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/fp16-instructions.ll

748 lines
24 KiB
LLVM
Raw Normal View History

[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
; SOFT:
[ARM] FullFP16 LowerReturn Fix Commit r323512 introduced an optimisation in LowerReturn for half-precision return values. A missing check caused a crash when the return value is "undef" (i.e. a node that has no operands). Differential Revision: https://reviews.llvm.org/D42743 llvm-svn: 323968
2018-02-01 21:48:40 +08:00
; RUN: llc < %s -mtriple=arm-none-eabi -float-abi=soft | FileCheck %s --check-prefixes=CHECK,CHECK-SOFT
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; RUN: llc < %s -mtriple=thumb-none-eabi -float-abi=soft | FileCheck %s --check-prefixes=CHECK,CHECK-SOFT
[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
; SOFTFP:
[ARM] FullFP16 LowerReturn Fix Commit r323512 introduced an optimisation in LowerReturn for half-precision return values. A missing check caused a crash when the return value is "undef" (i.e. a node that has no operands). Differential Revision: https://reviews.llvm.org/D42743 llvm-svn: 323968
2018-02-01 21:48:40 +08:00
; RUN: llc < %s -mtriple=arm-none-eabi -mattr=+vfp3 | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-VFP3
; RUN: llc < %s -mtriple=arm-none-eabi -mattr=+vfp4 | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-FP16
; RUN: llc < %s -mtriple=arm-none-eabi -mattr=+fullfp16 | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-FULLFP16
[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; RUN: llc < %s -mtriple=thumbv7-none-eabi -mattr=+vfp3 | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-VFP3
; RUN: llc < %s -mtriple=thumbv7-none-eabi -mattr=+vfp4 | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-FP16
; RUN: llc < %s -mtriple=thumbv7-none-eabi -mattr=+fullfp16 | FileCheck %s --check-prefixes=CHECK,CHECK-SOFTFP-FULLFP16
[ARM] f16 stack spill/reloads This adds support for handling f16 stack spills/reloads. Differential Revision: https://reviews.llvm.org/D43280 llvm-svn: 325130
2018-02-14 23:09:09 +08:00
; Test fast-isel
; RUN: llc < %s -mtriple=arm-none-eabi -mattr=+fullfp16 -O0 | FileCheck %s --check-prefixes=CHECK-SPILL-RELOAD
; RUN: llc < %s -mtriple=thumbv7-none-eabi -mattr=+fullfp16 -O0 | FileCheck %s --check-prefixes=CHECK-SPILL-RELOAD
[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
; HARD:
[ARM] FullFP16 LowerReturn Fix Commit r323512 introduced an optimisation in LowerReturn for half-precision return values. A missing check caused a crash when the return value is "undef" (i.e. a node that has no operands). Differential Revision: https://reviews.llvm.org/D42743 llvm-svn: 323968
2018-02-01 21:48:40 +08:00
; RUN: llc < %s -mtriple=arm-none-eabihf -mattr=+vfp3 | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-VFP3
; RUN: llc < %s -mtriple=arm-none-eabihf -mattr=+vfp4 | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-FP16
; RUN: llc < %s -mtriple=arm-none-eabihf -mattr=+fullfp16 | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-FULLFP16
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; RUN: llc < %s -mtriple=thumbv7-none-eabihf -mattr=+vfp3 | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-VFP3
; RUN: llc < %s -mtriple=thumbv7-none-eabihf -mattr=+vfp4 | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-FP16
; RUN: llc < %s -mtriple=thumbv7-none-eabihf -mattr=+fullfp16 | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-FULLFP16
; FP-CONTRACT=FAST
; RUN: llc < %s -mtriple=arm-none-eabihf -mattr=+fullfp16 -fp-contract=fast | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-FULLFP16-FAST
; RUN: llc < %s -mtriple=thumbv7-none-eabihf -mattr=+fullfp16 -fp-contract=fast | FileCheck %s --check-prefixes=CHECK,CHECK-HARDFP-FULLFP16-FAST
[ARM] FullFP16 LowerReturn Fix Commit r323512 introduced an optimisation in LowerReturn for half-precision return values. A missing check caused a crash when the return value is "undef" (i.e. a node that has no operands). Differential Revision: https://reviews.llvm.org/D42743 llvm-svn: 323968
2018-02-01 21:48:40 +08:00
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
define float @RetValBug(float %A.coerce) {
[ARM] FullFP16 LowerReturn Fix Commit r323512 introduced an optimisation in LowerReturn for half-precision return values. A missing check caused a crash when the return value is "undef" (i.e. a node that has no operands). Differential Revision: https://reviews.llvm.org/D42743 llvm-svn: 323968
2018-02-01 21:48:40 +08:00
entry:
ret float undef
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; Check thatLowerReturn can handle undef nodes (i.e. nodes which do not have
; any operands) when FullFP16 is enabled.
