maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-select.mir

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

1681 lines
86 KiB
Plaintext
Raw Normal View History

AMDGPU: Remove VS/SV mappings from select These would violate the constant bus restriction llvm-svn: 350517
2019-01-07 21:21:36 +08:00
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect -global-isel %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck -check-prefix=FAST %s
# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect -global-isel %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck -check-prefix=GREEDY %s
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_s32_scc_ss
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s32_scc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
; FAST: [[COPY3:%[0-9]+]]:sgpr(s32) = COPY $sgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:sgpr(s32) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_s32_scc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
; GREEDY: [[COPY3:%[0-9]+]]:sgpr(s32) = COPY $sgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:sgpr(s32) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s32) = COPY $sgpr2
%3:_(s32) = COPY $sgpr3
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s32) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_s32_scc_sv
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2, $vgpr0
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s32_scc_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; FAST: [[COPY5:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[COPY5]], [[COPY3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_s32_scc_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; GREEDY: [[COPY5:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[COPY5]], [[COPY3]]
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
AMDGPU: Remove VS/SV mappings from select These would violate the constant bus restriction llvm-svn: 350517
2019-01-07 21:21:36 +08:00
%2:_(s32) = COPY $sgpr2
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%3:_(s32) = COPY $vgpr0
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s32) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_s32_scc_vs
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2, $vgpr0
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s32_scc_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; FAST: [[COPY5:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[COPY3]], [[COPY5]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_s32_scc_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; GREEDY: [[COPY5:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[COPY3]], [[COPY5]]
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s32) = COPY $sgpr2
%3:_(s32) = COPY $vgpr0
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s32) = G_SELECT %4, %3, %2
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_s32_scc_vv
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s32_scc_vv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_s32_scc_vv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s32) = COPY $vgpr0
%3:_(s32) = COPY $vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s32) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_s32_vcc_ss
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s32_vcc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; FAST: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
; FAST: [[COPY5:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[COPY4]], [[COPY5]]
; GREEDY-LABEL: name: select_s32_vcc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; GREEDY: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
; GREEDY: [[COPY5:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[COPY4]], [[COPY5]]
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s32) = COPY $vgpr0
%3:_(s32) = COPY $vgpr1
%4:_(s1) = G_ICMP intpred(ne), %2, %3
%5:_(s32) = G_SELECT %4, %0, %1
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_s32_vcc_sv
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $vgpr0, $vgpr1, $vgpr2
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s32_vcc_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[COPY4]], [[COPY3]]
; GREEDY-LABEL: name: select_s32_vcc_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[COPY4]], [[COPY3]]
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(s32) = COPY $vgpr2
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(s32) = G_SELECT %4, %0, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_s32_vcc_vs
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $vgpr0, $vgpr1, $vgpr2
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s32_vcc_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[COPY3]], [[COPY4]]
; GREEDY-LABEL: name: select_s32_vcc_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[COPY3]], [[COPY4]]
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(s32) = COPY $vgpr2
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(s32) = G_SELECT %4, %3, %0
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_s32_vcc_vv
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
legalized: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s32_vcc_vv
; FAST: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[COPY2]], [[COPY3]]
; GREEDY-LABEL: name: select_s32_vcc_vv
; GREEDY: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Define instruction mapping for G_SELECT Reviewers: arsenm Reviewed By: arsenm Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye Differential Revision: https://reviews.llvm.org/D49737 llvm-svn: 341271
2018-09-01 10:41:19 +08:00
%0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1
%2:_(s32) = COPY $vgpr2
%3:_(s32) = COPY $vgpr3
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s32) = G_SELECT %4, %2, %3
...
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
---
name: select_s64_sss
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s64_sss
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:sgpr(s64) = COPY $sgpr4_sgpr5
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:sgpr(s64) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_s64_sss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:sgpr(s64) = COPY $sgpr4_sgpr5
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:sgpr(s64) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s64) = COPY $sgpr2_sgpr3
%3:_(s64) = COPY $sgpr4_sgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s64) = G_SELECT %4, %2, %3
...
