llvm-6502/test/CodeGen/X86/vector-idiv.ll

1274 lines
47 KiB
LLVM
Raw Normal View History

; RUN: llc -march=x86-64 -mcpu=core2 -mattr=+sse4.1 < %s | FileCheck %s --check-prefix=SSE41
; RUN: llc -march=x86-64 -mcpu=core2 < %s | FileCheck %s --check-prefix=SSE
; RUN: llc -march=x86-64 -mcpu=core-avx2 < %s | FileCheck %s --check-prefix=AVX
target triple = "x86_64-unknown-unknown"
define <4 x i32> @test1(<4 x i32> %a) #0 {
; SSE41-LABEL: test1:
; SSE41: # BB#0:
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [613566757,613566757,613566757,613566757]
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pmuludq %xmm2, %xmm3
; SSE41-NEXT: pmuludq %xmm0, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2,3],xmm1[4,5],xmm3[6,7]
; SSE41-NEXT: psubd %xmm1, %xmm0
; SSE41-NEXT: psrld $1, %xmm0
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: paddd %xmm1, %xmm0
; SSE41-NEXT: psrld $2, %xmm0
; SSE41-NEXT: retq
;
; SSE-LABEL: test1:
; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [613566757,613566757,613566757,613566757]
; SSE-NEXT: movdqa %xmm0, %xmm2
; SSE-NEXT: pmuludq %xmm1, %xmm2
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
; SSE-NEXT: pmuludq %xmm1, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: psubd %xmm2, %xmm0
; SSE-NEXT: psrld $1, %xmm0
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: paddd %xmm2, %xmm0
; SSE-NEXT: psrld $2, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: test1:
; AVX: # BB#0:
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %xmm1
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; AVX-NEXT: vpmuludq %xmm2, %xmm3, %xmm2
; AVX-NEXT: vpmuludq %xmm1, %xmm0, %xmm1
; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; AVX-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2],xmm2[3]
; AVX-NEXT: vpsubd %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpsrld $1, %xmm0, %xmm0
; AVX-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpsrld $2, %xmm0, %xmm0
; AVX-NEXT: retq
%div = udiv <4 x i32> %a, <i32 7, i32 7, i32 7, i32 7>
ret <4 x i32> %div
}
define <8 x i32> @test2(<8 x i32> %a) #0 {
; SSE41-LABEL: test2:
; SSE41: # BB#0:
; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [613566757,613566757,613566757,613566757]
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
; SSE41-NEXT: pmuludq %xmm3, %xmm4
; SSE41-NEXT: movdqa %xmm0, %xmm5
; SSE41-NEXT: pmuludq %xmm2, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm5 = xmm5[0,1],xmm4[2,3],xmm5[4,5],xmm4[6,7]
; SSE41-NEXT: psubd %xmm5, %xmm0
; SSE41-NEXT: psrld $1, %xmm0
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: paddd %xmm5, %xmm0
; SSE41-NEXT: psrld $2, %xmm0
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
; SSE41-NEXT: pmuludq %xmm3, %xmm4
; SSE41-NEXT: pmuludq %xmm1, %xmm2
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3],xmm2[4,5],xmm4[6,7]
; SSE41-NEXT: psubd %xmm2, %xmm1
; SSE41-NEXT: psrld $1, %xmm1
; SSE41-NEXT: paddd %xmm2, %xmm1
; SSE41-NEXT: psrld $2, %xmm1
; SSE41-NEXT: retq
;
; SSE-LABEL: test2:
; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [613566757,613566757,613566757,613566757]
; SSE-NEXT: movdqa %xmm0, %xmm3
; SSE-NEXT: pmuludq %xmm2, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,3,2,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
; SSE-NEXT: pmuludq %xmm4, %xmm5
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1]
; SSE-NEXT: psubd %xmm3, %xmm0
; SSE-NEXT: psrld $1, %xmm0
; SSE-NEXT: paddd %xmm3, %xmm0
; SSE-NEXT: psrld $2, %xmm0
; SSE-NEXT: pmuludq %xmm1, %xmm2
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,1,3,3]
; SSE-NEXT: pmuludq %xmm4, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
; SSE-NEXT: psubd %xmm2, %xmm1
; SSE-NEXT: psrld $1, %xmm1
; SSE-NEXT: paddd %xmm2, %xmm1
; SSE-NEXT: psrld $2, %xmm1
; SSE-NEXT: retq
;
; AVX-LABEL: test2:
; AVX: # BB#0:
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %ymm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} ymm2 = ymm1[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpshufd {{.*#+}} ymm3 = ymm0[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpmuludq %ymm2, %ymm3, %ymm2
; AVX-NEXT: vpmuludq %ymm1, %ymm0, %ymm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7]
; AVX-NEXT: vpsubd %ymm1, %ymm0, %ymm0
; AVX-NEXT: vpsrld $1, %ymm0, %ymm0
; AVX-NEXT: vpaddd %ymm1, %ymm0, %ymm0
; AVX-NEXT: vpsrld $2, %ymm0, %ymm0
; AVX-NEXT: retq
%div = udiv <8 x i32> %a, <i32 7, i32 7, i32 7, i32 7,i32 7, i32 7, i32 7, i32 7>
ret <8 x i32> %div
}
define <8 x i16> @test3(<8 x i16> %a) #0 {
; SSE41-LABEL: test3:
; SSE41: # BB#0:
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [9363,9363,9363,9363,9363,9363,9363,9363]
; SSE41-NEXT: pmulhuw %xmm0, %xmm1
; SSE41-NEXT: psubw %xmm1, %xmm0
; SSE41-NEXT: psrlw $1, %xmm0
; SSE41-NEXT: paddw %xmm1, %xmm0
; SSE41-NEXT: psrlw $2, %xmm0
; SSE41-NEXT: retq
;
; SSE-LABEL: test3:
; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [9363,9363,9363,9363,9363,9363,9363,9363]
; SSE-NEXT: pmulhuw %xmm0, %xmm1
; SSE-NEXT: psubw %xmm1, %xmm0
; SSE-NEXT: psrlw $1, %xmm0
; SSE-NEXT: paddw %xmm1, %xmm0
; SSE-NEXT: psrlw $2, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: test3:
; AVX: # BB#0:
; AVX-NEXT: vpmulhuw {{.*}}(%rip), %xmm0, %xmm1
; AVX-NEXT: vpsubw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpsrlw $1, %xmm0, %xmm0
; AVX-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpsrlw $2, %xmm0, %xmm0
; AVX-NEXT: retq
%div = udiv <8 x i16> %a, <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>
ret <8 x i16> %div
}
define <16 x i16> @test4(<16 x i16> %a) #0 {
; SSE41-LABEL: test4:
; SSE41: # BB#0:
; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [9363,9363,9363,9363,9363,9363,9363,9363]
; SSE41-NEXT: movdqa %xmm0, %xmm3
; SSE41-NEXT: pmulhuw %xmm2, %xmm3