[ARM] FullFP16 LowerReturn Fix Commit r323512 introduced an optimisation in LowerReturn for half-precision return values. A missing check caused a crash when the return value is "undef" (i.e. a node that has no operands). Differential Revision: https://reviews.llvm.org/D42743 llvm-svn: 323968
2018-02-01 21:48:40 +08:00
;
; CHECK-LABEL: RetValBug:
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; CHECK-HARDFP-FULLFP16: {{.*}} lr
[ARM] FullFP16 LowerReturn Fix Commit r323512 introduced an optimisation in LowerReturn for half-precision return values. A missing check caused a crash when the return value is "undef" (i.e. a node that has no operands). Differential Revision: https://reviews.llvm.org/D42743 llvm-svn: 323968
2018-02-01 21:48:40 +08:00
}
[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; 1. VABS: TODO
; 2. VADD
define float @Add(float %a.coerce, float %b.coerce) {
[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%add = fadd half %1, %3
%4 = bitcast half %add to i16
%tmp4.0.insert.ext = zext i16 %4 to i32
%5 = bitcast i32 %tmp4.0.insert.ext to float
ret float %5
[ARM] FullFP16 LowerReturn Fix Commit r323512 introduced an optimisation in LowerReturn for half-precision return values. A missing check caused a crash when the return value is "undef" (i.e. a node that has no operands). Differential Revision: https://reviews.llvm.org/D42743 llvm-svn: 323968
2018-02-01 21:48:40 +08:00
; CHECK-LABEL: Add:
[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
; CHECK-SOFT: bl __aeabi_h2f
; CHECK-SOFT: bl __aeabi_h2f
; CHECK-SOFT: bl __aeabi_fadd
; CHECK-SOFT: bl __aeabi_f2h
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: vadd.f32
; CHECK-SOFTFP-VFP3: bl __aeabi_f2h
; CHECK-SOFTFP-FP16: vmov [[S2:s[0-9]]], r1
; CHECK-SOFTFP-FP16: vmov [[S0:s[0-9]]], r0
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 [[S2]], [[S2]]
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 [[S0]], [[S0]]
; CHECK-SOFTFP-FP16: vadd.f32 [[S0]], [[S0]], [[S2]]
; CHECK-SOFTFP-FP16: vcvtb.f16.f32 [[S0]], [[S0]]
; CHECK-SOFTFP-FP16: vmov r0, s0
[ARM] Armv8.2-A FP16 code generation (part 2/3) Half-precision arguments and return values are passed as if it were an int or float for ARM. This results in truncates and bitcasts to/from i16 and f16 values, which are legalized very early to stack stores/loads. When FullFP16 is enabled, we want to avoid codegen for these bitcasts as it is unnecessary and inefficient. Differential Revision: https://reviews.llvm.org/D42580 llvm-svn: 323861
2018-01-31 18:18:29 +08:00
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S0:s[0-9]]], r1
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S2:s[0-9]]], r0
; CHECK-SOFTFP-FULLFP16: vadd.f16 [[S0]], [[S2]], [[S0]]
; CHECK-SOFTFP-FULLFP16-NEXT: vmov.f16 r0, s0
[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
; CHECK-HARDFP-VFP3: vmov r{{.}}, s0
; CHECK-HARDFP-VFP3: vmov{{.*}}, s1
; CHECK-HARDFP-VFP3: bl __aeabi_h2f
; CHECK-HARDFP-VFP3: bl __aeabi_h2f
; CHECK-HARDFP-VFP3: vadd.f32
; CHECK-HARDFP-VFP3: bl __aeabi_f2h
; CHECK-HARDFP-VFP3: vmov s0, r0
; CHECK-HARDFP-FP16: vcvtb.f32.f16 [[S2:s[0-9]]], s1
; CHECK-HARDFP-FP16: vcvtb.f32.f16 [[S0:s[0-9]]], s0
