llvm-project/llvm/test/CodeGen/X86/sse2.ll

327 lines
14 KiB
LLVM
Raw Normal View History

2009-08-16 01:28:09 +08:00
; Tests for SSE2 and below, without SSE3+.
; RUN: llc < %s -mtriple=i386-apple-darwin10 -mcpu=pentium4 -O3 | FileCheck %s
define void @test1(<2 x double>* %r, <2 x double>* %A, double %B) nounwind {
; CHECK-LABEL: test1:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
; CHECK-NEXT: movapd (%ecx), %xmm0
; CHECK-NEXT: movlpd {{[0-9]+}}(%esp), %xmm0
; CHECK-NEXT: movapd %xmm0, (%eax)
; CHECK-NEXT: retl
%tmp3 = load <2 x double>, <2 x double>* %A, align 16
%tmp7 = insertelement <2 x double> undef, double %B, i32 0
%tmp9 = shufflevector <2 x double> %tmp3, <2 x double> %tmp7, <2 x i32> < i32 2, i32 1 >
store <2 x double> %tmp9, <2 x double>* %r, align 16
ret void
}
define void @test2(<2 x double>* %r, <2 x double>* %A, double %B) nounwind {
; CHECK-LABEL: test2:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
; CHECK-NEXT: movapd (%ecx), %xmm0
; CHECK-NEXT: movhpd {{[0-9]+}}(%esp), %xmm0
; CHECK-NEXT: movapd %xmm0, (%eax)
; CHECK-NEXT: retl
%tmp3 = load <2 x double>, <2 x double>* %A, align 16
%tmp7 = insertelement <2 x double> undef, double %B, i32 0
%tmp9 = shufflevector <2 x double> %tmp3, <2 x double> %tmp7, <2 x i32> < i32 0, i32 2 >
store <2 x double> %tmp9, <2 x double>* %r, align 16
ret void
}
define void @test3(<4 x float>* %res, <4 x float>* %A, <4 x float>* %B) nounwind {
; CHECK-LABEL: test3:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %edx
; CHECK-NEXT: movaps (%edx), %xmm0
; CHECK-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; CHECK-NEXT: movaps %xmm0, (%eax)
; CHECK-NEXT: retl
%tmp = load <4 x float>, <4 x float>* %B ; <<4 x float>> [#uses=2]
%tmp3 = load <4 x float>, <4 x float>* %A ; <<4 x float>> [#uses=2]
%tmp.upgrd.1 = extractelement <4 x float> %tmp3, i32 0 ; <float> [#uses=1]
%tmp7 = extractelement <4 x float> %tmp, i32 0 ; <float> [#uses=1]
%tmp8 = extractelement <4 x float> %tmp3, i32 1 ; <float> [#uses=1]
%tmp9 = extractelement <4 x float> %tmp, i32 1 ; <float> [#uses=1]
%tmp10 = insertelement <4 x float> undef, float %tmp.upgrd.1, i32 0 ; <<4 x float>> [#uses=1]
%tmp11 = insertelement <4 x float> %tmp10, float %tmp7, i32 1 ; <<4 x float>> [#uses=1]
%tmp12 = insertelement <4 x float> %tmp11, float %tmp8, i32 2 ; <<4 x float>> [#uses=1]
%tmp13 = insertelement <4 x float> %tmp12, float %tmp9, i32 3 ; <<4 x float>> [#uses=1]
store <4 x float> %tmp13, <4 x float>* %res
ret void
}
define void @test4(<4 x float> %X, <4 x float>* %res) nounwind {
; CHECK-LABEL: test4:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,1,3,3]
; CHECK-NEXT: movaps %xmm0, (%eax)
; CHECK-NEXT: retl
%tmp5 = shufflevector <4 x float> %X, <4 x float> undef, <4 x i32> < i32 2, i32 6, i32 3, i32 7 > ; <<4 x float>> [#uses=1]
store <4 x float> %tmp5, <4 x float>* %res
ret void
}
define <4 x i32> @test5(i8** %ptr) nounwind {
; CHECK-LABEL: test5:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movl (%eax), %eax
; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
; CHECK-NEXT: pxor %xmm0, %xmm0
; CHECK-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]
; CHECK-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; CHECK-NEXT: retl
%tmp = load i8*, i8** %ptr ; <i8*> [#uses=1]
%tmp.upgrd.1 = bitcast i8* %tmp to float* ; <float*> [#uses=1]
