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/ARM/vector-promotion.ll

404 lines
17 KiB
LLVM
Raw Normal View History

[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; RUN: opt -codegenprepare -mtriple=thumbv7-apple-ios %s -o - -mattr=+neon -S | FileCheck --check-prefix=IR-BOTH --check-prefix=IR-NORMAL %s
; RUN: opt -codegenprepare -mtriple=thumbv7-apple-ios %s -o - -mattr=+neon -S -stress-cgp-store-extract | FileCheck --check-prefix=IR-BOTH --check-prefix=IR-STRESS %s
; RUN: llc -mtriple=thumbv7-apple-ios %s -o - -mattr=+neon | FileCheck --check-prefix=ASM %s
; IR-BOTH-LABEL: @simpleOneInstructionPromotion
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; IR-BOTH-NEXT: [[VECTOR_OR:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[LOAD]], <i32 undef, i32 1>
; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[VECTOR_OR]], i32 1
; IR-BOTH-NEXT: store i32 [[EXTRACT]], i32* %dest
; IR-BOTH-NEXT: ret
;
; Make sure we got rid of any expensive vmov.32 instructions.
; ASM-LABEL: simpleOneInstructionPromotion:
; ASM: vldr [[LOAD:d[0-9]+]], [r0]
; ASM-NEXT: vorr.i32 [[LOAD]], #0x1
; ASM-NEXT: vst1.32 {[[LOAD]][1]}, [r1:32]
; ASM-NEXT: bx
define void @simpleOneInstructionPromotion(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 1
%out = or i32 %extract, 1
store i32 %out, i32* %dest, align 4
ret void
}
; IR-BOTH-LABEL: @unsupportedInstructionForPromotion
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 0
; IR-BOTH-NEXT: [[CMP:%[a-zA-Z_0-9-]+]] = icmp eq i32 [[EXTRACT]], %in2
; IR-BOTH-NEXT: store i1 [[CMP]], i1* %dest
; IR-BOTH-NEXT: ret
;
; ASM-LABEL: unsupportedInstructionForPromotion:
; ASM: vldr [[LOAD:d[0-9]+]], [r0]
; ASM: vmov.32 {{r[0-9]+}}, [[LOAD]]
; ASM: bx
define void @unsupportedInstructionForPromotion(<2 x i32>* %addr1, i32 %in2, i1* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 0
%out = icmp eq i32 %extract, %in2
store i1 %out, i1* %dest, align 4
ret void
}
; IR-BOTH-LABEL: @unsupportedChainInDifferentBBs
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 0
; IR-BOTH-NEXT: br i1 %bool, label %bb2, label %end
; BB2
; IR-BOTH: [[OR:%[a-zA-Z_0-9-]+]] = or i32 [[EXTRACT]], 1
; IR-BOTH-NEXT: store i32 [[OR]], i32* %dest, align 4
; IR-BOTH: ret
;
; ASM-LABEL: unsupportedChainInDifferentBBs:
Don't conditionalize Neon instructions, even in IT blocks. This has been deprecated since ARMARM v7-AR, release C.b, published back in 2012. This also removes test/CodeGen/Thumb2/ifcvt-neon.ll that originally was introduced to check that conditionalization of Neon instructions did happen when generating Thumb2. However, the test had evolved and was no longer testing that. Rather than trying to adapt that test, this commit introduces test/CodeGen/Thumb2/ifcvt-neon-deprecated.mir, since we can now use the MIR framework to write nicer/more maintainable tests. llvm-svn: 305998
2017-06-22 20:11:38 +08:00
; ASM: vldr [[LOAD:d[0-9]+]], [r0]
; ASM: vmov.32 {{r[0-9]+}}, [[LOAD]]
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; ASM: bx
define void @unsupportedChainInDifferentBBs(<2 x i32>* %addr1, i32* %dest, i1 %bool) {
bb1:
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 0
br i1 %bool, label %bb2, label %end
bb2:
%out = or i32 %extract, 1
store i32 %out, i32* %dest, align 4
br label %end
end:
ret void
}
; IR-LABEL: @chainOfInstructionsToPromote
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; IR-BOTH-NEXT: [[VECTOR_OR1:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[LOAD]], <i32 1, i32 undef>
; IR-BOTH-NEXT: [[VECTOR_OR2:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR1]], <i32 1, i32 undef>
; IR-BOTH-NEXT: [[VECTOR_OR3:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR2]], <i32 1, i32 undef>
; IR-BOTH-NEXT: [[VECTOR_OR4:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR3]], <i32 1, i32 undef>
