; SSE2-LABEL: trunc_packus_v4i32_v4i8_store:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [255,255,255,255]
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: pcmpgtd %xmm0, %xmm2
; SSE2-NEXT: pand %xmm2, %xmm0
; SSE2-NEXT: pandn %xmm1, %xmm2
; SSE2-NEXT: por %xmm0, %xmm2
; SSE2-NEXT: pxor %xmm0, %xmm0
; SSE2-NEXT: movdqa %xmm2, %xmm1
; SSE2-NEXT: pcmpgtd %xmm0, %xmm1
; SSE2-NEXT: pand %xmm2, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: packuswb %xmm0, %xmm1
; SSE2-NEXT: packuswb %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, (%rdi)
; SSE2-NEXT: retq
;
; SSSE3-LABEL: trunc_packus_v4i32_v4i8_store:
; SSSE3: # %bb.0:
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [255,255,255,255]
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pandn %xmm1, %xmm2
; SSSE3-NEXT: por %xmm0, %xmm2
; SSSE3-NEXT: pxor %xmm0, %xmm0
; SSSE3-NEXT: movdqa %xmm2, %xmm1
; SSSE3-NEXT: pcmpgtd %xmm0, %xmm1
; SSSE3-NEXT: pand %xmm2, %xmm1
; SSSE3-NEXT: pshufb {{.*#+}} xmm1 = xmm1[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; SSSE3-NEXT: movd %xmm1, (%rdi)
; SSSE3-NEXT: retq
;
; SSE41-LABEL: trunc_packus_v4i32_v4i8_store:
; SSE41: # %bb.0:
; SSE41-NEXT: pminsd {{.*}}(%rip), %xmm0
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: pmaxsd %xmm0, %xmm1
; SSE41-NEXT: pshufb {{.*#+}} xmm1 = xmm1[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; SSE41-NEXT: movd %xmm1, (%rdi)
; SSE41-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v4i32_v4i8_store:
; AVX1: # %bb.0:
; AVX1-NEXT: vpminsd {{.*}}(%rip), %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX1-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX1-NEXT: vmovd %xmm0, (%rdi)
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v4i32_v4i8_store:
; AVX2: # %bb.0:
; AVX2-NEXT: vpbroadcastd {{.*#+}} xmm1 = [255,255,255,255]
; AVX2-NEXT: vpminsd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT: vmovd %xmm0, (%rdi)
; AVX2-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v4i32_v4i8_store:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastd {{.*#+}} xmm1 = [255,255,255,255]
; AVX512F-NEXT: vpminsd %xmm1, %xmm0, %xmm0
; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512F-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512F-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512F-NEXT: vmovd %xmm0, (%rdi)
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v4i32_v4i8_store:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512VL-NEXT: vpmovusdb %xmm0, (%rdi)
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v4i32_v4i8_store:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastd {{.*#+}} xmm1 = [255,255,255,255]
; AVX512BW-NEXT: vpminsd %xmm1, %xmm0, %xmm0
; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BW-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT: vmovd %xmm0, (%rdi)
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v4i32_v4i8_store:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsd %xmm1, %xmm0, %xmm0
; AVX512BWVL-NEXT: vpmovusdb %xmm0, (%rdi)
; AVX512BWVL-NEXT: retq
%1 = icmp slt <4 x i32> %a0, <i32 255, i32 255, i32 255, i32 255>
%2 = select <4 x i1> %1, <4 x i32> %a0, <4 x i32> <i32 255, i32 255, i32 255, i32 255>
%3 = icmp sgt <4 x i32> %2, zeroinitializer
%4 = select <4 x i1> %3, <4 x i32> %2, <4 x i32> zeroinitializer
%5 = trunc <4 x i32> %4 to <4 x i8>
store <4 x i8> %5, <4 x i8> *%p1
ret void
}
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
define <8 x i8> @trunc_packus_v8i32_v8i8(<8 x i32> %a0) {
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; SSE-LABEL: trunc_packus_v8i32_v8i8:
; SSE: # %bb.0:
; SSE-NEXT: packssdw %xmm1, %xmm0
; SSE-NEXT: packuswb %xmm0, %xmm0
; SSE-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
; AVX1-LABEL: trunc_packus_v8i32_v8i8:
; AVX1: # %bb.0:
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vpackssdw %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v8i32_v8i8:
; AVX2: # %bb.0:
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
Recommit r367901 "[X86] Enable -x86-experimental-vector-widening-legalization by default." The assert that caused this to be reverted should be fixed now. Original commit message: This patch changes our defualt legalization behavior for 16, 32, and 64 bit vectors with i8/i16/i32/i64 scalar types from promotion to widening. For example, v8i8 will now be widened to v16i8 instead of promoted to v8i16. This keeps the elements widths the same and pads with undef elements. We believe this is a better legalization strategy. But it carries some issues due to the fragmented vector ISA. For example, i8 shifts and multiplies get widened and then later have to be promoted/split into vXi16 vectors. This has the potential to cause regressions so we wanted to get it in early in the 10.0 cycle so we have plenty of time to address them. Next steps will be to merge tests that explicitly test the command line option. And then we can remove the option and its associated code. llvm-svn: 368183
2019-08-08 00:24:26 +08:00
; AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v8i32_v8i8:
; AVX512F: # %bb.0:
[X86] Improve the AVX512 bailout in combineTruncateWithSat to allow pack instructions in more situations. If we don't have VLX we won't end up selecting a saturating truncate for 256-bit or smaller vectors so we should just use the pack lowering. llvm-svn: 374487
2019-10-11 08:38:51 +08:00
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: vpackssdw %xmm1, %xmm0, %xmm0
; AVX512F-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v8i32_v8i8:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0
[X86] Add DAG combine to form saturating VTRUNCUS/VTRUNCS from VTRUNC We already do this for ISD::TRUNCATE, but we can do the same for X86ISD::VTRUNC Differential Revision: https://reviews.llvm.org/D68432 llvm-svn: 373765
2019-10-05 01:53:18 +08:00
; AVX512VL-NEXT: vpmovusdb %ymm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v8i32_v8i8:
; AVX512BW: # %bb.0:
[X86] Improve the AVX512 bailout in combineTruncateWithSat to allow pack instructions in more situations. If we don't have VLX we won't end up selecting a saturating truncate for 256-bit or smaller vectors so we should just use the pack lowering. llvm-svn: 374487
2019-10-11 08:38:51 +08:00
; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512BW-NEXT: vpackssdw %xmm1, %xmm0, %xmm0
; AVX512BW-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v8i32_v8i8:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0
[X86] Add DAG combine to form saturating VTRUNCUS/VTRUNCS from VTRUNC We already do this for ISD::TRUNCATE, but we can do the same for X86ISD::VTRUNC Differential Revision: https://reviews.llvm.org/D68432 llvm-svn: 373765
2019-10-05 01:53:18 +08:00
; AVX512BWVL-NEXT: vpmovusdb %ymm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
%1 = icmp slt <8 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
%2 = select <8 x i1> %1, <8 x i32> %a0, <8 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
%3 = icmp sgt <8 x i32> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i32> %2, <8 x i32> zeroinitializer
%5 = trunc <8 x i32> %4 to <8 x i8>
ret <8 x i8> %5
}
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
define void @trunc_packus_v8i32_v8i8_store(<8 x i32> %a0, <8 x i8> *%p1) {
[X86][SSE] combineTruncateWithSat - use truncateVectorWithPACK down to 64-bit subvectors Add support for chaining PACKSS/PACKUS down to 64-bit vectors by using only a single 128-bit input. llvm-svn: 325494
2018-02-19 21:29:20 +08:00
; SSE-LABEL: trunc_packus_v8i32_v8i8_store:
; SSE: # %bb.0:
; SSE-NEXT: packssdw %xmm1, %xmm0
; SSE-NEXT: packuswb %xmm0, %xmm0
; SSE-NEXT: movq %xmm0, (%rdi)
; SSE-NEXT: retq
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
;
; AVX1-LABEL: trunc_packus_v8i32_v8i8_store:
; AVX1: # %bb.0:
[X86][SSE] combineTruncateWithSat - use truncateVectorWithPACK down to 64-bit subvectors Add support for chaining PACKSS/PACKUS down to 64-bit vectors by using only a single 128-bit input. llvm-svn: 325494
2018-02-19 21:29:20 +08:00
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vpackssdw %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; AVX1-NEXT: vmovq %xmm0, (%rdi)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v8i32_v8i8_store:
; AVX2: # %bb.0:
[SelectionDAG] ComputeNumSignBits - add support for SMIN+SMAX clamp patterns If we have a clamp pattern, SMIN(SMAX(X, LO),HI) or SMAX(SMIN(X, HI),LO) then we can deduce that the number of signbits will be at least the minimum of the LO and HI constants. I haven't bothered with the UMIN/UMAX equivalent as (1) we don't have any current use cases and (2) I wonder if we'd be better off immediately falling back for ComputeKnownBits for UMIN/UMAX which already has optimization patterns useful for unsigned cases. Differential Revision: https://reviews.llvm.org/D43338 llvm-svn: 325450
2018-02-18 06:19:50 +08:00
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0
[X86][SSE] combineTruncateWithSat - use truncateVectorWithPACK down to 64-bit subvectors Add support for chaining PACKSS/PACKUS down to 64-bit vectors by using only a single 128-bit input. llvm-svn: 325494
2018-02-19 21:29:20 +08:00
; AVX2-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; AVX2-NEXT: vmovq %xmm0, (%rdi)
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v8i32_v8i8_store:
; AVX512F: # %bb.0:
[X86] Improve the AVX512 bailout in combineTruncateWithSat to allow pack instructions in more situations. If we don't have VLX we won't end up selecting a saturating truncate for 256-bit or smaller vectors so we should just use the pack lowering. llvm-svn: 374487
2019-10-11 08:38:51 +08:00
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: vpackssdw %xmm1, %xmm0, %xmm0
; AVX512F-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; AVX512F-NEXT: vmovq %xmm0, (%rdi)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v8i32_v8i8_store:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512VL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512VL-NEXT: vpmovusdb %ymm0, (%rdi)
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v8i32_v8i8_store:
; AVX512BW: # %bb.0:
[X86] Improve the AVX512 bailout in combineTruncateWithSat to allow pack instructions in more situations. If we don't have VLX we won't end up selecting a saturating truncate for 256-bit or smaller vectors so we should just use the pack lowering. llvm-svn: 374487
2019-10-11 08:38:51 +08:00
; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512BW-NEXT: vpackssdw %xmm1, %xmm0, %xmm0
; AVX512BW-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; AVX512BW-NEXT: vmovq %xmm0, (%rdi)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v8i32_v8i8_store:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsd %ymm1, %ymm0, %ymm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512BWVL-NEXT: vpmovusdb %ymm0, (%rdi)
[X86][SSE] Add saturated truncation tests for storing illegal v8i8 types Tests showing missing opportunities to use PACK instructions in cases where we need to truncate to illegal types for stores llvm-svn: 325270
2018-02-16 01:48:34 +08:00
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
%1 = icmp slt <8 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
%2 = select <8 x i1> %1, <8 x i32> %a0, <8 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
%3 = icmp sgt <8 x i32> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i32> %2, <8 x i32> zeroinitializer
%5 = trunc <8 x i32> %4 to <8 x i8>
store <8 x i8> %5, <8 x i8> *%p1
ret void
}
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
define <16 x i8> @trunc_packus_v16i32_v16i8(<16 x i32> %a0) {