%tmp.upgrd.2 = load float, float* %tmp.upgrd.1 ; <float> [#uses=1]
%tmp.upgrd.3 = insertelement <4 x float> undef, float %tmp.upgrd.2, i32 0 ; <<4 x float>> [#uses=1]
%tmp9 = insertelement <4 x float> %tmp.upgrd.3, float 0.000000e+00, i32 1 ; <<4 x float>> [#uses=1]
%tmp10 = insertelement <4 x float> %tmp9, float 0.000000e+00, i32 2 ; <<4 x float>> [#uses=1]
%tmp11 = insertelement <4 x float> %tmp10, float 0.000000e+00, i32 3 ; <<4 x float>> [#uses=1]
%tmp21 = bitcast <4 x float> %tmp11 to <16 x i8> ; <<16 x i8>> [#uses=1]
%tmp22 = shufflevector <16 x i8> %tmp21, <16 x i8> zeroinitializer, <16 x i32> < i32 0, i32 16, i32 1, i32 17, i32 2, i32 18, i32 3, i32 19, i32 4, i32 20, i32 5, i32 21, i32 6, i32 22, i32 7, i32 23 > ; <<16 x i8>> [#uses=1]
%tmp31 = bitcast <16 x i8> %tmp22 to <8 x i16> ; <<8 x i16>> [#uses=1]
%tmp.upgrd.4 = shufflevector <8 x i16> zeroinitializer, <8 x i16> %tmp31, <8 x i32> < i32 0, i32 8, i32 1, i32 9, i32 2, i32 10, i32 3, i32 11 > ; <<8 x i16>> [#uses=1]
%tmp36 = bitcast <8 x i16> %tmp.upgrd.4 to <4 x i32> ; <<4 x i32>> [#uses=1]
ret <4 x i32> %tmp36
}
define void @test6(<4 x float>* %res, <4 x float>* %A) nounwind {
; CHECK-LABEL: test6:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
; CHECK-NEXT: movaps (%ecx), %xmm0
; CHECK-NEXT: movaps %xmm0, (%eax)
; CHECK-NEXT: retl
%tmp1 = load <4 x float>, <4 x float>* %A ; <<4 x float>> [#uses=1]
%tmp2 = shufflevector <4 x float> %tmp1, <4 x float> undef, <4 x i32> < i32 0, i32 5, i32 6, i32 7 > ; <<4 x float>> [#uses=1]
store <4 x float> %tmp2, <4 x float>* %res
ret void
}
define void @test7() nounwind {
; CHECK-LABEL: test7:
; CHECK: ## BB#0:
; CHECK-NEXT: xorps %xmm0, %xmm0
; CHECK-NEXT: movaps %xmm0, 0
; CHECK-NEXT: retl
bitcast <4 x i32> zeroinitializer to <4 x float> ; <<4 x float>>:1 [#uses=1]
shufflevector <4 x float> %1, <4 x float> zeroinitializer, <4 x i32> zeroinitializer ; <<4 x float>>:2 [#uses=1]
store <4 x float> %2, <4 x float>* null
ret void
}
@x = external global [4 x i32]
define <2 x i64> @test8() nounwind {
; CHECK-LABEL: test8:
; CHECK: ## BB#0:
; CHECK-NEXT: movl L_x$non_lazy_ptr, %eax
; CHECK-NEXT: movups (%eax), %xmm0
; CHECK-NEXT: retl
%tmp = load i32, i32* getelementptr ([4 x i32], [4 x i32]* @x, i32 0, i32 0) ; <i32> [#uses=1]
%tmp3 = load i32, i32* getelementptr ([4 x i32], [4 x i32]* @x, i32 0, i32 1) ; <i32> [#uses=1]
%tmp5 = load i32, i32* getelementptr ([4 x i32], [4 x i32]* @x, i32 0, i32 2) ; <i32> [#uses=1]
%tmp7 = load i32, i32* getelementptr ([4 x i32], [4 x i32]* @x, i32 0, i32 3) ; <i32> [#uses=1]
%tmp.upgrd.1 = insertelement <4 x i32> undef, i32 %tmp, i32 0 ; <<4 x i32>> [#uses=1]
%tmp13 = insertelement <4 x i32> %tmp.upgrd.1, i32 %tmp3, i32 1 ; <<4 x i32>> [#uses=1]
%tmp14 = insertelement <4 x i32> %tmp13, i32 %tmp5, i32 2 ; <<4 x i32>> [#uses=1]
%tmp15 = insertelement <4 x i32> %tmp14, i32 %tmp7, i32 3 ; <<4 x i32>> [#uses=1]
%tmp16 = bitcast <4 x i32> %tmp15 to <2 x i64> ; <<2 x i64>> [#uses=1]
ret <2 x i64> %tmp16
}
define <4 x float> @test9(i32 %dummy, float %a, float %b, float %c, float %d) nounwind {
; CHECK-LABEL: test9:
; CHECK: ## BB#0:
; CHECK-NEXT: movups {{[0-9]+}}(%esp), %xmm0
; CHECK-NEXT: retl
%tmp = insertelement <4 x float> undef, float %a, i32 0 ; <<4 x float>> [#uses=1]
%tmp11 = insertelement <4 x float> %tmp, float %b, i32 1 ; <<4 x float>> [#uses=1]