; SSE41-NEXT: psubw %xmm3, %xmm0
; SSE41-NEXT: psrlw $1, %xmm0
; SSE41-NEXT: paddw %xmm3, %xmm0
; SSE41-NEXT: psrlw $2, %xmm0
; SSE41-NEXT: pmulhuw %xmm1, %xmm2
; SSE41-NEXT: psubw %xmm2, %xmm1
; SSE41-NEXT: psrlw $1, %xmm1
; SSE41-NEXT: paddw %xmm2, %xmm1
; SSE41-NEXT: psrlw $2, %xmm1
; SSE41-NEXT: retq
;
; SSE-LABEL: test4:
; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [9363,9363,9363,9363,9363,9363,9363,9363]
; SSE-NEXT: movdqa %xmm0, %xmm3
; SSE-NEXT: pmulhuw %xmm2, %xmm3
; SSE-NEXT: psubw %xmm3, %xmm0
; SSE-NEXT: psrlw $1, %xmm0
; SSE-NEXT: paddw %xmm3, %xmm0
; SSE-NEXT: psrlw $2, %xmm0
; SSE-NEXT: pmulhuw %xmm1, %xmm2
; SSE-NEXT: psubw %xmm2, %xmm1
; SSE-NEXT: psrlw $1, %xmm1
; SSE-NEXT: paddw %xmm2, %xmm1
; SSE-NEXT: psrlw $2, %xmm1
; SSE-NEXT: retq
;
; AVX-LABEL: test4:
; AVX: # BB#0:
; AVX-NEXT: vpmulhuw {{.*}}(%rip), %ymm0, %ymm1
; AVX-NEXT: vpsubw %ymm1, %ymm0, %ymm0
; AVX-NEXT: vpsrlw $1, %ymm0, %ymm0
; AVX-NEXT: vpaddw %ymm1, %ymm0, %ymm0
; AVX-NEXT: vpsrlw $2, %ymm0, %ymm0
; AVX-NEXT: retq
%div = udiv <16 x i16> %a, <i16 7, i16 7, i16 7, i16 7,i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7,i16 7, i16 7, i16 7, i16 7>
ret <16 x i16> %div
}
define <8 x i16> @test5(<8 x i16> %a) #0 {
; SSE41-LABEL: test5:
; SSE41: # BB#0:
; SSE41-NEXT: pmulhw {{.*}}(%rip), %xmm0
; SSE41-NEXT: movdqa %xmm0, %xmm1
; SSE41-NEXT: psrlw $15, %xmm1
; SSE41-NEXT: psraw $1, %xmm0
; SSE41-NEXT: paddw %xmm1, %xmm0
; SSE41-NEXT: retq
;
; SSE-LABEL: test5:
; SSE: # BB#0:
; SSE-NEXT: pmulhw {{.*}}(%rip), %xmm0
; SSE-NEXT: movdqa %xmm0, %xmm1
; SSE-NEXT: psrlw $15, %xmm1
; SSE-NEXT: psraw $1, %xmm0
; SSE-NEXT: paddw %xmm1, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: test5:
; AVX: # BB#0:
; AVX-NEXT: vpmulhw {{.*}}(%rip), %xmm0, %xmm0
; AVX-NEXT: vpsrlw $15, %xmm0, %xmm1
; AVX-NEXT: vpsraw $1, %xmm0, %xmm0
; AVX-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX-NEXT: retq
%div = sdiv <8 x i16> %a, <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>
ret <8 x i16> %div
}
define <16 x i16> @test6(<16 x i16> %a) #0 {
; SSE41-LABEL: test6:
; SSE41: # BB#0:
; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [18725,18725,18725,18725,18725,18725,18725,18725]
; SSE41-NEXT: pmulhw %xmm2, %xmm0
; SSE41-NEXT: movdqa %xmm0, %xmm3
; SSE41-NEXT: psrlw $15, %xmm3
; SSE41-NEXT: psraw $1, %xmm0
; SSE41-NEXT: paddw %xmm3, %xmm0
; SSE41-NEXT: pmulhw %xmm2, %xmm1
; SSE41-NEXT: movdqa %xmm1, %xmm2
; SSE41-NEXT: psrlw $15, %xmm2
; SSE41-NEXT: psraw $1, %xmm1
; SSE41-NEXT: paddw %xmm2, %xmm1
; SSE41-NEXT: retq
;
; SSE-LABEL: test6:
; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [18725,18725,18725,18725,18725,18725,18725,18725]
; SSE-NEXT: pmulhw %xmm2, %xmm0
; SSE-NEXT: movdqa %xmm0, %xmm3
; SSE-NEXT: psrlw $15, %xmm3
; SSE-NEXT: psraw $1, %xmm0
; SSE-NEXT: paddw %xmm3, %xmm0
; SSE-NEXT: pmulhw %xmm2, %xmm1
; SSE-NEXT: movdqa %xmm1, %xmm2
; SSE-NEXT: psrlw $15, %xmm2
; SSE-NEXT: psraw $1, %xmm1
; SSE-NEXT: paddw %xmm2, %xmm1
; SSE-NEXT: retq
;
; AVX-LABEL: test6:
; AVX: # BB#0:
; AVX-NEXT: vpmulhw {{.*}}(%rip), %ymm0, %ymm0
; AVX-NEXT: vpsrlw $15, %ymm0, %ymm1
; AVX-NEXT: vpsraw $1, %ymm0, %ymm0
; AVX-NEXT: vpaddw %ymm1, %ymm0, %ymm0
; AVX-NEXT: retq
%div = sdiv <16 x i16> %a, <i16 7, i16 7, i16 7, i16 7,i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7,i16 7, i16 7, i16 7, i16 7>
ret <16 x i16> %div
}
define <16 x i8> @test7(<16 x i8> %a) #0 {
; SSE41-LABEL: test7:
; SSE41: # BB#0:
; SSE41-NEXT: pextrb $1, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pextrb $0, %xmm0, %ecx
; SSE41-NEXT: movsbl %cl, %ecx
; SSE41-NEXT: imull $-109, %ecx, %edx
; SSE41-NEXT: shrl $8, %edx
; SSE41-NEXT: addb %dl, %cl
; SSE41-NEXT: movb %cl, %dl
; SSE41-NEXT: shrb $7, %dl
; SSE41-NEXT: sarb $2, %cl
; SSE41-NEXT: addb %dl, %cl
; SSE41-NEXT: movzbl %cl, %ecx
; SSE41-NEXT: movd %ecx, %xmm1
; SSE41-NEXT: pinsrb $1, %eax, %xmm1
; SSE41-NEXT: pextrb $2, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $2, %eax, %xmm1
; SSE41-NEXT: pextrb $3, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $3, %eax, %xmm1
; SSE41-NEXT: pextrb $4, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $4, %eax, %xmm1
; SSE41-NEXT: pextrb $5, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $5, %eax, %xmm1
; SSE41-NEXT: pextrb $6, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $6, %eax, %xmm1
; SSE41-NEXT: pextrb $7, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $7, %eax, %xmm1
; SSE41-NEXT: pextrb $8, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $8, %eax, %xmm1
; SSE41-NEXT: pextrb $9, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $9, %eax, %xmm1
; SSE41-NEXT: pextrb $10, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $10, %eax, %xmm1
; SSE41-NEXT: pextrb $11, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $11, %eax, %xmm1
; SSE41-NEXT: pextrb $12, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $12, %eax, %xmm1
; SSE41-NEXT: pextrb $13, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $13, %eax, %xmm1
; SSE41-NEXT: pextrb $14, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $14, %eax, %xmm1
; SSE41-NEXT: pextrb $15, %xmm0, %eax
; SSE41-NEXT: movsbl %al, %eax
; SSE41-NEXT: imull $-109, %eax, %ecx
; SSE41-NEXT: shrl $8, %ecx
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movb %al, %cl
; SSE41-NEXT: shrb $7, %cl
; SSE41-NEXT: sarb $2, %al
; SSE41-NEXT: addb %cl, %al
; SSE41-NEXT: movzbl %al, %eax
; SSE41-NEXT: pinsrb $15, %eax, %xmm1
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: retq
;
; SSE-LABEL: test7:
; SSE: # BB#0:
; SSE-NEXT: pushq %rbp
; SSE-NEXT: pushq %r14
; SSE-NEXT: pushq %rbx
; SSE-NEXT: movaps %xmm0, -{{[0-9]+}}(%rsp)
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %eax