[X86][SSE] combineTruncateWithSat - use truncateVectorWithPACK to chain PACKUS vXi32-vXi8 saturated truncation We can use PACKSS/PACKUS to saturate each stage of the chain: PACKSSDW down to [-32768,32767] and then PACKUSWB to [0,255]. llvm-svn: 325243
2018-02-15 22:37:59 +08:00
; SSE-LABEL: trunc_packus_v16i32_v16i8:
; SSE: # %bb.0:
; SSE-NEXT: packssdw %xmm3, %xmm2
; SSE-NEXT: packssdw %xmm1, %xmm0
; SSE-NEXT: packuswb %xmm2, %xmm0
; SSE-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
; AVX1-LABEL: trunc_packus_v16i32_v16i8:
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
[X86][SSE] combineTruncateWithSat - use truncateVectorWithPACK to chain PACKUS vXi32-vXi8 saturated truncation We can use PACKSS/PACKUS to saturate each stage of the chain: PACKSSDW down to [-32768,32767] and then PACKUSWB to [0,255]. llvm-svn: 325243
2018-02-15 22:37:59 +08:00
; AVX1-NEXT: vpackssdw %xmm2, %xmm1, %xmm1
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v16i32_v16i8:
; AVX2: # %bb.0:
[X86][SSE] combineTruncateWithSat - use truncateVectorWithPACK to chain PACKUS vXi32-vXi8 saturated truncation We can use PACKSS/PACKUS to saturate each stage of the chain: PACKSSDW down to [-32768,32767] and then PACKUSWB to [0,255]. llvm-svn: 325243
2018-02-15 22:37:59 +08:00
; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
; AVX512-LABEL: trunc_packus_v16i32_v16i8:
; AVX512: # %bb.0:
; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512-NEXT: vpmaxsd %zmm1, %zmm0, %zmm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512-NEXT: vpmovusdb %zmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
%1 = icmp slt <16 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
%2 = select <16 x i1> %1, <16 x i32> %a0, <16 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
%3 = icmp sgt <16 x i32> %2, zeroinitializer
%4 = select <16 x i1> %3, <16 x i32> %2, <16 x i32> zeroinitializer
%5 = trunc <16 x i32> %4 to <16 x i8>
ret <16 x i8> %5
}
[X86] Add more packus/ssat/usat truncate tests from legal vectors to less than 128-bit vectors. Some of these have sub-optimal codegen for avx512 relative to avx2. llvm-svn: 374505
2019-10-11 11:46:39 +08:00
define <8 x i8> @trunc_packus_v8i16_v8i8(<8 x i16> %a0) {
; SSE-LABEL: trunc_packus_v8i16_v8i8:
; SSE: # %bb.0:
; SSE-NEXT: packuswb %xmm0, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: trunc_packus_v8i16_v8i8:
; AVX: # %bb.0:
; AVX-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v8i16_v8i8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v8i16_v8i8:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v8i16_v8i8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v8i16_v8i8:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpminsw {{.*}}(%rip), %xmm0, %xmm0
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0
; AVX512BWVL-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX512BWVL-NEXT: retq
%1 = icmp slt <8 x i16> %a0, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
%2 = select <8 x i1> %1, <8 x i16> %a0, <8 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
%3 = icmp sgt <8 x i16> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i16> %2, <8 x i16> zeroinitializer
%5 = trunc <8 x i16> %4 to <8 x i8>
ret <8 x i8> %5
}
define void @trunc_packus_v8i16_v8i8_store(<8 x i16> %a0, <8 x i8> *%p1) {
; SSE-LABEL: trunc_packus_v8i16_v8i8_store:
; SSE: # %bb.0:
; SSE-NEXT: packuswb %xmm0, %xmm0
; SSE-NEXT: movq %xmm0, (%rdi)
; SSE-NEXT: retq
;
; AVX-LABEL: trunc_packus_v8i16_v8i8_store:
; AVX: # %bb.0:
; AVX-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX-NEXT: vmovq %xmm0, (%rdi)
; AVX-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v8i16_v8i8_store:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX512F-NEXT: vmovq %xmm0, (%rdi)
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v8i16_v8i8_store:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX512VL-NEXT: vmovq %xmm0, (%rdi)