%tmp12 = insertelement <4 x float> %tmp11, float %c, i32 2 ; <<4 x float>> [#uses=1]
%tmp13 = insertelement <4 x float> %tmp12, float %d, i32 3 ; <<4 x float>> [#uses=1]
ret <4 x float> %tmp13
}
define <4 x float> @test10(float %a, float %b, float %c, float %d) nounwind {
; CHECK-LABEL: test10:
; CHECK: ## BB#0:
; CHECK-NEXT: movaps {{[0-9]+}}(%esp), %xmm0
; CHECK-NEXT: retl
%tmp = insertelement <4 x float> undef, float %a, i32 0 ; <<4 x float>> [#uses=1]
%tmp11 = insertelement <4 x float> %tmp, float %b, i32 1 ; <<4 x float>> [#uses=1]
%tmp12 = insertelement <4 x float> %tmp11, float %c, i32 2 ; <<4 x float>> [#uses=1]
%tmp13 = insertelement <4 x float> %tmp12, float %d, i32 3 ; <<4 x float>> [#uses=1]
ret <4 x float> %tmp13
}
define <2 x double> @test11(double %a, double %b) nounwind {
; CHECK-LABEL: test11:
; CHECK: ## BB#0:
; CHECK-NEXT: movaps {{[0-9]+}}(%esp), %xmm0
; CHECK-NEXT: retl
%tmp = insertelement <2 x double> undef, double %a, i32 0 ; <<2 x double>> [#uses=1]
%tmp7 = insertelement <2 x double> %tmp, double %b, i32 1 ; <<2 x double>> [#uses=1]
ret <2 x double> %tmp7
}
define void @test12() nounwind {
; CHECK-LABEL: test12:
; CHECK: ## BB#0:
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: movapd 0, %xmm0
; CHECK-NEXT: movapd {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00]
; CHECK-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],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. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: xorpd %xmm2, %xmm2
; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; CHECK-NEXT: addps %xmm1, %xmm0
; CHECK-NEXT: movaps %xmm0, 0
; CHECK-NEXT: retl
%tmp1 = load <4 x float>, <4 x float>* null ; <<4 x float>> [#uses=2]
%tmp2 = shufflevector <4 x float> %tmp1, <4 x float> < float 1.000000e+00, float 1.000000e+00, float 1.000000e+00, float 1.000000e+00 >, <4 x i32> < i32 0, i32 1, i32 6, i32 7 > ; <<4 x float>> [#uses=1]
%tmp3 = shufflevector <4 x float> %tmp1, <4 x float> zeroinitializer, <4 x i32> < i32 2, i32 3, i32 6, i32 7 > ; <<4 x float>> [#uses=1]
%tmp4 = fadd <4 x float> %tmp2, %tmp3 ; <<4 x float>> [#uses=1]
store <4 x float> %tmp4, <4 x float>* null
ret void
}
define void @test13(<4 x float>* %res, <4 x float>* %A, <4 x float>* %B, <4 x float>* %C) nounwind {
; CHECK-LABEL: test13:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %edx
; CHECK-NEXT: movaps (%edx), %xmm0
; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],mem[0,1]
[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. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2,1,3]
; CHECK-NEXT: movaps %xmm0, (%eax)
; CHECK-NEXT: retl
%tmp3 = load <4 x float>, <4 x float>* %B ; <<4 x float>> [#uses=1]
%tmp5 = load <4 x float>, <4 x float>* %C ; <<4 x float>> [#uses=1]
%tmp11 = shufflevector <4 x float> %tmp3, <4 x float> %tmp5, <4 x i32> < i32 1, i32 4, i32 1, i32 5 > ; <<4 x float>> [#uses=1]
store <4 x float> %tmp11, <4 x float>* %res
ret void
}
define <4 x float> @test14(<4 x float>* %x, <4 x float>* %y) nounwind {
; CHECK-LABEL: test14:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
; CHECK-NEXT: movaps (%ecx), %xmm1
; CHECK-NEXT: movaps (%eax), %xmm2
; CHECK-NEXT: movaps %xmm2, %xmm0
; CHECK-NEXT: addps %xmm1, %xmm0
; CHECK-NEXT: subps %xmm1, %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. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm2[0]