---
name: select_s64_ssv
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s64_ssv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](s64)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_s64_ssv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](s64)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s64) = COPY $sgpr2_sgpr3
%3:_(s64) = COPY $vgpr0_vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s64) = G_SELECT %4, %2, %3
...
---
name: select_s64_svs
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s64_svs
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](s64)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_s64_svs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](s64)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s64) = COPY $sgpr2_sgpr3
%3:_(s64) = COPY $vgpr0_vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s64) = G_SELECT %4, %3, %2
...
---
name: select_s64_svv
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s64_svv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(s64) = COPY $vgpr0_vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](s64)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_s64_svv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s64) = COPY $vgpr0_vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](s64)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s64) = COPY $vgpr0_vgpr1
%3:_(s64) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s64) = G_SELECT %4, %2, %3
...
---
name: select_s64_vss
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $sgpr2_sgpr3, $vgpr0, $vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s64_vss
; FAST: [[COPY:%[0-9]+]]:sgpr(s64) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](s64)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY1]](s64)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_s64_vss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s64) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](s64)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY1]](s64)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s64) = COPY $sgpr0_sgpr1
%1:_(s64) = COPY $sgpr2_sgpr3
%2:_(s32) = COPY $vgpr0
%3:_(s32) = COPY $vgpr1
%4:_(s1) = G_ICMP intpred(ne), %2, %3
%5:_(s64) = G_SELECT %4, %0, %1
...
---
name: select_s64_vsv
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $vgpr0, $vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s64_vsv
; FAST: [[COPY:%[0-9]+]]:sgpr(s64) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](s64)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_s64_vsv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s64) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](s64)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s64) = COPY $sgpr0_sgpr1
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(s64) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(s64) = G_SELECT %4, %0, %3
...
---
name: select_s64_vvs
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $vgpr0, $vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s64_vvs
; FAST: [[COPY:%[0-9]+]]:sgpr(s64) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](s64)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_s64_vvs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s64) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](s64)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s64) = COPY $sgpr0_sgpr1
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(s64) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(s64) = G_SELECT %4, %3, %0
...
---
name: select_s64_vvv
legalized: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_s64_vvv
; FAST: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr4_vgpr5
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](s64)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_s64_vvv
; GREEDY: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s64) = COPY $vgpr4_vgpr5
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](s64)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](s64)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(s64) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1
%2:_(s64) = COPY $vgpr2_vgpr3
%3:_(s64) = COPY $vgpr4_vgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(s64) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v2s32_scc_ss
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v2s32_scc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr4_sgpr5
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:sgpr(<2 x s32>) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_v2s32_scc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr4_sgpr5
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:sgpr(<2 x s32>) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(<2 x s32>) = COPY $sgpr2_sgpr3
%3:_(<2 x s32>) = COPY $sgpr4_sgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<2 x s32>) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v2s32_scc_sv
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v2s32_scc_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](<2 x s32>)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_v2s32_scc_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](<2 x s32>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(<2 x s32>) = COPY $sgpr2_sgpr3
%3:_(<2 x s32>) = COPY $vgpr0_vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<2 x s32>) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v2s32_scc_vs
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v2s32_scc_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](<2 x s32>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_v2s32_scc_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](<2 x s32>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(<2 x s32>) = COPY $sgpr2_sgpr3
%3:_(<2 x s32>) = COPY $vgpr0_vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<2 x s32>) = G_SELECT %4, %3, %2
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v2s32_scc_vv
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v2s32_scc_vv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr0_vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr2_vgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](<2 x s32>)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_v2s32_scc_vv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr0_vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr2_vgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](<2 x s32>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(<2 x s32>) = COPY $vgpr0_vgpr1