; SSE-NEXT: imull $-109, %eax, %ecx
; SSE-NEXT: shrl $8, %ecx
; SSE-NEXT: addb %al, %cl
; SSE-NEXT: movb %cl, %al
; SSE-NEXT: shrb $7, %al
; SSE-NEXT: sarb $2, %cl
; SSE-NEXT: addb %al, %cl
; SSE-NEXT: movzbl %cl, %eax
; SSE-NEXT: movd %eax, %xmm0
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %r14d
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %edx
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %r9d
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %eax
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %r11d
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %ecx
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %r8d
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %esi
; SSE-NEXT: imull $-109, %esi, %edi
; SSE-NEXT: shrl $8, %edi
; SSE-NEXT: addb %sil, %dil
; SSE-NEXT: movb %dil, %bl
; SSE-NEXT: shrb $7, %bl
; SSE-NEXT: sarb $2, %dil
; SSE-NEXT: addb %bl, %dil
; SSE-NEXT: movzbl %dil, %esi
; SSE-NEXT: movd %esi, %xmm1
; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; SSE-NEXT: imull $-109, %eax, %esi
; SSE-NEXT: shrl $8, %esi
; SSE-NEXT: addb %al, %sil
; SSE-NEXT: movb %sil, %al
; SSE-NEXT: shrb $7, %al
; SSE-NEXT: sarb $2, %sil
; SSE-NEXT: addb %al, %sil
; SSE-NEXT: movzbl %sil, %eax
; SSE-NEXT: movd %eax, %xmm2
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %ebp
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %esi
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %r10d
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %edi
; SSE-NEXT: imull $-109, %edi, %ebx
; SSE-NEXT: shrl $8, %ebx
; SSE-NEXT: addb %dil, %bl
; SSE-NEXT: movb %bl, %al
; SSE-NEXT: shrb $7, %al
; SSE-NEXT: sarb $2, %bl
; SSE-NEXT: addb %al, %bl
; SSE-NEXT: movzbl %bl, %eax
; SSE-NEXT: movd %eax, %xmm0
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT: imull $-109, %edx, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %dl, %al
; SSE-NEXT: movb %al, %dl
; SSE-NEXT: shrb $7, %dl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %dl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm1
; SSE-NEXT: imull $-109, %esi, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %sil, %al
; SSE-NEXT: movb %al, %dl
; SSE-NEXT: shrb $7, %dl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %dl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm2
; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSE-NEXT: imull $-109, %ecx, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %cl, %al
; SSE-NEXT: movb %al, %cl
; SSE-NEXT: shrb $7, %cl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %cl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm3
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %ecx
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %eax
; SSE-NEXT: imull $-109, %eax, %edx
; SSE-NEXT: shrl $8, %edx
; SSE-NEXT: addb %al, %dl
; SSE-NEXT: movb %dl, %al
; SSE-NEXT: shrb $7, %al
; SSE-NEXT: sarb $2, %dl
; SSE-NEXT: addb %al, %dl
; SSE-NEXT: movzbl %dl, %eax
; SSE-NEXT: movd %eax, %xmm1
; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; SSE-NEXT: imull $-109, %r14d, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %r14b, %al
; SSE-NEXT: movb %al, %dl
; SSE-NEXT: shrb $7, %dl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %dl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm2
; SSE-NEXT: imull $-109, %ebp, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %bpl, %al
; SSE-NEXT: movb %al, %dl
; SSE-NEXT: shrb $7, %dl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %dl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm0
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; SSE-NEXT: imull $-109, %r11d, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %r11b, %al
; SSE-NEXT: movb %al, %dl
; SSE-NEXT: shrb $7, %dl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %dl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm3
; SSE-NEXT: imull $-109, %ecx, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %cl, %al
; SSE-NEXT: movb %al, %cl
; SSE-NEXT: shrb $7, %cl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %cl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm2
; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
; SSE-NEXT: imull $-109, %r9d, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %r9b, %al
; SSE-NEXT: movb %al, %cl
; SSE-NEXT: shrb $7, %cl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %cl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm0
; SSE-NEXT: imull $-109, %r10d, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %r10b, %al
; SSE-NEXT: movb %al, %cl
; SSE-NEXT: shrb $7, %cl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %cl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm3
; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
; SSE-NEXT: imull $-109, %r8d, %eax
; SSE-NEXT: shrl $8, %eax
; SSE-NEXT: addb %r8b, %al
; SSE-NEXT: movb %al, %cl
; SSE-NEXT: shrb $7, %cl
; SSE-NEXT: sarb $2, %al
; SSE-NEXT: addb %cl, %al
; SSE-NEXT: movzbl %al, %eax
; SSE-NEXT: movd %eax, %xmm4
; SSE-NEXT: movsbl -{{[0-9]+}}(%rsp), %eax
; SSE-NEXT: imull $-109, %eax, %ecx
; SSE-NEXT: shrl $8, %ecx
; SSE-NEXT: addb %al, %cl
; SSE-NEXT: movb %cl, %al
; SSE-NEXT: shrb $7, %al
; SSE-NEXT: sarb $2, %cl
; SSE-NEXT: addb %al, %cl
; SSE-NEXT: movzbl %cl, %eax
; SSE-NEXT: movd %eax, %xmm0
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3],xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT: popq %rbx
; SSE-NEXT: popq %r14
; SSE-NEXT: popq %rbp
; SSE-NEXT: retq
;
; AVX-LABEL: test7:
; AVX: # BB#0:
; AVX-NEXT: vpextrb $1, %xmm0, %eax
; AVX-NEXT: movsbl %al, %eax
; AVX-NEXT: imull $-109, %eax, %ecx
; AVX-NEXT: shrl $8, %ecx
; AVX-NEXT: addb %cl, %al
; AVX-NEXT: movb %al, %cl
; AVX-NEXT: shrb $7, %cl
; AVX-NEXT: sarb $2, %al
; AVX-NEXT: addb %cl, %al
; AVX-NEXT: movzbl %al, %eax