; CHECK-NEXT: retl
%tmp = load <4 x float>, <4 x float>* %y ; <<4 x float>> [#uses=2]
%tmp5 = load <4 x float>, <4 x float>* %x ; <<4 x float>> [#uses=2]
%tmp9 = fadd <4 x float> %tmp5, %tmp ; <<4 x float>> [#uses=1]
%tmp21 = fsub <4 x float> %tmp5, %tmp ; <<4 x float>> [#uses=1]
%tmp27 = shufflevector <4 x float> %tmp9, <4 x float> %tmp21, <4 x i32> < i32 0, i32 1, i32 4, i32 5 > ; <<4 x float>> [#uses=1]
ret <4 x float> %tmp27
}
define <4 x float> @test15(<4 x float>* %x, <4 x float>* %y) nounwind {
; CHECK-LABEL: test15:
; CHECK: ## BB#0: ## %entry
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: movapd (%ecx), %xmm0
; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],mem[1]
; CHECK-NEXT: retl
entry:
%tmp = load <4 x float>, <4 x float>* %y ; <<4 x float>> [#uses=1]
%tmp3 = load <4 x float>, <4 x float>* %x ; <<4 x float>> [#uses=1]
%tmp4 = shufflevector <4 x float> %tmp3, <4 x float> %tmp, <4 x i32> < i32 2, i32 3, i32 6, i32 7 > ; <<4 x float>> [#uses=1]
ret <4 x float> %tmp4
}
; PR8900
define <2 x double> @test16(<4 x double> * nocapture %srcA, <2 x double>* nocapture %dst) {
; CHECK-LABEL: test16:
; CHECK: ## BB#0:
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movapd 96(%eax), %xmm0
; CHECK-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
; CHECK-NEXT: retl
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction One of several parallel first steps to remove the target type of pointers, replacing them with a single opaque pointer type. This adds an explicit type parameter to the gep instruction so that when the first parameter becomes an opaque pointer type, the type to gep through is still available to the instructions. * This doesn't modify gep operators, only instructions (operators will be handled separately) * Textual IR changes only. Bitcode (including upgrade) and changing the in-memory representation will be in separate changes. * geps of vectors are transformed as: getelementptr <4 x float*> %x, ... ->getelementptr float, <4 x float*> %x, ... Then, once the opaque pointer type is introduced, this will ultimately look like: getelementptr float, <4 x ptr> %x with the unambiguous interpretation that it is a vector of pointers to float. * address spaces remain on the pointer, not the type: getelementptr float addrspace(1)* %x ->getelementptr float, float addrspace(1)* %x Then, eventually: getelementptr float, ptr addrspace(1) %x Importantly, the massive amount of test case churn has been automated by same crappy python code. I had to manually update a few test cases that wouldn't fit the script's model (r228970,r229196,r229197,r229198). The python script just massages stdin and writes the result to stdout, I then wrapped that in a shell script to handle replacing files, then using the usual find+xargs to migrate all the files. update.py: import fileinput import sys import re ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))") normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))") def conv(match, line): if not match: return line line = match.groups()[0] if len(match.groups()[5]) == 0: line += match.groups()[2] line += match.groups()[3] line += ", " line += match.groups()[1] line += "\n" return line for line in sys.stdin: if line.find("getelementptr ") == line.find("getelementptr