%3:_(<2 x s32>) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<2 x s32>) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v2s32_vcc_ss
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $sgpr2_sgpr3, $vgpr0, $vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v2s32_vcc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr2_sgpr3
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY1]](<2 x s32>)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_v2s32_vcc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY1]](<2 x s32>)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(<2 x s32>) = COPY $sgpr0_sgpr1
%1:_(<2 x s32>) = COPY $sgpr2_sgpr3
%2:_(s32) = COPY $vgpr0
%3:_(s32) = COPY $vgpr1
%4:_(s1) = G_ICMP intpred(ne), %2, %3
%5:_(<2 x s32>) = G_SELECT %4, %0, %1
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v2s32_vcc_sv
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $vgpr0, $vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v2s32_vcc_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr2_vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_v2s32_vcc_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr2_vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(<2 x s32>) = COPY $sgpr0_sgpr1
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(<2 x s32>) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(<2 x s32>) = G_SELECT %4, %0, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v2s32_vcc_vs
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $vgpr0, $vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v2s32_vcc_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr2_vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_v2s32_vcc_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(<2 x s32>) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr2_vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(<2 x s32>) = COPY $sgpr0_sgpr1
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(<2 x s32>) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(<2 x s32>) = G_SELECT %4, %3, %0
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v2s32_vcc_vv
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
legalized: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v2s32_vcc_vv
; FAST: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr2_vgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr4_vgpr5
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](<2 x s32>)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_v2s32_vcc_vv
; GREEDY: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr2_vgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<2 x s32>) = COPY $vgpr4_vgpr5
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](<2 x s32>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](<2 x s32>)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[BUILD_VECTOR:%[0-9]+]]:vgpr(<2 x s32>) = G_BUILD_VECTOR [[SELECT]](s32), [[SELECT1]](s32)
AMDGPU/GlobalISel: Handle split for 64-bit VALU select llvm-svn: 354065
2019-02-15 05:58:12 +08:00
%0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1
%2:_(<2 x s32>) = COPY $vgpr2_vgpr3
%3:_(<2 x s32>) = COPY $vgpr4_vgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<2 x s32>) = G_SELECT %4, %2, %3
...
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v4s16_scc_ss
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v4s16_scc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr4_sgpr5
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:sgpr(<4 x s16>) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_v4s16_scc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr4_sgpr5
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:sgpr(<4 x s16>) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(<4 x s16>) = COPY $sgpr2_sgpr3
%3:_(<4 x s16>) = COPY $sgpr4_sgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<4 x s16>) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v4s16_scc_sv
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v4s16_scc_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
; FAST: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
; GREEDY-LABEL: name: select_v4s16_scc_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(<4 x s16>) = COPY $sgpr2_sgpr3
%3:_(<4 x s16>) = COPY $vgpr0_vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<4 x s16>) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v4s16_scc_vs
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v4s16_scc_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
; FAST: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
; GREEDY-LABEL: name: select_v4s16_scc_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(<4 x s16>) = COPY $sgpr2_sgpr3
%3:_(<4 x s16>) = COPY $vgpr0_vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<4 x s16>) = G_SELECT %4, %3, %2
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v4s16_scc_vv
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v4s16_scc_vv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
; FAST: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
; GREEDY-LABEL: name: select_v4s16_scc_vv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(<4 x s16>) = COPY $vgpr0_vgpr1
%3:_(<4 x s16>) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<4 x s16>) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v4s16_vcc_ss
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $sgpr2_sgpr3, $vgpr0, $vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v4s16_vcc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; FAST: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
; FAST: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
; FAST: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
; GREEDY-LABEL: name: select_v4s16_vcc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(<4 x s16>) = COPY $sgpr0_sgpr1
%1:_(<4 x s16>) = COPY $sgpr2_sgpr3
%2:_(s32) = COPY $vgpr0
%3:_(s32) = COPY $vgpr1
%4:_(s1) = G_ICMP intpred(ne), %2, %3
%5:_(<4 x s16>) = G_SELECT %4, %0, %1