; AVX-NEXT: vpextrb $0, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %dl
; AVX-NEXT: shrb $7, %dl
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movzbl %cl, %ecx
; AVX-NEXT: vmovd %ecx, %xmm1
; AVX-NEXT: vpextrb $2, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $1, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $3, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $2, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $4, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $3, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $5, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $6, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $5, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $7, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $6, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $8, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $7, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $9, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $10, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $9, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $11, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $10, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $12, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $11, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $13, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $14, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $13, %eax, %xmm1, %xmm1
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpextrb $15, %xmm0, %ecx
; AVX-NEXT: movsbl %cl, %ecx
; AVX-NEXT: imull $-109, %ecx, %edx
; AVX-NEXT: vpinsrb $14, %eax, %xmm1, %xmm0
; AVX-NEXT: shrl $8, %edx
; AVX-NEXT: addb %dl, %cl
; AVX-NEXT: movb %cl, %al
; AVX-NEXT: shrb $7, %al
; AVX-NEXT: sarb $2, %cl
; AVX-NEXT: addb %al, %cl
; AVX-NEXT: movzbl %cl, %eax
; AVX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; AVX-NEXT: retq
%div = sdiv <16 x i8> %a, <i8 7, i8 7, i8 7, i8 7,i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7,i8 7, i8 7, i8 7, i8 7>
ret <16 x i8> %div
}
define <4 x i32> @test8(<4 x i32> %a) #0 {
; SSE41-LABEL: test8:
; SSE41: # BB#0:
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [2454267027,2454267027,2454267027,2454267027]
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
; SSE41-NEXT: pmuldq %xmm2, %xmm3
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pmuldq %xmm0, %xmm1
; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2,3],xmm1[4,5],xmm3[6,7]
; SSE41-NEXT: paddd %xmm0, %xmm1
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: psrld $31, %xmm0
; SSE41-NEXT: psrad $2, %xmm1
; SSE41-NEXT: paddd %xmm0, %xmm1
; SSE41-NEXT: movdqa %xmm1, %xmm0
; SSE41-NEXT: retq
;
; SSE-LABEL: test8:
; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [2454267027,2454267027,2454267027,2454267027]
; SSE-NEXT: movdqa %xmm0, %xmm2
; SSE-NEXT: psrad $31, %xmm2
; SSE-NEXT: pand %xmm1, %xmm2
; SSE-NEXT: movdqa %xmm0, %xmm3
; SSE-NEXT: pmuludq %xmm1, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
; SSE-NEXT: psrad $31, %xmm1
; SSE-NEXT: pand %xmm0, %xmm1
; SSE-NEXT: paddd %xmm1, %xmm2
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,3,2,3]
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
; SSE-NEXT: pmuludq %xmm4, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
; SSE-NEXT: psubd %xmm2, %xmm1
; SSE-NEXT: paddd %xmm0, %xmm1
; SSE-NEXT: movdqa %xmm1, %xmm0
; SSE-NEXT: psrld $31, %xmm0
; SSE-NEXT: psrad $2, %xmm1
; SSE-NEXT: paddd %xmm0, %xmm1
; SSE-NEXT: movdqa %xmm1, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: test8:
; AVX: # BB#0:
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %xmm1
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; AVX-NEXT: vpmuldq %xmm2, %xmm3, %xmm2
; AVX-NEXT: vpmuldq %xmm1, %xmm0, %xmm1
; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; AVX-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2],xmm2[3]
; AVX-NEXT: vpaddd %xmm0, %xmm1, %xmm0
; AVX-NEXT: vpsrld $31, %xmm0, %xmm1
; AVX-NEXT: vpsrad $2, %xmm0, %xmm0
; AVX-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX-NEXT: retq
%div = sdiv <4 x i32> %a, <i32 7, i32 7, i32 7, i32 7>
ret <4 x i32> %div
}
define <8 x i32> @test9(<8 x i32> %a) #0 {
; SSE41-LABEL: test9:
; SSE41: # BB#0:
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [2454267027,2454267027,2454267027,2454267027]
; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
; SSE41-NEXT: pmuldq %xmm4, %xmm5
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: pmuldq %xmm3, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3],xmm2[4,5],xmm5[6,7]
; SSE41-NEXT: paddd %xmm0, %xmm2
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: psrld $31, %xmm0
; SSE41-NEXT: psrad $2, %xmm2
; SSE41-NEXT: paddd %xmm0, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
; SSE41-NEXT: pmuldq %xmm4, %xmm0
; SSE41-NEXT: pmuldq %xmm1, %xmm3
; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,3],xmm3[4,5],xmm0[6,7]
; SSE41-NEXT: paddd %xmm1, %xmm3
; SSE41-NEXT: movdqa %xmm3, %xmm0
; SSE41-NEXT: psrld $31, %xmm0
; SSE41-NEXT: psrad $2, %xmm3
; SSE41-NEXT: paddd %xmm0, %xmm3
; SSE41-NEXT: movdqa %xmm2, %xmm0
; SSE41-NEXT: movdqa %xmm3, %xmm1
; SSE41-NEXT: retq
;
; SSE-LABEL: test9:
; SSE: # BB#0:
; SSE-NEXT: movdqa %xmm0, %xmm2
; SSE-NEXT: movdqa {{.*#+}} xmm3 = [2454267027,2454267027,2454267027,2454267027]
; SSE-NEXT: movdqa %xmm3, %xmm4
; SSE-NEXT: psrad $31, %xmm4
; SSE-NEXT: movdqa %xmm4, %xmm0
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: movdqa %xmm2, %xmm5
; SSE-NEXT: psrad $31, %xmm5
; SSE-NEXT: pand %xmm3, %xmm5
; SSE-NEXT: paddd %xmm0, %xmm5
; SSE-NEXT: movdqa %xmm2, %xmm0
; SSE-NEXT: pmuludq %xmm3, %xmm0
