RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-23 19:23:23 +01:00
Code Issues Projects Releases Wiki Activity
369ae03621
llvm-mirror/test/CodeGen/X86/vec_cast2.ll

179 lines
7.2 KiB
LLVM
Raw Normal View History

Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
; RUN: llc < %s -mtriple=i386-apple-darwin10 -mcpu=corei7-avx -mattr=+avx | FileCheck %s
[x86] Fix a bug in my new zext-vector-inreg DAG trickery where we were not widening the input type to the node sufficiently to let the ext take place in a register. This would in turn result in a mysterious bitcast assertion failure downstream. First change here is to add back the helpful assert I had in an earlier version of the code to catch this immediately. Next change is to add support to the type legalization to detect when we have widened the operand either too little or too much (for whatever reason) and find a size-matched legal vector type to convert it to first. This can also fail so we get a new fallback path, but that seems OK. With this, we no longer crash on vec_cast2.ll when using widening. I've also added the CHECK lines for the zero-extend cases here. We still need to support sign-extend and trunc (or something) to get plausible code for the other two thirds of this test which is one of the regression tests that showed the most scalarization when widening was force-enabled. Slowly closing in on widening being a viable legalization strategy without it resorting to scalarization at every turn. =] llvm-svn: 212614
2014-07-09 14:36:54 +02:00
; RUN: llc < %s -mtriple=i386-apple-darwin10 -mcpu=corei7-avx -mattr=+avx -x86-experimental-vector-widening-legalization | FileCheck %s --check-prefix=CHECK-WIDE
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
define <8 x float> @foo1_8(<8 x i8> %src) {
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-LABEL: foo1_8:
; CHECK: ## BB#0:
; CHECK-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm0[4,4,5,5,6,6,7,7]
; CHECK-NEXT: vpmovzxwd %xmm0, %xmm0
; CHECK-NEXT: vpslld $24, %xmm0, %xmm0
; CHECK-NEXT: vpsrad $24, %xmm0, %xmm0
; CHECK-NEXT: vpslld $24, %xmm1, %xmm1
; CHECK-NEXT: vpsrad $24, %xmm1, %xmm1
; CHECK-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; CHECK-NEXT: vcvtdq2ps %ymm0, %ymm0
; CHECK-NEXT: retl
;
; CHECK-WIDE-LABEL: foo1_8:
; CHECK-WIDE: ## BB#0:
; CHECK-WIDE-NEXT: vpmovzxbd %xmm0, %xmm1
; CHECK-WIDE-NEXT: vpslld $24, %xmm1, %xmm1
; CHECK-WIDE-NEXT: vpsrad $24, %xmm1, %xmm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; CHECK-WIDE-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4,4,5,5,6,6,7,7]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vpslld $24, %xmm0, %xmm0
; CHECK-WIDE-NEXT: vpsrad $24, %xmm0, %xmm0
; CHECK-WIDE-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; CHECK-WIDE-NEXT: vcvtdq2ps %ymm0, %ymm0
; CHECK-WIDE-NEXT: retl
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
%res = sitofp <8 x i8> %src to <8 x float>
ret <8 x float> %res
}
define <4 x float> @foo1_4(<4 x i8> %src) {
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-LABEL: foo1_4:
; CHECK: ## BB#0:
; CHECK-NEXT: vpslld $24, %xmm0, %xmm0
; CHECK-NEXT: vpsrad $24, %xmm0, %xmm0
; CHECK-NEXT: vcvtdq2ps %xmm0, %xmm0
; CHECK-NEXT: retl
;
; CHECK-WIDE-LABEL: foo1_4:
; CHECK-WIDE: ## BB#0:
; CHECK-WIDE-NEXT: vpmovzxbd %xmm0, %xmm0
; CHECK-WIDE-NEXT: vpslld $24, %xmm0, %xmm0
; CHECK-WIDE-NEXT: vpsrad $24, %xmm0, %xmm0
; CHECK-WIDE-NEXT: vcvtdq2ps %xmm0, %xmm0
; CHECK-WIDE-NEXT: retl
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
%res = sitofp <4 x i8> %src to <4 x float>
ret <4 x float> %res
}
define <8 x float> @foo2_8(<8 x i8> %src) {
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-LABEL: foo2_8:
; 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 05:52:55 +02:00
; CHECK-NEXT: vpmovzxwd %xmm0, %xmm1
; CHECK-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4,4,5,5,6,6,7,7]
; CHECK-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-NEXT: vandps LCPI2_0, %ymm0, %ymm0
; CHECK-NEXT: vcvtdq2ps %ymm0, %ymm0
; CHECK-NEXT: retl
;
; CHECK-WIDE-LABEL: foo2_8:
; CHECK-WIDE: ## 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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vxorps %xmm1, %xmm1, %xmm1
; CHECK-WIDE-NEXT: vextractf128 $1, %ymm1, %xmm2
; CHECK-WIDE-NEXT: vmovdqa {{.*#+}} xmm3 = <1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u>
; CHECK-WIDE-NEXT: vpshufb %xmm3, %xmm2, %xmm4