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v4s16_vcc_sv
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $vgpr0, $vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; CHECK-LABEL: name: select_v4s16_vcc_sv
; FAST-LABEL: name: select_v4s16_vcc_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
; FAST: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
; FAST: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
; GREEDY-LABEL: name: select_v4s16_vcc_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(<4 x s16>) = COPY $sgpr0_sgpr1
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(<4 x s16>) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(<4 x s16>) = G_SELECT %4, %0, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v4s16_vcc_vs
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $vgpr0, $vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v4s16_vcc_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
; FAST: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
; FAST: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
; GREEDY-LABEL: name: select_v4s16_vcc_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(<4 x s16>) = COPY $sgpr0_sgpr1
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(<4 x s16>) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(<4 x s16>) = G_SELECT %4, %3, %0
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_v4s16_vcc_vv
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_v4s16_vcc_vv
; FAST: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr4_vgpr5
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
; FAST: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
; FAST: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
; GREEDY-LABEL: name: select_v4s16_vcc_vv
; GREEDY: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr4_vgpr5
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(<2 x s16>), [[UV1:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(<2 x s16>), [[UV3:%[0-9]+]]:vgpr(<2 x s16>) = G_UNMERGE_VALUES [[COPY3]](<4 x s16>)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(<2 x s16>) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<4 x s16>) = G_CONCAT_VECTORS [[SELECT]](<2 x s16>), [[SELECT1]](<2 x s16>)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1
%2:_(<4 x s16>) = COPY $vgpr2_vgpr3
%3:_(<4 x s16>) = COPY $vgpr4_vgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(<4 x s16>) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_p1_scc_ss
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_p1_scc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(p1) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:sgpr(p1) = COPY $sgpr4_sgpr5
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:sgpr(p1) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_p1_scc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(p1) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:sgpr(p1) = COPY $sgpr4_sgpr5
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:sgpr(p1) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(p1) = COPY $sgpr2_sgpr3
%3:_(p1) = COPY $sgpr4_sgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(p1) = G_SELECT %4, %2, %3
...
AMDGPU/GlobalISel: Fix scalar G_SELECT for arbitrary pointers 4e85ca9562a588eba491e44bcbf73cb2f419780f missed updating the legal condition type set for pointers with any unrecognized address space.
2020-01-08 05:24:42 +08:00
---
name: select_p999_scc_ss
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5
; FAST-LABEL: name: select_p999_scc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(p999) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:sgpr(p999) = COPY $sgpr4_sgpr5
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:sgpr(p999) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
; GREEDY-LABEL: name: select_p999_scc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(p999) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:sgpr(p999) = COPY $sgpr4_sgpr5
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:sgpr(p999) = G_SELECT [[ZEXT]](s32), [[COPY2]], [[COPY3]]
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(p999) = COPY $sgpr2_sgpr3
%3:_(p999) = COPY $sgpr4_sgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(p999) = G_SELECT %4, %2, %3
...
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_p1_scc_sv
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_p1_scc_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(p1) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p1)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_p1_scc_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(p1) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p1)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(p1) = COPY $sgpr2_sgpr3
%3:_(p1) = COPY $vgpr0_vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(p1) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_p1_scc_vs
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_p1_scc_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(p1) = COPY $sgpr2_sgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p1)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_p1_scc_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(p1) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr0_vgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p1)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(p1) = COPY $sgpr2_sgpr3
%3:_(p1) = COPY $vgpr0_vgpr1
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(p1) = G_SELECT %4, %3, %2
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_p1_scc_vv
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2_sgpr3, $vgpr0_vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_p1_scc_vv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(p1) = COPY $vgpr0_vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr2_vgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; FAST: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p1)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_p1_scc_vv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(p1) = COPY $vgpr0_vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr2_vgpr3