inbounds"): if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("): line = conv(re.match(ibrep, line), line) elif line.find("getelementptr ") != line.find("getelementptr ("): line = conv(re.match(normrep, line), line) sys.stdout.write(line) apply.sh: for name in "$@" do python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name" rm -f "$name.tmp" done The actual commands: From llvm/src: find test/ -name *.ll | xargs ./apply.sh From llvm/src/tools/clang: find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}" From llvm/src/tools/polly: find test/ -name *.ll | xargs ./apply.sh After that, check-all (with llvm, clang, clang-tools-extra, lld, compiler-rt, and polly all checked out). The extra 'rm' in the apply.sh script is due to a few files in clang's test suite using interesting unicode stuff that my python script was throwing exceptions on. None of those files needed to be migrated, so it seemed sufficient to ignore those cases. Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7636 llvm-svn: 230786
2015-02-28 03:29:02 +08:00
%i5 = getelementptr inbounds <4 x double>, <4 x double>* %srcA, i32 3
%i6 = load <4 x double>, <4 x double>* %i5, align 32
%i7 = shufflevector <4 x double> %i6, <4 x double> undef, <2 x i32> <i32 0, i32 2>
ret <2 x double> %i7
}
; PR9009
define fastcc void @test17() nounwind {
; CHECK-LABEL: test17:
; CHECK: ## BB#0: ## %entry
; CHECK-NEXT: movaps {{.*#+}} xmm0 = <u,u,32768,32768>
; CHECK-NEXT: movaps %xmm0, (%eax)
; CHECK-NEXT: retl
entry:
%0 = insertelement <4 x i32> undef, i32 undef, i32 1
%1 = shufflevector <4 x i32> <i32 undef, i32 undef, i32 32768, i32 32768>, <4 x i32> %0, <4 x i32> <i32 4, i32 5, i32 2, i32 3>
%2 = bitcast <4 x i32> %1 to <4 x float>
store <4 x float> %2, <4 x float> * undef
ret void
}
; PR9210
define <4 x float> @f(<4 x double>) nounwind {
; CHECK-LABEL: f:
; CHECK: ## BB#0: ## %entry
; CHECK-NEXT: cvtpd2ps %xmm1, %xmm1
; CHECK-NEXT: cvtpd2ps %xmm0, %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. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; CHECK-NEXT: retl
entry:
%double2float.i = fptrunc <4 x double> %0 to <4 x float>
ret <4 x float> %double2float.i
}
define <2 x i64> @test_insert_64_zext(<2 x i64> %i) {
; CHECK-LABEL: test_insert_64_zext:
; CHECK: ## BB#0:
; CHECK-NEXT: movq {{.*#+}} xmm0 = xmm0[0],zero
; CHECK-NEXT: retl
%1 = shufflevector <2 x i64> %i, <2 x i64> <i64 0, i64 undef>, <2 x i32> <i32 0, i32 2>
ret <2 x i64> %1
}
define <4 x i32> @PR19721(<4 x i32> %i) {
; CHECK-LABEL: PR19721:
; CHECK: ## BB#0:
; CHECK-NEXT: andps LCPI19_0, %xmm0
; CHECK-NEXT: retl
%bc = bitcast <4 x i32> %i to i128
%insert = and i128 %bc, -4294967296
%bc2 = bitcast i128 %insert to <4 x i32>
ret <4 x i32> %bc2
}
define <4 x i32> @test_mul(<4 x i32> %x, <4 x i32> %y) {
; CHECK-LABEL: test_mul:
; CHECK: ## BB#0:
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
; CHECK-NEXT: pmuludq %xmm1, %xmm0
; CHECK-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
[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. llvm-svn: 219046
2014-10-04 11:52:55 +08:00
; CHECK-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
; CHECK-NEXT: pmuludq %xmm2, %xmm1
; CHECK-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; CHECK-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; CHECK-NEXT: retl
%m = mul <4 x i32> %x, %y
ret <4 x i32> %m
}