; CHECK-HARDFP-FP16: vadd.f32 [[S0]], [[S0]], [[S2]]
; CHECK-HARDFP-FP16: vcvtb.f16.f32 [[S0]], [[S0]]
; CHECK-HARDFP-FULLFP16: vadd.f16 s0, s0, s1
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
}
; 3. VCMP
define zeroext i1 @VCMP(float %F.coerce, float %G.coerce) {
entry:
%0 = bitcast float %F.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %G.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%cmp = fcmp ogt half %1, %3
ret i1 %cmp
; CHECK-LABEL: VCMP:
; CHECK-SOFT: bl __aeabi_fcmpgt
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: vcmpe.f32 s{{.}}, s{{.}}
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 s{{.}}, s{{.}}
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 s{{.}}, s{{.}}
; CHECK-SOFTFP-FP16: vcmpe.f32 s{{.}}, s{{.}}
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S2:s[0-9]]], r0
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S0:s[0-9]]], r1
; CHECK-SOFTFP-FULLFP16: vcmpe.f16 [[S2]], [[S0]]
; CHECK-SOFTFP-FULLFP16-NOT: vmov.f16 s{{.}}, r0
; CHECK-SOFTFP-FULLFP16-NOT: vmov.f16 s{{.}}, r1
; CHECK-HARDFP-FULLFP16: vcmpe.f16 s0, s1
}
; 4. VCMPE
; FIXME: enable when constant pool is fixed
;
;define i32 @VCMPE_IMM(float %F.coerce) {
;entry:
; %0 = bitcast float %F.coerce to i32
; %tmp.0.extract.trunc = trunc i32 %0 to i16
; %1 = bitcast i16 %tmp.0.extract.trunc to half
; %tmp = fcmp olt half %1, 1.000000e+00
; %tmp1 = zext i1 %tmp to i32
; ret i32 %tmp1
;}
define i32 @VCMPE(float %F.coerce, float %G.coerce) {
entry:
%0 = bitcast float %F.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %G.coerce to i32
%tmp.1.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp.1.extract.trunc to half
%tmp = fcmp olt half %1, %3
%tmp1 = zext i1 %tmp to i32
ret i32 %tmp1
; CHECK-LABEL: VCMPE:
}
; 5. VCVT (between floating-point and fixed-point)
; Only assembly/disassembly support
; 6. VCVT (between floating-point and integer, both directions)
define i32 @fptosi(i32 %A.coerce) {
entry:
%tmp.0.extract.trunc = trunc i32 %A.coerce to i16
%0 = bitcast i16 %tmp.0.extract.trunc to half
%conv = fptosi half %0 to i32
ret i32 %conv
; CHECK-LABEL: fptosi:
; CHECK-HARDFP-FULLFP16: vmov.f16 s0, r0
; CHECK-HARDFP-FULLFP16-NEXT: vcvt.s32.f16 s0, s0
; CHECK-HARDFP-FULLFP16-NEXT: vmov r0, s0
}
define i32 @fptoui(i32 %A.coerce) {
entry:
%tmp.0.extract.trunc = trunc i32 %A.coerce to i16
%0 = bitcast i16 %tmp.0.extract.trunc to half
%conv = fptoui half %0 to i32
ret i32 %conv
; CHECK-HARDFP-FULLFP16: vcvt.u32.f16 s0, s0
; CHECK-HARDFP-FULLFP16-NEXT: vmov r0, s0
}
define float @UintToH(i32 %a, i32 %b) {
entry:
%0 = uitofp i32 %a to half
%1 = bitcast half %0 to i16
%tmp0.insert.ext = zext i16 %1 to i32
%2 = bitcast i32 %tmp0.insert.ext to float
ret float %2
; CHECK-LABEL: UintToH:
; CHECK-HARDFP-FULLFP16: vmov s0, r0
; CHECK-HARDFP-FULLFP16-NEXT: vcvt.f16.u32 s0, s0
}
define float @SintToH(i32 %a, i32 %b) {
entry:
%0 = sitofp i32 %a to half
%1 = bitcast half %0 to i16
%tmp0.insert.ext = zext i16 %1 to i32
%2 = bitcast i32 %tmp0.insert.ext to float
ret float %2
; CHECK-LABEL: SintToH:
; CHECK-HARDFP-FULLFP16: vmov s0, r0
; CHECK-HARDFP-FULLFP16-NEXT: vcvt.f16.s32 s0, s0
}