; IR-BOTH-NEXT: [[VECTOR_OR5:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR4]], <i32 1, i32 undef>
; IR-BOTH-NEXT: [[VECTOR_OR6:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR5]], <i32 1, i32 undef>
; IR-BOTH-NEXT: [[VECTOR_OR7:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[VECTOR_OR6]], <i32 1, i32 undef>
; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[VECTOR_OR7]], i32 0
; IR-BOTH-NEXT: store i32 [[EXTRACT]], i32* %dest
; IR-BOTH-NEXT: ret
;
; ASM-LABEL: chainOfInstructionsToPromote:
; ASM: vldr [[LOAD:d[0-9]+]], [r0]
; ASM-NOT: vmov.32 {{r[0-9]+}}, [[LOAD]]
; ASM: bx
define void @chainOfInstructionsToPromote(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 0
%out1 = or i32 %extract, 1
%out2 = or i32 %out1, 1
%out3 = or i32 %out2, 1
%out4 = or i32 %out3, 1
%out5 = or i32 %out4, 1
%out6 = or i32 %out5, 1
%out7 = or i32 %out6, 1
store i32 %out7, i32* %dest, align 4
ret void
}
; IR-BOTH-LABEL: @unsupportedMultiUses
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 1
; IR-BOTH-NEXT: [[OR:%[a-zA-Z_0-9-]+]] = or i32 [[EXTRACT]], 1
; IR-BOTH-NEXT: store i32 [[OR]], i32* %dest
; IR-BOTH-NEXT: ret i32 [[OR]]
;
; ASM-LABEL: unsupportedMultiUses:
; ASM: vldr [[LOAD:d[0-9]+]], [r0]
; ASM: vmov.32 {{r[0-9]+}}, [[LOAD]]
; ASM: bx
define i32 @unsupportedMultiUses(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 1
%out = or i32 %extract, 1
store i32 %out, i32* %dest, align 4
ret i32 %out
}
; Check that we promote we a splat constant when this is a division.
; The NORMAL mode does not promote anything as divisions are not legal.
; IR-BOTH-LABEL: @udivCase
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = udiv i32 [[EXTRACT]], 7
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = udiv <2 x i32> [[LOAD]], <i32 7, i32 7>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store i32 [[RES]], i32* %dest
; IR-BOTH-NEXT: ret
define void @udivCase(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 1
%out = udiv i32 %extract, 7
store i32 %out, i32* %dest, align 4
ret void
}
; IR-BOTH-LABEL: @uremCase
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = urem i32 [[EXTRACT]], 7
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = urem <2 x i32> [[LOAD]], <i32 7, i32 7>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store i32 [[RES]], i32* %dest
; IR-BOTH-NEXT: ret
define void @uremCase(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 1
%out = urem i32 %extract, 7
store i32 %out, i32* %dest, align 4
ret void
}
; IR-BOTH-LABEL: @sdivCase
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = sdiv i32 [[EXTRACT]], 7
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = sdiv <2 x i32> [[LOAD]], <i32 7, i32 7>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store i32 [[RES]], i32* %dest
; IR-BOTH-NEXT: ret
define void @sdivCase(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 1
%out = sdiv i32 %extract, 7
store i32 %out, i32* %dest, align 4
ret void
}
; IR-BOTH-LABEL: @sremCase
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = srem i32 [[EXTRACT]], 7
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = srem <2 x i32> [[LOAD]], <i32 7, i32 7>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store i32 [[RES]], i32* %dest
; IR-BOTH-NEXT: ret
define void @sremCase(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 1
%out = srem i32 %extract, 7
store i32 %out, i32* %dest, align 4
ret void
}
; IR-BOTH-LABEL: @fdivCase
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x float>, <2 x float>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = fdiv float [[EXTRACT]], 7.0
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = fdiv <2 x float> [[LOAD]], <float 7.000000e+00, float 7.000000e+00>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store float [[RES]], float* %dest