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm3[1,1,3,3]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
; SSE-NEXT: pmuludq %xmm6, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]
; SSE-NEXT: psubd %xmm5, %xmm0
; SSE-NEXT: paddd %xmm2, %xmm0
; SSE-NEXT: movdqa %xmm0, %xmm2
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: psrld $31, %xmm2
; SSE-NEXT: psrad $2, %xmm0
; SSE-NEXT: paddd %xmm2, %xmm0
; SSE-NEXT: pand %xmm1, %xmm4
; SSE-NEXT: movdqa %xmm1, %xmm5
; SSE-NEXT: psrad $31, %xmm5
; SSE-NEXT: pand %xmm3, %xmm5
; SSE-NEXT: paddd %xmm4, %xmm5
; SSE-NEXT: pmuludq %xmm1, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,3,2,3]
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,1,3,3]
; SSE-NEXT: pmuludq %xmm6, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
; SSE-NEXT: psubd %xmm5, %xmm2
; SSE-NEXT: paddd %xmm1, %xmm2
; SSE-NEXT: movdqa %xmm2, %xmm1
; SSE-NEXT: psrld $31, %xmm1
; SSE-NEXT: psrad $2, %xmm2
; SSE-NEXT: paddd %xmm1, %xmm2
; SSE-NEXT: movdqa %xmm2, %xmm1
; SSE-NEXT: retq
;
; AVX-LABEL: test9:
; AVX: # BB#0:
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %ymm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} ymm2 = ymm1[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpshufd {{.*#+}} ymm3 = ymm0[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpmuldq %ymm2, %ymm3, %ymm2
; AVX-NEXT: vpmuldq %ymm1, %ymm0, %ymm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7]
; AVX-NEXT: vpaddd %ymm0, %ymm1, %ymm0
; AVX-NEXT: vpsrld $31, %ymm0, %ymm1
; AVX-NEXT: vpsrad $2, %ymm0, %ymm0
; AVX-NEXT: vpaddd %ymm1, %ymm0, %ymm0
; AVX-NEXT: retq
%div = sdiv <8 x i32> %a, <i32 7, i32 7, i32 7, i32 7,i32 7, i32 7, i32 7, i32 7>
ret <8 x i32> %div
}
define <8 x i32> @test10(<8 x i32> %a) #0 {
; SSE41-LABEL: test10:
; SSE41: # BB#0:
; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [613566757,613566757,613566757,613566757]
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
; SSE41-NEXT: pmuludq %xmm3, %xmm4
; SSE41-NEXT: movdqa %xmm0, %xmm5
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pmuludq %xmm2, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm5 = xmm5[0,1],xmm4[2,3],xmm5[4,5],xmm4[6,7]
; SSE41-NEXT: movdqa %xmm0, %xmm4
; SSE41-NEXT: psubd %xmm5, %xmm4
; SSE41-NEXT: psrld $1, %xmm4
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: paddd %xmm5, %xmm4
; SSE41-NEXT: psrld $2, %xmm4
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: movdqa {{.*#+}} xmm5 = [7,7,7,7]
; SSE41-NEXT: pmulld %xmm5, %xmm4
; SSE41-NEXT: psubd %xmm4, %xmm0
; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
; SSE41-NEXT: pmuludq %xmm3, %xmm4
; SSE41-NEXT: pmuludq %xmm1, %xmm2
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3],xmm2[4,5],xmm4[6,7]
; SSE41-NEXT: movdqa %xmm1, %xmm3
; SSE41-NEXT: psubd %xmm2, %xmm3
; SSE41-NEXT: psrld $1, %xmm3
; SSE41-NEXT: paddd %xmm2, %xmm3
; SSE41-NEXT: psrld $2, %xmm3
; SSE41-NEXT: pmulld %xmm5, %xmm3
; SSE41-NEXT: psubd %xmm3, %xmm1
; SSE41-NEXT: retq
;
; SSE-LABEL: test10:
; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm3 = [613566757,613566757,613566757,613566757]
; SSE-NEXT: movdqa %xmm0, %xmm2
; SSE-NEXT: pmuludq %xmm3, %xmm2
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
; SSE-NEXT: pmuludq %xmm4, %xmm5
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
; SSE-NEXT: movdqa %xmm0, %xmm5
; SSE-NEXT: psubd %xmm2, %xmm5
; SSE-NEXT: psrld $1, %xmm5
; SSE-NEXT: paddd %xmm2, %xmm5
; SSE-NEXT: psrld $2, %xmm5
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [7,7,7,7]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
; SSE-NEXT: pmuludq %xmm2, %xmm5
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3]
; SSE-NEXT: pmuludq %xmm2, %xmm6
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]
; SSE-NEXT: psubd %xmm5, %xmm0
; SSE-NEXT: pmuludq %xmm1, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,3,2,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
; SSE-NEXT: pmuludq %xmm4, %xmm5
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm5[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]
; SSE-NEXT: movdqa %xmm1, %xmm4
; SSE-NEXT: psubd %xmm3, %xmm4
; SSE-NEXT: psrld $1, %xmm4
; SSE-NEXT: paddd %xmm3, %xmm4
; SSE-NEXT: psrld $2, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
; SSE-NEXT: pmuludq %xmm2, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,2,2,3]
; SSE-NEXT: pmuludq %xmm2, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,2,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: psubd %xmm4, %xmm1
; SSE-NEXT: retq
;
; AVX-LABEL: test10:
; AVX: # BB#0:
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %ymm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} ymm2 = ymm1[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpshufd {{.*#+}} ymm3 = ymm0[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpmuludq %ymm2, %ymm3, %ymm2
; AVX-NEXT: vpmuludq %ymm1, %ymm0, %ymm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7]
; AVX-NEXT: vpsubd %ymm1, %ymm0, %ymm2
; AVX-NEXT: vpsrld $1, %ymm2, %ymm2
; AVX-NEXT: vpaddd %ymm1, %ymm2, %ymm1
; AVX-NEXT: vpsrld $2, %ymm1, %ymm1
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %ymm2
; AVX-NEXT: vpmulld %ymm2, %ymm1, %ymm1
; AVX-NEXT: vpsubd %ymm1, %ymm0, %ymm0
; AVX-NEXT: retq
%rem = urem <8 x i32> %a, <i32 7, i32 7, i32 7, i32 7,i32 7, i32 7, i32 7, i32 7>
ret <8 x i32> %rem
}
define <8 x i32> @test11(<8 x i32> %a) #0 {
; SSE41-LABEL: test11:
; SSE41: # BB#0:
; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [2454267027,2454267027,2454267027,2454267027]
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
; SSE41-NEXT: pmuldq %xmm3, %xmm4
; SSE41-NEXT: movdqa %xmm0, %xmm5
; SSE41-NEXT: pmuldq %xmm2, %xmm5
; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm5 = xmm5[0,1],xmm4[2,3],xmm5[4,5],xmm4[6,7]