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ICMP:%[0-9]+]]:sgpr(s32) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[ICMP]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p1)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY4]](s1), [[UV1]], [[UV3]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(p1) = COPY $vgpr0_vgpr1
%3:_(p1) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(p1) = G_SELECT %4, %2, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_p1_vcc_ss
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $sgpr2_sgpr3, $vgpr0, $vgpr1
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_p1_vcc_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(p1) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:sgpr(p1) = COPY $sgpr2_sgpr3
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](p1)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY1]](p1)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_p1_vcc_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(p1) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(p1) = COPY $sgpr2_sgpr3
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[COPY3]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](p1)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY1]](p1)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(p1) = COPY $sgpr0_sgpr1
%1:_(p1) = COPY $sgpr2_sgpr3
%2:_(s32) = COPY $vgpr0
%3:_(s32) = COPY $vgpr1
%4:_(s1) = G_ICMP intpred(ne), %2, %3
%5:_(p1) = G_SELECT %4, %0, %1
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_p1_vcc_sv
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $vgpr0, $vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_p1_vcc_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(p1) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr2_vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](p1)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_p1_vcc_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(p1) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr2_vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](p1)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(p1) = COPY $sgpr0_sgpr1
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(p1) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(p1) = G_SELECT %4, %0, %3
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_p1_vcc_vs
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $sgpr0_sgpr1, $vgpr0, $vgpr1, $vgpr2_vgpr3
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_p1_vcc_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(p1) = COPY $sgpr0_sgpr1
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr2_vgpr3
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](p1)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_p1_vcc_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(p1) = COPY $sgpr0_sgpr1
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr2_vgpr3
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY1]](s32), [[COPY2]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY]](p1)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(p1) = COPY $sgpr0_sgpr1
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(p1) = COPY $vgpr2_vgpr3
%4:_(s1) = G_ICMP intpred(ne), %1, %2
%5:_(p1) = G_SELECT %4, %3, %0
...
---
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
name: select_p1_vcc_vv
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
legalized: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; FAST-LABEL: name: select_p1_vcc_vv
; FAST: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(p1) = COPY $vgpr2_vgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr4_vgpr5
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p1)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_p1_vcc_vv
; GREEDY: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(p1) = COPY $vgpr2_vgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(p1) = COPY $vgpr4_vgpr5
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p1)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p1)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(p1) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
RegBankSelect: Handle slightly more complex value mappings Try to use concat_vectors. Also remove unnecessary assert on pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers for AMDGPU. llvm-svn: 354828
2019-02-26 06:24:13 +08:00
%0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1
%2:_(p1) = COPY $vgpr2_vgpr3
%3:_(p1) = COPY $vgpr4_vgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(p1) = G_SELECT %4, %2, %3
...
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
AMDGPU/GlobalISel: Fix scalar G_SELECT for arbitrary pointers 4e85ca9562a588eba491e44bcbf73cb2f419780f missed updating the legal condition type set for pointers with any unrecognized address space.
2020-01-08 05:24:42 +08:00
---
name: select_p999_vcc_vv
legalized: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
; FAST-LABEL: name: select_p999_vcc_vv
; FAST: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(p999) = COPY $vgpr2_vgpr3
; FAST: [[COPY3:%[0-9]+]]:vgpr(p999) = COPY $vgpr4_vgpr5
; FAST: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; FAST: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p999)
; FAST: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p999)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; FAST: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; FAST: [[MV:%[0-9]+]]:vgpr(p999) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
; GREEDY-LABEL: name: select_p999_vcc_vv
; GREEDY: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(p999) = COPY $vgpr2_vgpr3
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(p999) = COPY $vgpr4_vgpr5
; GREEDY: [[ICMP:%[0-9]+]]:vcc(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
; GREEDY: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY2]](p999)
; GREEDY: [[UV2:%[0-9]+]]:vgpr(s32), [[UV3:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[COPY3]](p999)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
; GREEDY: [[SELECT1:%[0-9]+]]:vgpr(s32) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
; GREEDY: [[MV:%[0-9]+]]:vgpr(p999) = G_MERGE_VALUES [[SELECT]](s32), [[SELECT1]](s32)
%0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1
%2:_(p999) = COPY $vgpr2_vgpr3
%3:_(p999) = COPY $vgpr4_vgpr5
%4:_(s1) = G_ICMP intpred(ne), %0, %1
%5:_(p999) = G_SELECT %4, %2, %3
...