; IR-BOTH-NEXT: ret
define void @fdivCase(<2 x float>* %addr1, float* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x float>, <2 x float>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x float> %in1, i32 1
%out = fdiv float %extract, 7.0
store float %out, float* %dest, align 4
ret void
}
; IR-BOTH-LABEL: @fremCase
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x float>, <2 x float>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = frem float [[EXTRACT]], 7.0
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = frem <2 x float> [[LOAD]], <float 7.000000e+00, float 7.000000e+00>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store float [[RES]], float* %dest
; IR-BOTH-NEXT: ret
define void @fremCase(<2 x float>* %addr1, float* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x float>, <2 x float>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x float> %in1, i32 1
%out = frem float %extract, 7.0
store float %out, float* %dest, align 4
ret void
}
; Check that we do not promote when we may introduce undefined behavior
; like division by zero.
; IR-BOTH-LABEL: @undefDivCase
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 1
; IR-BOTH-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = udiv i32 7, [[EXTRACT]]
; IR-BOTH-NEXT: store i32 [[RES]], i32* %dest
; IR-BOTH-NEXT: ret
define void @undefDivCase(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 1
%out = udiv i32 7, %extract
store i32 %out, i32* %dest, align 4
ret void
}
; Check that we do not promote when we may introduce undefined behavior
; like division by zero.
; IR-BOTH-LABEL: @undefRemCase
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 1
; IR-BOTH-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = srem i32 7, [[EXTRACT]]
; IR-BOTH-NEXT: store i32 [[RES]], i32* %dest
; IR-BOTH-NEXT: ret
define void @undefRemCase(<2 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 1
%out = srem i32 7, %extract
store i32 %out, i32* %dest, align 4
ret void
}
; Check that we use an undef mask for undefined behavior if the fast-math
; flag is set.
; IR-BOTH-LABEL: @undefConstantFRemCaseWithFastMath
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x float>, <2 x float>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = frem nnan float [[EXTRACT]], 7.0
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = frem nnan <2 x float> [[LOAD]], <float undef, float 7.000000e+00>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store float [[RES]], float* %dest
; IR-BOTH-NEXT: ret
define void @undefConstantFRemCaseWithFastMath(<2 x float>* %addr1, float* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x float>, <2 x float>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x float> %in1, i32 1
%out = frem nnan float %extract, 7.0
store float %out, float* %dest, align 4
ret void
}
; Check that we use an undef mask for undefined behavior if the fast-math
; flag is set.
; IR-BOTH-LABEL: @undefVectorFRemCaseWithFastMath
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x float>, <2 x float>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = frem nnan float 7.000000e+00, [[EXTRACT]]
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = frem nnan <2 x float> <float undef, float 7.000000e+00>, [[LOAD]]
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store float [[RES]], float* %dest
; IR-BOTH-NEXT: ret
define void @undefVectorFRemCaseWithFastMath(<2 x float>* %addr1, float* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x float>, <2 x float>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x float> %in1, i32 1
%out = frem nnan float 7.0, %extract
store float %out, float* %dest, align 4
ret void
}
; Check that we are able to promote floating point value.
; This requires the STRESS mode, as floating point value are
; not promote on armv7.