; SSE41-NEXT: paddd %xmm0, %xmm5
; SSE41-NEXT: movdqa %xmm5, %xmm4
; SSE41-NEXT: psrld $31, %xmm4
; SSE41-NEXT: psrad $2, %xmm5
; SSE41-NEXT: paddd %xmm4, %xmm5
; SSE41-NEXT: movdqa {{.*#+}} xmm4 = [7,7,7,7]
; SSE41-NEXT: pmulld %xmm4, %xmm5
; SSE41-NEXT: psubd %xmm5, %xmm0
; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
; SSE41-NEXT: pmuldq %xmm3, %xmm5
; SSE41-NEXT: pmuldq %xmm1, %xmm2
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3],xmm2[4,5],xmm5[6,7]
; SSE41-NEXT: paddd %xmm1, %xmm2
; SSE41-NEXT: movdqa %xmm2, %xmm3
; SSE41-NEXT: psrld $31, %xmm3
; SSE41-NEXT: psrad $2, %xmm2
; SSE41-NEXT: paddd %xmm3, %xmm2
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pmulld %xmm4, %xmm2
; SSE41-NEXT: psubd %xmm2, %xmm1
; SSE41-NEXT: retq
;
; SSE-LABEL: test11:
; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [2454267027,2454267027,2454267027,2454267027]
; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: psrad $31, %xmm3
; SSE-NEXT: movdqa %xmm3, %xmm4
; SSE-NEXT: pand %xmm0, %xmm4
; SSE-NEXT: movdqa %xmm0, %xmm6
; SSE-NEXT: psrad $31, %xmm6
; SSE-NEXT: pand %xmm2, %xmm6
; SSE-NEXT: paddd %xmm4, %xmm6
; SSE-NEXT: movdqa %xmm0, %xmm4
; SSE-NEXT: pmuludq %xmm2, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm4[1,3,2,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
; SSE-NEXT: pmuludq %xmm5, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm4[0],xmm7[1],xmm4[1]
; SSE-NEXT: psubd %xmm6, %xmm7
; SSE-NEXT: paddd %xmm0, %xmm7
; SSE-NEXT: movdqa %xmm7, %xmm4
; SSE-NEXT: psrld $31, %xmm4
; SSE-NEXT: psrad $2, %xmm7
; SSE-NEXT: paddd %xmm4, %xmm7
; SSE-NEXT: movdqa {{.*#+}} xmm4 = [7,7,7,7]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm7[1,1,3,3]
; SSE-NEXT: pmuludq %xmm4, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3]
; SSE-NEXT: pmuludq %xmm4, %xmm6
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: psubd %xmm7, %xmm0
; SSE-NEXT: pand %xmm1, %xmm3
; SSE-NEXT: movdqa %xmm1, %xmm6
; SSE-NEXT: psrad $31, %xmm6
; SSE-NEXT: pand %xmm2, %xmm6
; SSE-NEXT: paddd %xmm3, %xmm6
; SSE-NEXT: pmuludq %xmm1, %xmm2
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,1,3,3]
; SSE-NEXT: pmuludq %xmm5, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
; SSE-NEXT: psubd %xmm6, %xmm2
; SSE-NEXT: paddd %xmm1, %xmm2
; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: psrld $31, %xmm3
; SSE-NEXT: psrad $2, %xmm2
; SSE-NEXT: paddd %xmm3, %xmm2
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
; SSE-NEXT: pmuludq %xmm4, %xmm2
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
; SSE-NEXT: pmuludq %xmm4, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,2,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
; SSE-NEXT: psubd %xmm2, %xmm1
; SSE-NEXT: retq
;
; AVX-LABEL: test11:
; AVX: # BB#0:
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %ymm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} ymm2 = ymm1[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpshufd {{.*#+}} ymm3 = ymm0[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpmuldq %ymm2, %ymm3, %ymm2
; AVX-NEXT: vpmuldq %ymm1, %ymm0, %ymm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; AVX-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,1,3,3,5,5,7,7]
; AVX-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7]
; AVX-NEXT: vpaddd %ymm0, %ymm1, %ymm1
; AVX-NEXT: vpsrld $31, %ymm1, %ymm2
; AVX-NEXT: vpsrad $2, %ymm1, %ymm1
; AVX-NEXT: vpaddd %ymm2, %ymm1, %ymm1
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %ymm2
; AVX-NEXT: vpmulld %ymm2, %ymm1, %ymm1
; AVX-NEXT: vpsubd %ymm1, %ymm0, %ymm0
; AVX-NEXT: retq
%rem = srem <8 x i32> %a, <i32 7, i32 7, i32 7, i32 7,i32 7, i32 7, i32 7, i32 7>
ret <8 x i32> %rem
}
define <2 x i16> @test12() #0 {
; SSE41-LABEL: test12:
; SSE41: # BB#0:
; SSE41-NEXT: xorps %xmm0, %xmm0
; SSE41-NEXT: retq
;
; SSE-LABEL: test12:
; SSE: # BB#0:
; SSE-NEXT: xorps %xmm0, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: test12:
; AVX: # BB#0:
; AVX-NEXT: vxorps %xmm0, %xmm0, %xmm0
; AVX-NEXT: retq
%I8 = insertelement <2 x i16> zeroinitializer, i16 -1, i32 0
%I9 = insertelement <2 x i16> %I8, i16 -1, i32 1
%B9 = urem <2 x i16> %I9, %I9
ret <2 x i16> %B9
}
[x86] Fix PR20355 (for real). There are many layers to this bug. The tale starts with r212808 which attempted to fix inversion of the low and high bits when lowering MUL_LOHI. Sadly, that commit did not include any positive test cases, and just removed some operations from a test case where the actual logic being changed isn't fully visible from the test. What this commit did was two things. First, it reversed the low and high results in the formation of the MERGE_VALUES node for the multiple results. This is entirely correct. Second it changed the shuffles for extracting the low and high components from the i64 results of the multiplies to extract them assuming a big-endian-style encoding of the multiply results. This second change is wrong. There is no big-endian encoding in x86, the results of the multiplies are normal v2i64s: when cast to v4i32, the low i32s are at offsets 0 and 2, and the high i32s are at offsets 1 and 3. However, the first change wasn't enough to actually fix the bug, which is (I assume) why the second change was also made. There was another bug in the MERGE_VALUES formation: we weren't using a VTList, and so were getting a single result node! When grabbing the *second* result from the node, we got... well.. colud be anything. I think this *appeared* to invert things, but had to be causing other problems as well. Fortunately, I fixed the MERGE_VALUES issue in r213931, so we should have been fine, right? NOOOPE! Because the core bug was never addressed, the test in vector-idiv failed when I fixed the MERGE_VALUES node. Because there are essentially no docs for this node, I had to guess at how to fix it and tried swapping the operands, restoring the order of the original code before r212808. While this "fixed" the test case (in that we produced the write instructions) we were still extracting the wrong elements of the i64s, and thus PR20355 was still broken. This commit essentially reverts the big-endian-style extraction part of r212808 and goes back to the original masks which were correct. Now that the MERGE_VALUES node formation is also correct, everything works. I've also included a more detailed test from PR20355 to make sure this stays fixed. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@214011 91177308-0d34-0410-b5e6-96231b3b80d8