; CHECK-WIDE-NEXT: vmovdqa {{.*#+}} xmm5 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u>
; CHECK-WIDE-NEXT: vpshufb %xmm5, %xmm2, %xmm2
[x86] Don't form overly fragmented blends when splitting and re-combining shuffles because nothing was available in the wider vector type. The key observation (which I've put in the comments for future maintainers) is that at this point, no further combining is really possible. And so even though these shuffles trivially could be combined, we need to actually do that as we produce them when producing them this late in the lowering. This fixes another (huge) part of the Halide vector shuffle regressions. As it happens, this was already well covered by the tests, but I hadn't noticed how bad some of these got. The specific patterns that turn directly into unpckl/h patterns were occurring *many* times in common vector processing code. There are still more problems here sadly, but trying to incrementally tease them apart and it looks like this is the core of the problem in the splitting logic. There is some chance of regression here, you can see it in the test changes. Specifically, where we stop forming pshufb in some cases, it is possible that pshufb was in fact faster. Intel "says" that pshufb is slower than the instruction sequences replacing it. llvm-svn: 221852
2014-11-13 03:42:08 +01:00
; CHECK-WIDE-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[1,1,2,3]
[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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3],xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
; CHECK-WIDE-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; CHECK-WIDE-NEXT: vpshufb %xmm3, %xmm1, %xmm3
; CHECK-WIDE-NEXT: vpshufb %xmm5, %xmm1, %xmm1
; CHECK-WIDE-NEXT: vpunpcklbw {{.*#+}} 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]
; CHECK-WIDE-NEXT: vpunpcklbw {{.*#+}} 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]
; CHECK-WIDE-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvtdq2ps %ymm0, %ymm0
; CHECK-WIDE-NEXT: retl
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
%res = uitofp <8 x i8> %src to <8 x float>
ret <8 x float> %res
}
define <4 x float> @foo2_4(<4 x i8> %src) {
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-LABEL: foo2_4:
; CHECK: ## BB#0:
; CHECK-NEXT: vandps LCPI3_0, %xmm0, %xmm0
; CHECK-NEXT: vcvtdq2ps %xmm0, %xmm0
; CHECK-NEXT: retl
;
; CHECK-WIDE-LABEL: foo2_4:
; CHECK-WIDE: ## 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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpmovzxbd %xmm0, %xmm0
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvtdq2ps %xmm0, %xmm0
; CHECK-WIDE-NEXT: retl
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
%res = uitofp <4 x i8> %src to <4 x float>
ret <4 x float> %res
}
define <8 x i8> @foo3_8(<8 x float> %src) {
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-LABEL: foo3_8:
; CHECK: ## BB#0:
; CHECK-NEXT: vcvttps2dq %ymm0, %ymm0
; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 05:52:55 +02:00
; CHECK-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; CHECK-NEXT: vpshufb %xmm2, %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 05:52:55 +02:00
; CHECK-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retl
;
; CHECK-WIDE-LABEL: foo3_8:
; CHECK-WIDE: ## 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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[3,1,2,3]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm1, %eax
; CHECK-WIDE-NEXT: shll $8, %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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm1, %ecx
; CHECK-WIDE-NEXT: movzbl %cl, %ecx
; CHECK-WIDE-NEXT: orl %eax, %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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[1,1,2,3]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm1, %eax
; CHECK-WIDE-NEXT: shll $8, %eax
; CHECK-WIDE-NEXT: vcvttss2si %xmm0, %edx
; CHECK-WIDE-NEXT: movzbl %dl, %edx
; CHECK-WIDE-NEXT: orl %eax, %edx
; CHECK-WIDE-NEXT: vpinsrw $0, %edx, %xmm0, %xmm1
; CHECK-WIDE-NEXT: vpinsrw $1, %ecx, %xmm1, %xmm1
; CHECK-WIDE-NEXT: vextractf128 $1, %ymm0, %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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilps {{.*#+}} xmm2 = xmm0[1,1,2,3]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm2, %eax
; CHECK-WIDE-NEXT: shll $8, %eax
; CHECK-WIDE-NEXT: vcvttss2si %xmm0, %ecx