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v8i16_v8i8_store:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
; AVX512BW-NEXT: vmovq %xmm0, (%rdi)
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v8i16_v8i8_store:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0
; AVX512BWVL-NEXT: vpmovuswb %xmm0, (%rdi)
; AVX512BWVL-NEXT: retq
%1 = icmp slt <8 x i16> %a0, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
%2 = select <8 x i1> %1, <8 x i16> %a0, <8 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
%3 = icmp sgt <8 x i16> %2, zeroinitializer
%4 = select <8 x i1> %3, <8 x i16> %2, <8 x i16> zeroinitializer
%5 = trunc <8 x i16> %4 to <8 x i8>
store <8 x i8> %5, <8 x i8> *%p1
ret void
}
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
define <16 x i8> @trunc_packus_v16i16_v16i8(<16 x i16> %a0) {
[X86][SSE] Relax type legality for combineTruncateWithSat PACKSS/PACKUS truncation While the AVX512 VTRUNCS/VTRUNCUS instructions require legal types, truncateVectorWithPACK handles cases with multiples of legal types through splitting/concatenation. So we just need to ensure that the src/dst scalar types are correct and leave truncateVectorWithPACK to handle the rest of it. llvm-svn: 325127
2018-02-14 22:14:29 +08:00
; SSE-LABEL: trunc_packus_v16i16_v16i8:
; SSE: # %bb.0:
; SSE-NEXT: packuswb %xmm1, %xmm0
; SSE-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
; AVX1-LABEL: trunc_packus_v16i16_v16i8:
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
[X86][SSE] Add PACKUS support for truncation of clamped values Followup to D42544 that matches PACKUSWB cases for non-AVX512, SSE and PACKUSDW cases will have to wait until we can add support for general SMIN/SMAX matching. llvm-svn: 324347
2018-02-06 22:07:46 +08:00
; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v16i16_v16i8:
; AVX2: # %bb.0:
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
[X86][SSE] Add PACKUS support for truncation of clamped values Followup to D42544 that matches PACKUSWB cases for non-AVX512, SSE and PACKUSDW cases will have to wait until we can add support for general SMIN/SMAX matching. llvm-svn: 324347
2018-02-06 22:07:46 +08:00
; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v16i16_v16i8:
; AVX512F: # %bb.0:
[X86] Allow combineTruncateWithSat to use pack instructions for i16->i8 without AVX512BW. We need AVX512BW to be able to truncate an i16 vector. If we don't have that we have to extend i16->i32, then trunc, i32->i8. But we won't be able to remove the min/max if we do that. At least not without more special handling. llvm-svn: 368623
2019-08-13 06:18:23 +08:00
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v16i16_v16i8:
; AVX512VL: # %bb.0:
[X86] Allow combineTruncateWithSat to use pack instructions for i16->i8 without AVX512BW. We need AVX512BW to be able to truncate an i16 vector. If we don't have that we have to extend i16->i32, then trunc, i32->i8. But we won't be able to remove the min/max if we do that. At least not without more special handling. llvm-svn: 368623
2019-08-13 06:18:23 +08:00
; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512VL-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v16i16_v16i8:
; AVX512BW: # %bb.0:
[X86] Improve the AVX512 bailout in combineTruncateWithSat to allow pack instructions in more situations. If we don't have VLX we won't end up selecting a saturating truncate for 256-bit or smaller vectors so we should just use the pack lowering. llvm-svn: 374487
2019-10-11 08:38:51 +08:00
; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512BW-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v16i16_v16i8:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512BWVL-NEXT: vpmovuswb %ymm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
%1 = icmp slt <16 x i16> %a0, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
%2 = select <16 x i1> %1, <16 x i16> %a0, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