[ARM] f16 conversions This is a follow up of r324321, adding f16 <-> f32 and f16 <-> f64 conversion match patterns. Differential Revision: https://reviews.llvm.org/D42954 llvm-svn: 324360
2018-02-07 00:28:43 +08:00
define i32 @f2h(float %f) {
entry:
%conv = fptrunc float %f to half
%0 = bitcast half %conv to i16
%tmp.0.insert.ext = zext i16 %0 to i32
ret i32 %tmp.0.insert.ext
; CHECK-LABEL: f2h:
; CHECK-HARDFP-FULLFP16: vcvtb.f16.f32 s0, s0
}
define float @h2f(i32 %h.coerce) {
entry:
%tmp.0.extract.trunc = trunc i32 %h.coerce to i16
%0 = bitcast i16 %tmp.0.extract.trunc to half
%conv = fpext half %0 to float
ret float %conv
; CHECK-LABEL: h2f:
; CHECK-HARDFP-FULLFP16: vcvtb.f32.f16 s0, s0
}
define double @h2d(i32 %h.coerce) {
entry:
%tmp.0.extract.trunc = trunc i32 %h.coerce to i16
%0 = bitcast i16 %tmp.0.extract.trunc to half
%conv = fpext half %0 to double
ret double %conv
; CHECK-LABEL: h2d:
; CHECK-HARDFP-FULLFP16: vcvtb.f64.f16 d{{.*}}, s{{.}}
}
define i32 @d2h(double %d) {
entry:
%conv = fptrunc double %d to half
%0 = bitcast half %conv to i16
%tmp.0.insert.ext = zext i16 %0 to i32
ret i32 %tmp.0.insert.ext
; CHECK-LABEL: d2h:
; CHECK-HARDFP-FULLFP16: vcvtb.f16.f64 s0, d{{.*}}
}
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; TODO:
; 7. VCVTA
; 8. VCVTM
; 9. VCVTN
; 10. VCVTP
; 11. VCVTR
; 12. VDIV
define float @Div(float %a.coerce, float %b.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%add = fdiv half %1, %3
%4 = bitcast half %add to i16
%tmp4.0.insert.ext = zext i16 %4 to i32
%5 = bitcast i32 %tmp4.0.insert.ext to float
ret float %5
; CHECK-LABEL: Div:
; CHECK-SOFT: bl __aeabi_h2f
; CHECK-SOFT: bl __aeabi_h2f
; CHECK-SOFT: bl __aeabi_fdiv
; CHECK-SOFT: bl __aeabi_f2h
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: vdiv.f32
; CHECK-SOFTFP-VFP3: bl __aeabi_f2h
; CHECK-SOFTFP-FP16: vmov [[S2:s[0-9]]], r1
; CHECK-SOFTFP-FP16: vmov [[S0:s[0-9]]], r0
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 [[S2]], [[S2]]
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 [[S0]], [[S0]]
; CHECK-SOFTFP-FP16: vdiv.f32 [[S0]], [[S0]], [[S2]]
; CHECK-SOFTFP-FP16: vcvtb.f16.f32 [[S0]], [[S0]]
; CHECK-SOFTFP-FP16: vmov r0, s0
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S0:s[0-9]]], r1
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S2:s[0-9]]], r0
; CHECK-SOFTFP-FULLFP16: vdiv.f16 [[S0]], [[S2]], [[S0]]
; CHECK-SOFTFP-FULLFP16-NEXT: vmov.f16 r0, s0
; CHECK-HARDFP-VFP3: vmov r{{.}}, s0
; CHECK-HARDFP-VFP3: vmov{{.*}}, s1
; CHECK-HARDFP-VFP3: bl __aeabi_h2f
; CHECK-HARDFP-VFP3: bl __aeabi_h2f
; CHECK-HARDFP-VFP3: vdiv.f32
; CHECK-HARDFP-VFP3: bl __aeabi_f2h
; CHECK-HARDFP-VFP3: vmov s0, r0
; CHECK-HARDFP-FP16: vcvtb.f32.f16 [[S2:s[0-9]]], s1
; CHECK-HARDFP-FP16: vcvtb.f32.f16 [[S0:s[0-9]]], s0
; CHECK-HARDFP-FP16: vdiv.f32 [[S0]], [[S0]], [[S2]]
; CHECK-HARDFP-FP16: vcvtb.f16.f32 [[S0]], [[S0]]
; CHECK-HARDFP-FULLFP16: vdiv.f16 s0, s0, s1
}
; 13. VFMA
define float @VFMA(float %a.coerce, float %b.coerce, float %c.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%4 = bitcast float %c.coerce to i32
%tmp2.0.extract.trunc = trunc i32 %4 to i16
%5 = bitcast i16 %tmp2.0.extract.trunc to half