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
---
name: select_s32_vgpr_vv
legalized: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
; FAST-LABEL: name: select_s32_vgpr_vv
; FAST: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[TRUNC:%[0-9]+]]:vcc(s1) = G_TRUNC [[COPY]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[TRUNC]](s1), [[COPY1]], [[COPY2]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_s32_vgpr_vv
; GREEDY: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[TRUNC:%[0-9]+]]:vcc(s1) = G_TRUNC [[COPY]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[TRUNC]](s1), [[COPY1]], [[COPY2]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
%0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1
%2:_(s32) = COPY $vgpr2
%3:_(s1) = G_TRUNC %0
%4:_(s32) = G_SELECT %3, %1, %2
...
---
name: select_s32_vgpr_ss
legalized: true
body: |
bb.0:
liveins: $vgpr0, $sgpr0, $sgpr1
; FAST-LABEL: name: select_s32_vgpr_ss
; FAST: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[TRUNC:%[0-9]+]]:vcc(s1) = G_TRUNC [[COPY]](s32)
; FAST: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
; FAST: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[TRUNC]](s1), [[COPY3]], [[COPY4]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_s32_vgpr_ss
; GREEDY: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[TRUNC:%[0-9]+]]:vcc(s1) = G_TRUNC [[COPY]](s32)
; GREEDY: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
; GREEDY: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[TRUNC]](s1), [[COPY3]], [[COPY4]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
%0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $sgpr0
%2:_(s32) = COPY $sgpr1
%3:_(s1) = G_TRUNC %0
%4:_(s32) = G_SELECT %3, %1, %2
...
---
name: select_s32_sgpr_vv
legalized: true
body: |
bb.0:
liveins: $sgpr0, $vgpr0, $vgpr1
; FAST-LABEL: name: select_s32_sgpr_vv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[COPY]](s32)
; FAST: [[COPY3:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY3]](s1), [[COPY1]], [[COPY2]]
; GREEDY-LABEL: name: select_s32_sgpr_vv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[COPY]](s32)
; GREEDY: [[COPY3:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY3]](s1), [[COPY1]], [[COPY2]]
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $vgpr1
%3:_(s1) = G_TRUNC %0
%4:_(s32) = G_SELECT %3, %1, %2
...
---
name: select_s32_sgpr_vs
legalized: true
body: |
bb.0:
liveins: $sgpr0, $vgpr0, $sgpr1
; FAST-LABEL: name: select_s32_sgpr_vs
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[COPY]](s32)
; FAST: [[COPY3:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; FAST: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY3]](s1), [[COPY1]], [[COPY4]]
; GREEDY-LABEL: name: select_s32_sgpr_vs
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[COPY]](s32)
; GREEDY: [[COPY3:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; GREEDY: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY3]](s1), [[COPY1]], [[COPY4]]
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $vgpr0
%2:_(s32) = COPY $sgpr1
%3:_(s1) = G_TRUNC %0
%4:_(s32) = G_SELECT %3, %1, %2
...
---
name: select_s32_sgpr_sv
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr0, $vgpr0
; FAST-LABEL: name: select_s32_sgpr_sv
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[COPY]](s32)
; FAST: [[COPY3:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; FAST: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
; FAST: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY3]](s1), [[COPY4]], [[COPY2]]
; GREEDY-LABEL: name: select_s32_sgpr_sv
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[COPY]](s32)
; GREEDY: [[COPY3:%[0-9]+]]:vcc(s1) = COPY [[TRUNC]](s1)
; GREEDY: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
; GREEDY: [[SELECT:%[0-9]+]]:vgpr(s32) = G_SELECT [[COPY3]](s1), [[COPY4]], [[COPY2]]
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s32) = COPY $vgpr0
%3:_(s1) = G_TRUNC %0
%4:_(s32) = G_SELECT %3, %1, %2
...