%3 = icmp sgt <16 x i16> %2, zeroinitializer
%4 = select <16 x i1> %3, <16 x i16> %2, <16 x i16> zeroinitializer
%5 = trunc <16 x i16> %4 to <16 x i8>
ret <16 x i8> %5
}
define <32 x i8> @trunc_packus_v32i16_v32i8(<32 x i16> %a0) {
[X86][SSE] Relax type legality for combineTruncateWithSat PACKSS/PACKUS truncation While the AVX512 VTRUNCS/VTRUNCUS instructions require legal types, truncateVectorWithPACK handles cases with multiples of legal types through splitting/concatenation. So we just need to ensure that the src/dst scalar types are correct and leave truncateVectorWithPACK to handle the rest of it. llvm-svn: 325127
2018-02-14 22:14:29 +08:00
; SSE-LABEL: trunc_packus_v32i16_v32i8:
; SSE: # %bb.0:
; SSE-NEXT: packuswb %xmm1, %xmm0
; SSE-NEXT: packuswb %xmm3, %xmm2
; SSE-NEXT: movdqa %xmm2, %xmm1
; SSE-NEXT: retq
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
;
; AVX1-LABEL: trunc_packus_v32i16_v32i8:
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
[X86][SSE] Add PACKUS support for truncation of clamped values Followup to D42544 that matches PACKUSWB cases for non-AVX512, SSE and PACKUSDW cases will have to wait until we can add support for general SMIN/SMAX matching. llvm-svn: 324347
2018-02-06 22:07:46 +08:00
; AVX1-NEXT: vpackuswb %xmm2, %xmm1, %xmm1
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
; AVX1-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v32i16_v32i8:
; AVX2: # %bb.0:
[X86][SSE] Relax type legality for combineTruncateWithSat PACKSS/PACKUS truncation While the AVX512 VTRUNCS/VTRUNCUS instructions require legal types, truncateVectorWithPACK handles cases with multiples of legal types through splitting/concatenation. So we just need to ensure that the src/dst scalar types are correct and leave truncateVectorWithPACK to handle the rest of it. llvm-svn: 325127
2018-02-14 22:14:29 +08:00
; AVX2-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX2-NEXT: retq
;
; AVX512F-LABEL: trunc_packus_v32i16_v32i8:
; AVX512F: # %bb.0:
[X86] Pass v32i16/v64i8 in zmm registers on KNL target. gcc and icc pass these types in zmm registers in zmm registers. This patch implements a quick hack to override the register type before calling convention handling to one that is legal. Longer term we might want to do something similar to 256-bit integer registers on AVX1 where we just split all the operations. Fixes PR42957 Differential Revision: https://reviews.llvm.org/D66708 llvm-svn: 370495
2019-08-31 01:35:08 +08:00
; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
[X86] Allow combineTruncateWithSat to use pack instructions for i16->i8 without AVX512BW. We need AVX512BW to be able to truncate an i16 vector. If we don't have that we have to extend i16->i32, then trunc, i32->i8. But we won't be able to remove the min/max if we do that. At least not without more special handling. llvm-svn: 368623
2019-08-13 06:18:23 +08:00
; AVX512F-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
; AVX512F-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: trunc_packus_v32i16_v32i8:
; AVX512VL: # %bb.0:
[X86] Pass v32i16/v64i8 in zmm registers on KNL target. gcc and icc pass these types in zmm registers in zmm registers. This patch implements a quick hack to override the register type before calling convention handling to one that is legal. Longer term we might want to do something similar to 256-bit integer registers on AVX1 where we just split all the operations. Fixes PR42957 Differential Revision: https://reviews.llvm.org/D66708 llvm-svn: 370495
2019-08-31 01:35:08 +08:00
; AVX512VL-NEXT: vextracti64x4 $1, %zmm0, %ymm1
[X86] Allow combineTruncateWithSat to use pack instructions for i16->i8 without AVX512BW. We need AVX512BW to be able to truncate an i16 vector. If we don't have that we have to extend i16->i32, then trunc, i32->i8. But we won't be able to remove the min/max if we do that. At least not without more special handling. llvm-svn: 368623