%mul = fmul half %1, %3
%add = fadd half %mul, %5
%6 = bitcast half %add to i16
%tmp4.0.insert.ext = zext i16 %6 to i32
%7 = bitcast i32 %tmp4.0.insert.ext to float
ret float %7
; CHECK-LABEL: VFMA:
; CHECK-HARDFP-FULLFP16-FAST: vfma.f16 s2, s0, s1
; CHECK-HARDFP-FULLFP16-FAST-NEXT: vmov.f32 s0, s2
}
; 14. VFMS
define float @VFMS(float %a.coerce, float %b.coerce, float %c.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%4 = bitcast float %c.coerce to i32
%tmp2.0.extract.trunc = trunc i32 %4 to i16
%5 = bitcast i16 %tmp2.0.extract.trunc to half
%mul = fmul half %1, %3
%sub = fsub half %5, %mul
%6 = bitcast half %sub to i16
%tmp4.0.insert.ext = zext i16 %6 to i32
%7 = bitcast i32 %tmp4.0.insert.ext to float
ret float %7
; CHECK-LABEL: VFMS:
; CHECK-HARDFP-FULLFP16-FAST: vfms.f16 s2, s0, s1
; CHECK-HARDFP-FULLFP16-FAST-NEXT: vmov.f32 s0, s2
}
; 15. VFNMA
define float @VFNMA(float %a.coerce, float %b.coerce, float %c.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%4 = bitcast float %c.coerce to i32
%tmp2.0.extract.trunc = trunc i32 %4 to i16
%5 = bitcast i16 %tmp2.0.extract.trunc to half
%mul = fmul half %1, %3
%sub = fsub half -0.0, %mul
%sub2 = fsub half %sub, %5
%6 = bitcast half %sub2 to i16
%tmp4.0.insert.ext = zext i16 %6 to i32
%7 = bitcast i32 %tmp4.0.insert.ext to float
ret float %7
; CHECK-LABEL: VFNMA:
; CHECK-HARDFP-FULLFP16-FAST: vfnma.f16 s2, s0, s1
; CHECK-HARDFP-FULLFP16-FAST-NEXT: vmov.f32 s0, s2
}
; 16. VFNMS
define float @VFNMS(float %a.coerce, float %b.coerce, float %c.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%4 = bitcast float %c.coerce to i32
%tmp2.0.extract.trunc = trunc i32 %4 to i16
%5 = bitcast i16 %tmp2.0.extract.trunc to half
%mul = fmul half %1, %3
%sub2 = fsub half %mul, %5
%6 = bitcast half %sub2 to i16
%tmp4.0.insert.ext = zext i16 %6 to i32
%7 = bitcast i32 %tmp4.0.insert.ext to float
ret float %7
; CHECK-LABEL: VFNMS:
; CHECK-HARDFP-FULLFP16-FAST: vfnms.f16 s2, s0, s1
; CHECK-HARDFP-FULLFP16-FAST-NEXT: vmov.f32 s0, s2
}
; TODO:
; 17. VMAXNM
; 18. VMINNM
; 19. VMLA
define float @VMLA(float %a.coerce, float %b.coerce, float %c.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%4 = bitcast float %c.coerce to i32
%tmp2.0.extract.trunc = trunc i32 %4 to i16
%5 = bitcast i16 %tmp2.0.extract.trunc to half
%mul = fmul half %1, %3
%add = fadd half %5, %mul
%6 = bitcast half %add to i16
%tmp4.0.insert.ext = zext i16 %6 to i32
%7 = bitcast i32 %tmp4.0.insert.ext to float
ret float %7
; CHECK-LABEL: VMLA:
; CHECK-HARDFP-FULLFP16: vmla.f16 s2, s0, s1
; CHECK-HARDFP-FULLFP16-NEXT: vmov.f32 s0, s2
}
; 20. VMLS
define float @VMLS(float %a.coerce, float %b.coerce, float %c.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%4 = bitcast float %c.coerce to i32
%tmp2.0.extract.trunc = trunc i32 %4 to i16
%5 = bitcast i16 %tmp2.0.extract.trunc to half
%mul = fmul half %1, %3
%add = fsub half %5, %mul
%6 = bitcast half %add to i16
%tmp4.0.insert.ext = zext i16 %6 to i32
%7 = bitcast i32 %tmp4.0.insert.ext to float
ret float %7
; CHECK-LABEL: VMLS:
; CHECK-HARDFP-FULLFP16: vmls.f16 s2, s0, s1
; CHECK-HARDFP-FULLFP16-NEXT: vmov.f32 s0, s2
}
; TODO: fix immediates.