2014-07-26 03:46:57 +00:00
define <4 x i32> @PR20355(<4 x i32> %a) #0 {
; SSE41-LABEL: PR20355:
; SSE41: # BB#0: # %entry
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [1431655766,1431655766,1431655766,1431655766]
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
; SSE41-NEXT: pmuldq %xmm2, %xmm3
; SSE41-NEXT: pmuldq %xmm1, %xmm0
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2,3],xmm1[4,5],xmm3[6,7]
; SSE41-NEXT: movdqa %xmm1, %xmm0
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; SSE41-NEXT: psrld $31, %xmm0
; SSE41-NEXT: paddd %xmm1, %xmm0
; SSE41-NEXT: retq
;
[x86] Fix PR20355 (for real). There are many layers to this bug. The tale starts with r212808 which attempted to fix inversion of the low and high bits when lowering MUL_LOHI. Sadly, that commit did not include any positive test cases, and just removed some operations from a test case where the actual logic being changed isn't fully visible from the test. What this commit did was two things. First, it reversed the low and high results in the formation of the MERGE_VALUES node for the multiple results. This is entirely correct. Second it changed the shuffles for extracting the low and high components from the i64 results of the multiplies to extract them assuming a big-endian-style encoding of the multiply results. This second change is wrong. There is no big-endian encoding in x86, the results of the multiplies are normal v2i64s: when cast to v4i32, the low i32s are at offsets 0 and 2, and the high i32s are at offsets 1 and 3. However, the first change wasn't enough to actually fix the bug, which is (I assume) why the second change was also made. There was another bug in the MERGE_VALUES formation: we weren't using a VTList, and so were getting a single result node! When grabbing the *second* result from the node, we got... well.. colud be anything. I think this *appeared* to invert things, but had to be causing other problems as well. Fortunately, I fixed the MERGE_VALUES issue in r213931, so we should have been fine, right? NOOOPE! Because the core bug was never addressed, the test in vector-idiv failed when I fixed the MERGE_VALUES node. Because there are essentially no docs for this node, I had to guess at how to fix it and tried swapping the operands, restoring the order of the original code before r212808. While this "fixed" the test case (in that we produced the write instructions) we were still extracting the wrong elements of the i64s, and thus PR20355 was still broken. This commit essentially reverts the big-endian-style extraction part of r212808 and goes back to the original masks which were correct. Now that the MERGE_VALUES node formation is also correct, everything works. I've also included a more detailed test from PR20355 to make sure this stays fixed. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@214011 91177308-0d34-0410-b5e6-96231b3b80d8
2014-07-26 03:46:57 +00:00
; SSE-LABEL: PR20355:
; SSE: # BB#0: # %entry
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [1431655766,1431655766,1431655766,1431655766]
; SSE-NEXT: movdqa %xmm1, %xmm2
; SSE-NEXT: psrad $31, %xmm2
; SSE-NEXT: pand %xmm0, %xmm2
; SSE-NEXT: movdqa %xmm0, %xmm3
; SSE-NEXT: psrad $31, %xmm3
; SSE-NEXT: pand %xmm1, %xmm3
; SSE-NEXT: paddd %xmm2, %xmm3
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219046 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-04 03:52:55 +00:00
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
; SSE-NEXT: pmuludq %xmm1, %xmm0
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,3,2,3]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
; SSE-NEXT: pmuludq %xmm2, %xmm0
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1]
; SSE-NEXT: psubd %xmm3, %xmm4
; SSE-NEXT: movdqa %xmm4, %xmm0
; SSE-NEXT: psrld $31, %xmm0
; SSE-NEXT: paddd %xmm4, %xmm0
[x86] Fix PR20355 (for real). There are many layers to this bug. The tale starts with r212808 which attempted to fix inversion of the low and high bits when lowering MUL_LOHI. Sadly, that commit did not include any positive test cases, and just removed some operations from a test case where the actual logic being changed isn't fully visible from the test. What this commit did was two things. First, it reversed the low and high results in the formation of the MERGE_VALUES node for the multiple results. This is entirely correct. Second it changed the shuffles for extracting the low and high components from the i64 results of the multiplies to extract them assuming a big-endian-style encoding of the multiply results. This second change is wrong. There is no big-endian encoding in x86, the results of the multiplies are normal v2i64s: when cast to v4i32, the low i32s are at offsets 0 and 2, and the high i32s are at offsets 1 and 3. However, the first change wasn't enough to actually fix the bug, which is (I assume) why the second change was also made. There was another bug in the MERGE_VALUES formation: we weren't using a VTList, and so were getting a single result node! When grabbing the *second* result from the node, we got... well.. colud be anything. I think this *appeared* to invert things, but had to be causing other problems as well. Fortunately, I fixed the MERGE_VALUES issue in r213931, so we should have been fine, right? NOOOPE! Because the core bug was never addressed, the test in vector-idiv failed when I fixed the MERGE_VALUES node. Because there are essentially no docs for this node, I had to guess at how to fix it and tried swapping the operands, restoring the order of the original code before r212808. While this "fixed" the test case (in that we produced the write instructions) we were still extracting the wrong elements of the i64s, and thus PR20355 was still broken. This commit essentially reverts the big-endian-style extraction part of r212808 and goes back to the original masks which were correct. Now that the MERGE_VALUES node formation is also correct, everything works. I've also included a more detailed test from PR20355 to make sure this stays fixed. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@214011 91177308-0d34-0410-b5e6-96231b3b80d8
2014-07-26 03:46:57 +00:00
; SSE-NEXT: retq
;
; AVX-LABEL: PR20355:
; AVX: # BB#0: # %entry
; AVX-NEXT: vpbroadcastd {{.*}}(%rip), %xmm1
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
; AVX-NEXT: vpmuldq %xmm2, %xmm3, %xmm2
; AVX-NEXT: vpmuldq %xmm1, %xmm0, %xmm0
[x86] Teach the 128-bit vector shuffle lowering routines to take advantage of the existence of a reasonable blend instruction. The 256-bit vector shuffle lowering has leveraged the general technique of decomposed shuffles and blends for quite some time, but this never made it back into the 128-bit code, and there are a large number of patterns where this is substantially better. For example, this removes almost all domain crossing in vector shuffles that involve some blend and some permutation with SSE4.1 and later. See the massive reduction in 'shufps' for integer test cases in this commit. This isn't perfect yet for a few reasons: 1) The v8i16 shuffle lowering continues to plague me. We don't always form an unpack-based blend when that would be better. But the wins pretty drastically outstrip the losses here. 2) The v16i8 shuffle lowering is just a disaster here. I never went and implemented blend support here for some terrible reason. I'll do that next probably. I've not updated it for now. More variations on this technique are coming as well -- we don't shuffle-into-unpack or shuffle-into-palignr, both of which would also be profitable. Note that some test cases grow significantly in the number of instructions, but I expect to actually be faster. We use pshufd+pshufd+blendw instead of a single shufps, but the pshufd's are very likely to pipeline well (two ports on most modern intel chips) and the blend is a *very* fast instruction. The domain switch penalty will essentially always be more than a blend instruction, which is the only increase in tree height. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229350 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-16 01:52:02 +00:00
; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
; AVX-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3]
; AVX-NEXT: vpsrld $31, %xmm0, %xmm1
; AVX-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX-NEXT: retq
[x86] Fix PR20355 (for real). There are many layers to this bug. The tale starts with r212808 which attempted to fix inversion of the low and high bits when lowering MUL_LOHI. Sadly, that commit did not include any positive test cases, and just removed some operations from a test case where the actual logic being changed isn't fully visible from the test. What this commit did was two things. First, it reversed the low and high results in the formation of the MERGE_VALUES node for the multiple results. This is entirely correct. Second it changed the shuffles for extracting the low and high components from the i64 results of the multiplies to extract them assuming a big-endian-style encoding of the multiply results. This second change is wrong. There is no big-endian encoding in x86, the results of the multiplies are normal v2i64s: when cast to v4i32, the low i32s are at offsets 0 and 2, and the high i32s are at offsets 1 and 3. However, the first change wasn't enough to actually fix the bug, which is (I assume) why the second change was also made. There was another bug in the MERGE_VALUES formation: we weren't using a VTList, and so were getting a single result node! When grabbing the *second* result from the node, we got... well.. colud be anything. I think this *appeared* to invert things, but had to be causing other problems as well. Fortunately, I fixed the MERGE_VALUES issue in r213931, so we should have been fine, right? NOOOPE! Because the core bug was never addressed, the test in vector-idiv failed when I fixed the MERGE_VALUES node. Because there are essentially no docs for this node, I had to guess at how to fix it and tried swapping the operands, restoring the order of the original code before r212808. While this "fixed" the test case (in that we produced the write instructions) we were still extracting the wrong elements of the i64s, and thus PR20355 was still broken. This commit essentially reverts the big-endian-style extraction part of r212808 and goes back to the original masks which were correct. Now that the MERGE_VALUES node formation is also correct, everything works. I've also included a more detailed test from PR20355 to make sure this stays fixed. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@214011 91177308-0d34-0410-b5e6-96231b3b80d8
2014-07-26 03:46:57 +00:00
entry:
%sdiv = sdiv <4 x i32> %a, <i32 3, i32 3, i32 3, i32 3>
ret <4 x i32> %sdiv
}
attributes #0 = { nounwind }