; CHECK-WIDE-NEXT: movzbl %cl, %ecx
; CHECK-WIDE-NEXT: orl %eax, %ecx
; CHECK-WIDE-NEXT: vpinsrw $2, %ecx, %xmm1, %xmm1
[x86] Enable the new vector shuffle lowering by default. Update the entire regression test suite for the new shuffles. Remove most of the old testing which was devoted to the old shuffle lowering path and is no longer relevant really. Also remove a few other random tests that only really exercised shuffles and only incidently or without any interesting aspects to them. Benchmarking that I have done shows a few small regressions with this on LNT, zero measurable regressions on real, large applications, and for several benchmarks where the loop vectorizer fires in the hot path it shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy Bridge machines. Running on AMD machines shows even more dramatic improvements. When using newer ISA vector extensions the gains are much more modest, but the code is still better on the whole. There are a few regressions being tracked (PR21137, PR21138, PR21139) but by and large this is expected to be a win for x86 generated code performance. It is also more correct than the code it replaces. I have fuzz tested this extensively with ISA extensions up through AVX2 and found no crashes or miscompiles (yet...). The old lowering had a few miscompiles and crashers after a somewhat smaller amount of fuzz testing. There is one significant area where the new code path lags behind and that is in AVX-512 support. However, there was *extremely little* support for that already and so this isn't a significant step backwards and the new framework will probably make it easier to implement lowering that uses the full power of AVX-512's table-based shuffle+blend (IMO). Many thanks to Quentin, Andrea, Robert, and others for benchmarking assistance. Thanks to Adam and others for help with AVX-512. Thanks to Hal, Eric, and *many* others for answering my incessant questions about how the backend actually works. =] I will leave the old code path in the tree until the 3 PRs above are at least resolved to folks' satisfaction. Then I will rip it (and 1000s of lines of code) out. =] I don't expect this flag to stay around for very long. It may not survive next week. llvm-svn: 219046
2014-10-04 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilps {{.*#+}} xmm2 = xmm0[3,1,2,3]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm2, %eax
; CHECK-WIDE-NEXT: shll $8, %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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm0, %ecx
; CHECK-WIDE-NEXT: movzbl %cl, %ecx
; CHECK-WIDE-NEXT: orl %eax, %ecx
; CHECK-WIDE-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm0
; CHECK-WIDE-NEXT: vzeroupper
; CHECK-WIDE-NEXT: retl
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
%res = fptosi <8 x float> %src to <8 x i8>
ret <8 x i8> %res
}
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
define <4 x i8> @foo3_4(<4 x float> %src) {
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-LABEL: foo3_4:
; CHECK: ## BB#0:
; CHECK-NEXT: vcvttps2dq %xmm0, %xmm0
; CHECK-NEXT: retl
;
; CHECK-WIDE-LABEL: foo3_4:
; CHECK-WIDE: ## 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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[3,1,2,3]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm1, %eax
; CHECK-WIDE-NEXT: shll $8, %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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm1, %ecx
; CHECK-WIDE-NEXT: movzbl %cl, %ecx
; CHECK-WIDE-NEXT: orl %eax, %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 05:52:55 +02:00
; CHECK-WIDE-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[1,1,2,3]
[x86] Regenerate precise checks for a couple of test cases and remove a test case that was just grepping the debug stats output rather than actually checking the generated code for anything useful. llvm-svn: 218951
2014-10-03 03:50:08 +02:00
; CHECK-WIDE-NEXT: vcvttss2si %xmm1, %eax
; CHECK-WIDE-NEXT: shll $8, %eax
; CHECK-WIDE-NEXT: vcvttss2si %xmm0, %edx
; CHECK-WIDE-NEXT: movzbl %dl, %edx
; CHECK-WIDE-NEXT: orl %eax, %edx
; CHECK-WIDE-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
; CHECK-WIDE-NEXT: vpinsrw $1, %ecx, %xmm0, %xmm0
; CHECK-WIDE-NEXT: retl
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics). llvm-svn: 155397
2012-04-23 23:53:37 +02:00
%res = fptosi <4 x float> %src to <4 x i8>
ret <4 x i8> %res
}
Reference in New Issue Copy Permalink