; 21. VMOV (between general-purpose register and half-precision register)
[ARM] FP16 mov imm pattern This is a follow up of r324321, adding a match pattern for mov with a FP16 immediate (also fixing operand vfp_f16imm that wasn't even compiling). Differential Revision: https://reviews.llvm.org/D42973 llvm-svn: 324456
2018-02-07 16:37:17 +08:00
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; 22. VMOV (immediate)
[ARM] FP16 mov imm pattern This is a follow up of r324321, adding a match pattern for mov with a FP16 immediate (also fixing operand vfp_f16imm that wasn't even compiling). Differential Revision: https://reviews.llvm.org/D42973 llvm-svn: 324456
2018-02-07 16:37:17 +08:00
define i32 @movi(i32 %a.coerce) {
entry:
%tmp.0.extract.trunc = trunc i32 %a.coerce to i16
%0 = bitcast i16 %tmp.0.extract.trunc to half
%add = fadd half %0, 0xHC000
%1 = bitcast half %add to i16
%tmp2.0.insert.ext = zext i16 %1 to i32
ret i32 %tmp2.0.insert.ext
; CHECK-LABEL: movi:
; CHECK-HARDFP-FULLFP16: vmov.f16 s0, #-2.000000e+00
}
[ARM] Armv8.2-A FP16 code generation (part 3/3) This adds most of the FP16 codegen support, but these areas need further work: - FP16 literals and immediates are not properly supported yet (e.g. literal pool needs work), - Instructions that are generated from intrinsics (e.g. vabs) haven't been added. This will be addressed in follow-up patches. Differential Revision: https://reviews.llvm.org/D42849 llvm-svn: 324321
2018-02-06 16:43:56 +08:00
; 23. VMUL
define float @Mul(float %a.coerce, float %b.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%add = fmul half %1, %3
%4 = bitcast half %add to i16
%tmp4.0.insert.ext = zext i16 %4 to i32
%5 = bitcast i32 %tmp4.0.insert.ext to float
ret float %5
; CHECK-LABEL: Mul:
; CHECK-SOFT: bl __aeabi_h2f
; CHECK-SOFT: bl __aeabi_h2f
; CHECK-SOFT: bl __aeabi_fmul
; CHECK-SOFT: bl __aeabi_f2h
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: vmul.f32
; CHECK-SOFTFP-VFP3: bl __aeabi_f2h
; CHECK-SOFTFP-FP16: vmov [[S2:s[0-9]]], r1
; CHECK-SOFTFP-FP16: vmov [[S0:s[0-9]]], r0
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 [[S2]], [[S2]]
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 [[S0]], [[S0]]
; CHECK-SOFTFP-FP16: vmul.f32 [[S0]], [[S0]], [[S2]]
; CHECK-SOFTFP-FP16: vcvtb.f16.f32 [[S0]], [[S0]]
; CHECK-SOFTFP-FP16: vmov r0, s0
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S0:s[0-9]]], r1
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S2:s[0-9]]], r0
; CHECK-SOFTFP-FULLFP16: vmul.f16 [[S0]], [[S2]], [[S0]]
; CHECK-SOFTFP-FULLFP16-NEXT: vmov.f16 r0, s0
; CHECK-HARDFP-VFP3: vmov r{{.}}, s0
; CHECK-HARDFP-VFP3: vmov{{.*}}, s1
; CHECK-HARDFP-VFP3: bl __aeabi_h2f
; CHECK-HARDFP-VFP3: bl __aeabi_h2f
; CHECK-HARDFP-VFP3: vmul.f32
; CHECK-HARDFP-VFP3: bl __aeabi_f2h
; CHECK-HARDFP-VFP3: vmov s0, r0
; CHECK-HARDFP-FP16: vcvtb.f32.f16 [[S2:s[0-9]]], s1
; CHECK-HARDFP-FP16: vcvtb.f32.f16 [[S0:s[0-9]]], s0
; CHECK-HARDFP-FP16: vmul.f32 [[S0]], [[S0]], [[S2]]
; CHECK-HARDFP-FP16: vcvtb.f16.f32 [[S0]], [[S0]]
; CHECK-HARDFP-FULLFP16: vmul.f16 s0, s0, s1
}
; 24. VNEG
define float @Neg(float %a.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = fsub half -0.000000e+00, %1
%3 = bitcast half %2 to i16
%tmp4.0.insert.ext = zext i16 %3 to i32
%4 = bitcast i32 %tmp4.0.insert.ext to float
ret float %4
; CHECK-LABEL: Neg:
; CHECK-HARDFP-FULLFP16: vneg.f16 s0, s0
}
; 25. VNMLA
define float @VNMLA(float %a.coerce, float %b.coerce, float %c.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%4 = bitcast float %c.coerce to i32
%tmp2.0.extract.trunc = trunc i32 %4 to i16
%5 = bitcast i16 %tmp2.0.extract.trunc to half