; IR-BOTH-LABEL: @simpleOneInstructionPromotionFloat
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x float>, <2 x float>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = fadd float [[EXTRACT]], 1.0
; Vector version:
; IR-STRESS-NEXT: [[DIV:%[a-zA-Z_0-9-]+]] = fadd <2 x float> [[LOAD]], <float undef, float 1.000000e+00>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x float> [[DIV]], i32 1
;
; IR-BOTH-NEXT: store float [[RES]], float* %dest
; IR-BOTH-NEXT: ret
define void @simpleOneInstructionPromotionFloat(<2 x float>* %addr1, float* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x float>, <2 x float>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x float> %in1, i32 1
%out = fadd float %extract, 1.0
store float %out, float* %dest, align 4
ret void
}
; Check that we correctly use a splat constant when we cannot
; determine at compile time the index of the extract.
; This requires the STRESS modes, as variable index are expensive
; to lower.
; IR-BOTH-LABEL: @simpleOneInstructionPromotionVariableIdx
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <2 x i32>, <2 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[LOAD]], i32 %idx
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = or i32 [[EXTRACT]], 1
; Vector version:
; IR-STRESS-NEXT: [[OR:%[a-zA-Z_0-9-]+]] = or <2 x i32> [[LOAD]], <i32 1, i32 1>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <2 x i32> [[OR]], i32 %idx
;
; IR-BOTH-NEXT: store i32 [[RES]], i32* %dest
; IR-BOTH-NEXT: ret
define void @simpleOneInstructionPromotionVariableIdx(<2 x i32>* %addr1, i32* %dest, i32 %idx) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <2 x i32>, <2 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <2 x i32> %in1, i32 %idx
%out = or i32 %extract, 1
store i32 %out, i32* %dest, align 4
ret void
}
; Check a vector with more than 2 elements.
; This requires the STRESS mode because currently 'or v8i8' is not marked
; as legal or custom, althought the actual assembly is better if we were
; promoting it.
; IR-BOTH-LABEL: @simpleOneInstructionPromotion8x8
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <8 x i8>, <8 x i8>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; Scalar version:
; IR-NORMAL-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <8 x i8> [[LOAD]], i32 1
; IR-NORMAL-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = or i8 [[EXTRACT]], 1
; Vector version:
; IR-STRESS-NEXT: [[OR:%[a-zA-Z_0-9-]+]] = or <8 x i8> [[LOAD]], <i8 undef, i8 1, i8 undef, i8 undef, i8 undef, i8 undef, i8 undef, i8 undef>
; IR-STRESS-NEXT: [[RES:%[a-zA-Z_0-9-]+]] = extractelement <8 x i8> [[OR]], i32 1
;
; IR-BOTH-NEXT: store i8 [[RES]], i8* %dest
; IR-BOTH-NEXT: ret
define void @simpleOneInstructionPromotion8x8(<8 x i8>* %addr1, i8* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <8 x i8>, <8 x i8>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <8 x i8> %in1, i32 1
%out = or i8 %extract, 1
store i8 %out, i8* %dest, align 4
ret void
}
; Check that we optimized the sequence correctly when it can be
; lowered on a Q register.
; IR-BOTH-LABEL: @simpleOneInstructionPromotion
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; IR-BOTH: [[LOAD:%[a-zA-Z_0-9-]+]] = load <4 x i32>, <4 x i32>* %addr1
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
; IR-BOTH-NEXT: [[VECTOR_OR:%[a-zA-Z_0-9-]+]] = or <4 x i32> [[LOAD]], <i32 undef, i32 1, i32 undef, i32 undef>
; IR-BOTH-NEXT: [[EXTRACT:%[a-zA-Z_0-9-]+]] = extractelement <4 x i32> [[VECTOR_OR]], i32 1
; IR-BOTH-NEXT: store i32 [[EXTRACT]], i32* %dest
; IR-BOTH-NEXT: ret
;
; Make sure we got rid of any expensive vmov.32 instructions.