2019-08-13 06:18:23 +08:00
; AVX512VL-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
; AVX512VL-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: trunc_packus_v32i16_v32i8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BW-NEXT: vpmaxsw %zmm1, %zmm0, %zmm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512BW-NEXT: vpmovuswb %zmm0, %ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_packus_v32i16_v32i8:
; AVX512BWVL: # %bb.0:
; AVX512BWVL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX512BWVL-NEXT: vpmaxsw %zmm1, %zmm0, %zmm0
[X86] Improve unsigned saturation downconvert detection. Summary: New unsigned saturation downconvert patterns detection was implemented in X86 Codegen: (truncate (smin (smax (x, C1), C2)) to dest_type), where C1 >= 0 and C2 is unsigned max of destination type. (truncate (smax (smin (x, C2), C1)) to dest_type) where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2. These two patterns are equivalent to: (truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type) Reviewers: RKSimon Subscribers: llvm-commits, a.elovikov Differential Revision: https://reviews.llvm.org/D45315 llvm-svn: 332336
2018-05-15 18:24:12 +08:00
; AVX512BWVL-NEXT: vpmovuswb %zmm0, %ymm0
[X86][SSE] Add tests for vector truncation with PACKUS style signed saturation PACKUS - truncates signed value, saturating to [0,unsigned_max_trunc] llvm-svn: 323531
2018-01-26 22:58:50 +08:00
; AVX512BWVL-NEXT: retq
%1 = icmp slt <32 x i16> %a0, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
%2 = select <32 x i1> %1, <32 x i16> %a0, <32 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
%3 = icmp sgt <32 x i16> %2, zeroinitializer
%4 = select <32 x i1> %3, <32 x i16> %2, <32 x i16> zeroinitializer
%5 = trunc <32 x i16> %4 to <32 x i8>
ret <32 x i8> %5
}
[X86] Add test cases for packus/ssat/usat 32i32->v32i8 test cases. NFC llvm-svn: 374459
2019-10-11 05:46:44 +08:00
define <32 x i8> @trunc_packus_v32i32_v32i8(<32 x i32> %a0) {
; SSE-LABEL: trunc_packus_v32i32_v32i8:
; SSE: # %bb.0:
; SSE-NEXT: packssdw %xmm3, %xmm2
; SSE-NEXT: packssdw %xmm1, %xmm0
; SSE-NEXT: packuswb %xmm2, %xmm0
; SSE-NEXT: packssdw %xmm7, %xmm6
; SSE-NEXT: packssdw %xmm5, %xmm4
; SSE-NEXT: packuswb %xmm6, %xmm4
; SSE-NEXT: movdqa %xmm4, %xmm1
; SSE-NEXT: retq
;
; AVX1-LABEL: trunc_packus_v32i32_v32i8:
; AVX1: # %bb.0:
; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm4
; AVX1-NEXT: vpackssdw %xmm4, %xmm3, %xmm3
; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm4
; AVX1-NEXT: vpackssdw %xmm4, %xmm2, %xmm2
; AVX1-NEXT: vpackuswb %xmm3, %xmm2, %xmm2
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3
; AVX1-NEXT: vpackssdw %xmm3, %xmm1, %xmm1
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
; AVX1-NEXT: vpackssdw %xmm3, %xmm0, %xmm0
; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: trunc_packus_v32i32_v32i8:
; AVX2: # %bb.0:
; AVX2-NEXT: vpackssdw %ymm3, %ymm2, %ymm2
; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]
; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-NEXT: vpackuswb %ymm2, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
; AVX2-NEXT: retq
;
; AVX512-LABEL: trunc_packus_v32i32_v32i8:
; AVX512: # %bb.0:
; AVX512-NEXT: vpxor %xmm2, %xmm2, %xmm2
; AVX512-NEXT: vpmaxsd %zmm2, %zmm0, %zmm0
; AVX512-NEXT: vpmovusdb %zmm0, %xmm0
; AVX512-NEXT: vpmaxsd %zmm2, %zmm1, %zmm1
; AVX512-NEXT: vpmovusdb %zmm1, %xmm1
; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512-NEXT: retq
%1 = icmp slt <32 x i32> %a0, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
%2 = select <32 x i1> %1, <32 x i32> %a0, <32 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
%3 = icmp sgt <32 x i32> %2, zeroinitializer
%4 = select <32 x i1> %3, <32 x i32> %2, <32 x i32> zeroinitializer
%5 = trunc <32 x i32> %4 to <32 x i8>
ret <32 x i8> %5
}