%add = fmul half %1, %3
%add2 = fsub half -0.000000e+00, %add
%add3 = fsub half %add2, %5
%6 = bitcast half %add3 to i16
%tmp4.0.insert.ext = zext i16 %6 to i32
%7 = bitcast i32 %tmp4.0.insert.ext to float
ret float %7
; CHECK-LABEL: VNMLA:
; CHECK-HARDFP-FULLFP16: vnmla.f16 s2, s0, s1
; CHECK-HARDFP-FULLFP16: vmov.f32 s0, s2
}
; 26. VNMLS
define float @VNMLS(float %a.coerce, float %b.coerce, float %c.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%4 = bitcast float %c.coerce to i32
%tmp2.0.extract.trunc = trunc i32 %4 to i16
%5 = bitcast i16 %tmp2.0.extract.trunc to half
%add = fmul half %1, %3
%add2 = fsub half %add, %5
%6 = bitcast half %add2 to i16
%tmp4.0.insert.ext = zext i16 %6 to i32
%7 = bitcast i32 %tmp4.0.insert.ext to float
ret float %7
; CHECK-LABEL: VNMLS:
; CHECK-HARDFP-FULLFP16: vnmls.f16 s2, s0, s1
; CHECK-HARDFP-FULLFP16: vmov.f32 s0, s2
}
; 27. VNMUL
define float @NMul(float %a.coerce, float %b.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%add = fmul half %1, %3
%add2 = fsub half -0.0, %add
%4 = bitcast half %add2 to i16
%tmp4.0.insert.ext = zext i16 %4 to i32
%5 = bitcast i32 %tmp4.0.insert.ext to float
ret float %5
; CHECK-LABEL: NMul:
; CHECK-HARDFP-FULLFP16: vnmul.f16 s0, s0, s1
}
; 28. VRINTA
; 29. VRINTM
; 30. VRINTN
; 31. VRINTP
; 32. VRINTR
; 33. VRINTX
; 34. VRINTZ
; 35. VSELEQ
; 36. VSELGE
; 37. VSELGT
; 38. VSELVS
; 39. VSQRT
; 40. VSUB
define float @Sub(float %a.coerce, float %b.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
%2 = bitcast float %b.coerce to i32
%tmp1.0.extract.trunc = trunc i32 %2 to i16
%3 = bitcast i16 %tmp1.0.extract.trunc to half
%add = fsub half %1, %3
%4 = bitcast half %add to i16
%tmp4.0.insert.ext = zext i16 %4 to i32
%5 = bitcast i32 %tmp4.0.insert.ext to float
ret float %5
; CHECK-LABEL: Sub:
; CHECK-SOFT: bl __aeabi_h2f
; CHECK-SOFT: bl __aeabi_h2f
; CHECK-SOFT: bl __aeabi_fsub
; CHECK-SOFT: bl __aeabi_f2h
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: bl __aeabi_h2f
; CHECK-SOFTFP-VFP3: vsub.f32
; CHECK-SOFTFP-VFP3: bl __aeabi_f2h
; CHECK-SOFTFP-FP16: vmov [[S2:s[0-9]]], r1
; CHECK-SOFTFP-FP16: vmov [[S0:s[0-9]]], r0
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 [[S2]], [[S2]]
; CHECK-SOFTFP-FP16: vcvtb.f32.f16 [[S0]], [[S0]]
; CHECK-SOFTFP-FP16: vsub.f32 [[S0]], [[S0]], [[S2]]
; CHECK-SOFTFP-FP16: vcvtb.f16.f32 [[S0]], [[S0]]
; CHECK-SOFTFP-FP16: vmov r0, s0
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S0:s[0-9]]], r1
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S2:s[0-9]]], r0
; CHECK-SOFTFP-FULLFP16: vsub.f16 [[S0]], [[S2]], [[S0]]
; CHECK-SOFTFP-FULLFP16-NEXT: vmov.f16 r0, s0
; CHECK-HARDFP-VFP3: vmov r{{.}}, s0
; CHECK-HARDFP-VFP3: vmov{{.*}}, s1
; CHECK-HARDFP-VFP3: bl __aeabi_h2f
; CHECK-HARDFP-VFP3: bl __aeabi_h2f
; CHECK-HARDFP-VFP3: vsub.f32
; CHECK-HARDFP-VFP3: bl __aeabi_f2h
; CHECK-HARDFP-VFP3: vmov s0, r0
; CHECK-HARDFP-FP16: vcvtb.f32.f16 [[S2:s[0-9]]], s1
; CHECK-HARDFP-FP16: vcvtb.f32.f16 [[S0:s[0-9]]], s0
; CHECK-HARDFP-FP16: vsub.f32 [[S0]], [[S0]], [[S2]]
; CHECK-HARDFP-FP16: vcvtb.f16.f32 [[S0]], [[S0]]
; CHECK-HARDFP-FULLFP16: vsub.f16 s0, s0, s1