; ASM-LABEL: simpleOneInstructionPromotion4x32:
; ASM: vld1.64 {[[LOAD:d[0-9]+]], d{{[0-9]+}}}, [r0]
; The Q register used here must be [[LOAD]] / 2, but we cannot express that.
; ASM-NEXT: vorr.i32 q{{[[0-9]+}}, #0x1
; ASM-NEXT: vst1.32 {[[LOAD]][1]}, [r1]
; ASM-NEXT: bx
define void @simpleOneInstructionPromotion4x32(<4 x i32>* %addr1, i32* %dest) {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%in1 = load <4 x i32>, <4 x i32>* %addr1, align 8
[CodeGenPrepare] Move extractelement close to store if they can be combined. This patch adds an optimization in CodeGenPrepare to move an extractelement right before a store when the target can combine them. The optimization may promote any scalar operations to vector operations in the way to make that possible. ** Context ** Some targets use different register files for both vector and scalar operations. This means that transitioning from one domain to another may incur copy from one register file to another. These copies are not coalescable and may be expensive. For example, according to the scheduling model, on cortex-A8 a vector to GPR move is 20 cycles. ** Motivating Example ** Let us consider an example: define void @foo(<2 x i32>* %addr1, i32* %dest) { %in1 = load <2 x i32>* %addr1, align 8 %extract = extractelement <2 x i32> %in1, i32 1 %out = or i32 %extract, 1 store i32 %out, i32* %dest, align 4 ret void } As it is, this IR generates the following assembly on armv7: vldr d16, [r0] @vector load vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles orr r0, r0, #1 @ scalar bitwise or str r0, [r1] @ scalar store bx lr Whereas we could generate much faster code: vldr d16, [r0] @ vector load vorr.i32 d16, #0x1 @ vector bitwise or vst1.32 {d16[1]}, [r1:32] @ vector extract + store bx lr Half of the computation made in the vector is useless, but this allows to get rid of the expensive cross-register-file copy. ** Proposed Solution ** To avoid this cross-register-copy penalty, we promote the scalar operations to vector operations. The penalty will be removed if we manage to promote the whole chain of computation in the vector domain. Currently, we do that only when the chain of computation ends by a store and the target is able to combine an extract with a store. Stores are the most likely candidates, because other instructions produce values that would need to be promoted and so, extracted as some point[1]. Moreover, this is customary that targets feature stores that perform a vector extract (see AArch64 and X86 for instance). The proposed implementation relies on the TargetTransformInfo to decide whether or not it is beneficial to promote a chain of computation in the vector domain. Unfortunately, this interface is rather inaccurate for this level of details and although this optimization may be beneficial for X86 and AArch64, the inaccuracy will lead to the optimization being too aggressive. Basically in TargetTransformInfo, everything that is legal has a cost of 1, whereas, even if a vector type is legal, usually a vector operation is slightly more expensive than its scalar counterpart. That will lead to too many promotions that may not be counter balanced by the saving of the cross-register-file copy. For instance, on AArch64 this penalty is just 4 cycles. For now, the optimization is just enabled for ARM prior than v8, since those processors have a larger penalty on cross-register-file copies, and the scope is limited to basic blocks. Because of these two factors, we limit the effects of the inaccuracy. Indeed, I did not want to build up a fancy cost model with block frequency and everything on top of that. [1] We can imagine targets that can combine an extractelement with other instructions than just stores. If we want to go into that direction, the current interfaces must be augmented and, moreover, I think this becomes a global isel problem. Differential Revision: http://reviews.llvm.org/D5921 <rdar://problem/14170854> llvm-svn: 220978
2014-11-01 01:52:53 +08:00
%extract = extractelement <4 x i32> %in1, i32 1
%out = or i32 %extract, 1
store i32 %out, i32* %dest, align 1
ret void
}