# FIXME: greedy unnecessairly uses vcc/vgpr
---
name: select_s32_sgpr_ss
legalized: true
body: |
bb.0:
liveins: $sgpr0, $sgpr1, $sgpr2
; FAST-LABEL: name: select_s32_sgpr_ss
; FAST: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; FAST: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; FAST: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
; FAST: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[COPY]](s32)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; FAST: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; FAST: [[SELECT:%[0-9]+]]:sgpr(s32) = G_SELECT [[ZEXT]](s32), [[COPY1]], [[COPY2]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
; GREEDY-LABEL: name: select_s32_sgpr_ss
; GREEDY: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; GREEDY: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
; GREEDY: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
; GREEDY: [[TRUNC:%[0-9]+]]:sgpr(s1) = G_TRUNC [[COPY]](s32)
AMDGPU/GlobalISel: Replace handling of boolean values This solves selection failures with generated selection patterns, which would fail due to inferring the SGPR reg bank for virtual registers with a set register class instead of VCC bank. Use instruction selection would constrain the virtual register to a specific class, so when the def was selected later the bank no longer was set to VCC. Remove the SCC reg bank. SCC isn't directly addressable, so it requires copying from SCC to an allocatable 32-bit register during selection, so these might as well be treated as 32-bit SGPR values. Now any scalar boolean value that will produce an outupt in SCC should be widened during RegBankSelect to s32. Any s1 value should be a vector boolean during selection. This makes the vcc register bank unambiguous with a normal SGPR during selection. Summary of how this should now work: - G_TRUNC is always a no-op, and never should use a vcc bank result. - SALU boolean operations should be promoted to s32 in RegBankSelect apply mapping - An s1 value means vcc bank at selection. The exception is for legalization artifacts that use s1, which are never VCC. All other contexts should infer the VCC register classes for s1 typed registers. The LLT for the register is now needed to infer the correct register class. Extensions with vcc sources should be legalized to a select of constants during RegBankSelect. - Copy from non-vcc to vcc ensures high bits of the input value are cleared during selection. - SALU boolean inputs should ensure the inputs are 0/1. This includes select, conditional branches, and carry-ins. There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT selection ignores the usual register-bank from register class functions, and can't handle truncates with VCC result banks. I think this is OK, since the artifacts are specially treated anyway. This does require some care to avoid producing cases with vcc. There will also be no 100% reliable way to verify this rule is followed in selection in case of register classes, and violations manifests themselves as invalid copy instructions much later. Standard phi handling also only considers the bank of the result register, and doesn't insert copies to make the source banks match. This doesn't work for vcc, so we have to manually correct phi inputs in this case. We should add a verifier check to make sure there are no phis with mixed vcc and non-vcc register bank inputs. There's also some duplication with the LegalizerHelper, and some code which should live in the helper. I don't see a good way to share special knowledge about what types to use for intermediate operations depending on the bank for example. Using the helper to replace extensions with selects also seems somewhat awkward to me. Another issue is there are some contexts calling getRegBankFromRegClass that apparently don't have the LLT type for the register, but I haven't yet run into a real issue from this. This also introduces new unnecessary instructions in most cases, since we don't yet try to optimize out the zext when the source is known to come from a compare.
2019-11-03 00:30:59 +08:00
; GREEDY: [[ZEXT:%[0-9]+]]:sgpr(s32) = G_ZEXT [[TRUNC]](s1)
; GREEDY: [[SELECT:%[0-9]+]]:sgpr(s32) = G_SELECT [[ZEXT]](s32), [[COPY1]], [[COPY2]]
AMDGPU/GlobalISel: Fix allowing non-boolean conditions for G_SELECT The condition register bank must be scc or vcc so that a copy will be inserted, which will be lowered to a compare. Currently greedy unnecessarily forces using a VCC select. llvm-svn: 364825
2019-07-02 02:13:12 +08:00
%0:_(s32) = COPY $sgpr0
%1:_(s32) = COPY $sgpr1
%2:_(s32) = COPY $sgpr2
%3:_(s1) = G_TRUNC %0
%4:_(s32) = G_SELECT %3, %1, %2
...