[ARM] Armv8.2-A FP16 code generation (part 1/3) This is the groundwork for Armv8.2-A FP16 code generation . Clang passes and returns _Float16 values as floats, together with the required bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318. We will implement half-precision argument passing/returning lowering in the ARM backend soon, but for now this means that this: _Float16 sub(_Float16 a, _Float16 b) { return a + b; } gets lowered to this: define float @sub(float %a.coerce, float %b.coerce) { entry: %0 = bitcast float %a.coerce to i32 %tmp.0.extract.trunc = trunc i32 %0 to i16 %1 = bitcast i16 %tmp.0.extract.trunc to half <SNIP> %add = fadd half %1, %3 <SNIP> } When FullFP16 is *not* supported, we don't make f16 a legal type, and we get legalization for "free", i.e. nothing changes and everything works as before. And also f16 argument passing/returning is handled. When FullFP16 is supported, we do make f16 a legal type, and have 2 places that we need to patch up: f16 argument passing and returning, which involves minor tweaks to avoid unnecessary code generation for some bitcasts. As a "demonstrator" that this works for the different FP16, FullFP16, softfp modes, etc., I've added match rules to the VSUB instruction description showing that we can codegen this instruction from IR, but more importantly, also to some conversion instructions. These conversions were causing issue before in the FP16 and FullFP16 cases. I've also added match rules to the VLDRH and VSTRH desriptions, so that we can actually compile the entire half-precision sub code example above. This showed that these loads and stores had the wrong addressing mode specified: AddrMode5 instead of AddrMode5FP16, which turned out not be implemented at all, so that has also been added. This is the minimal patch that shows all the different moving parts. In patch 2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the remaining Armv8.2-A FP16 instruction descriptions. Thanks to Sam Parker and Oliver Stannard for their help and reviews! Differential Revision: https://reviews.llvm.org/D38315 llvm-svn: 323512
2018-01-26 17:26:40 +08:00
}
[Thumb] Handle addressing mode AddrMode5FP16 This addressing mode wasn't checked, so we were running in an assert. Differential Revision: https://reviews.llvm.org/D43179 llvm-svn: 324996
2018-02-13 18:29:03 +08:00
; Check for VSTRH with a FCONSTH, this checks that addressing mode
; AddrMode5FP16 is supported.
define i32 @ThumbAddrMode5FP16(i32 %A.coerce) {
entry:
%S = alloca half, align 2
%tmp.0.extract.trunc = trunc i32 %A.coerce to i16
%0 = bitcast i16 %tmp.0.extract.trunc to half
%S.0.S.0..sroa_cast = bitcast half* %S to i8*
store volatile half 0xH3C00, half* %S, align 2
%S.0.S.0. = load volatile half, half* %S, align 2
%add = fadd half %S.0.S.0., %0
%1 = bitcast half %add to i16
%tmp2.0.insert.ext = zext i16 %1 to i32
ret i32 %tmp2.0.insert.ext
; CHECK-LABEL: ThumbAddrMode5FP16
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S0:s[0-9]]], #1.000000e+00
; CHECK-SOFTFP-FULLFP16: vstr.16 [[S0]], [sp, #{{.}}]
; CHECK-SOFTFP-FULLFP16: vmov.f16 [[S0_2:s[0-9]]], r0
; CHECK-SOFTFP-FULLFP16: vldr.16 [[S2:s[0-9]]], [sp, #{{.}}]
; CHECK-SOFTFP-FULLFP16: vadd.f16 s{{.}}, [[S2]], [[S0_2]]
}
[ARM] f16 stack spill/reloads This adds support for handling f16 stack spills/reloads. Differential Revision: https://reviews.llvm.org/D43280 llvm-svn: 325130
2018-02-14 23:09:09 +08:00
; Test function calls to check store/load reg to/from stack
define i32 @fn1() {
entry:
%coerce = alloca half, align 2
%tmp2 = alloca i32, align 4
store half 0xH7C00, half* %coerce, align 2
%0 = load i32, i32* %tmp2, align 4
%call = call i32 bitcast (i32 (...)* @fn2 to i32 (i32)*)(i32 %0)
store half 0xH7C00, half* %coerce, align 2
%1 = load i32, i32* %tmp2, align 4
%call3 = call i32 bitcast (i32 (...)* @fn3 to i32 (i32)*)(i32 %1)
ret i32 %call3
; CHECK-SPILL-RELOAD-LABEL: fn1:
; CHECK-SPILL-RELOAD: vstr.16 s0, [sp, #{{.}}] @ 2-byte Spill
; CHECK-SPILL-RELOAD-NEXT: bl fn2
; CHECK-SPILL-RELOAD-NEXT: vldr.16 s0, [sp, #{{.}}] @ 2-byte Reload
}
declare dso_local i32 @fn2(...)
declare dso_local i32 @fn3(...)