Initial support for single-precision FP using NEON. Added "neonfp" attribute to enable. Added patterns for some binary FP operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@78081 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
David Goodwin 2009-08-04 17:53:06 +00:00
parent c7a6da6e14
commit 42a83f2d15
16 changed files with 168 additions and 16 deletions

View File

@ -32,6 +32,9 @@ def ArchV6T2 : SubtargetFeature<"v6t2", "ARMArchVersion", "V6T2",
"ARM v6t2">;
def ArchV7A : SubtargetFeature<"v7a", "ARMArchVersion", "V7A",
"ARM v7A">;
def FeatureNEONFP : SubtargetFeature<"neonfp", "UseNEONForSinglePrecisionFP",
"true",
"Use NEON for single-precision FP">;
def FeatureVFP2 : SubtargetFeature<"vfp2", "ARMFPUType", "VFPv2",
"Enable VFP2 instructions">;
def FeatureVFP3 : SubtargetFeature<"vfp3", "ARMFPUType", "VFPv3",

View File

@ -1080,6 +1080,14 @@ class ASbI<bits<8> opcod, dag oops, dag iops, string opc,
let Inst{11-8} = 0b1010;
}
// Single precision, binary if no NEON
// Same as ASbI except not available if NEON is enabled
class ASbIn<bits<8> opcod, dag oops, dag iops, string opc,
string asm, list<dag> pattern>
: ASbI<opcod, oops, iops, opc, asm, pattern> {
list<Predicate> Predicates = [HasVFP2,DontUseNEONForFP];
}
// VFP conversion instructions
class AVConv1I<bits<8> opcod1, bits<4> opcod2, bits<4> opcod3,
dag oops, dag iops, string opc, string asm, list<dag> pattern>
@ -1220,3 +1228,9 @@ class NVSetLane<bits<8> opcod1, bits<4> opcod2, bits<2> opcod3,
class NVDup<bits<8> opcod1, bits<4> opcod2, bits<2> opcod3,
dag oops, dag iops, string opc, string asm, list<dag> pattern>
: NVLaneOp<opcod1, opcod2, opcod3, oops, iops, NEONDupFrm, opc, asm, pattern>;
// NEONFPPat - Same as Pat<>, but requires that the compiler be using NEON
// for single-precision FP.
class NEONFPPat<dag pattern, dag result> : Pat<pattern, result> {
list<Predicate> Predicates = [HasNEON,UseNEONForFP];
}

View File

@ -104,6 +104,8 @@ def HasV7 : Predicate<"Subtarget->hasV7Ops()">;
def HasVFP2 : Predicate<"Subtarget->hasVFP2()">;
def HasVFP3 : Predicate<"Subtarget->hasVFP3()">;
def HasNEON : Predicate<"Subtarget->hasNEON()">;
def UseNEONForFP : Predicate<"Subtarget->useNEONForSinglePrecisionFP()">;
def DontUseNEONForFP : Predicate<"!Subtarget->useNEONForSinglePrecisionFP()">;
def IsThumb : Predicate<"Subtarget->isThumb()">;
def IsThumb1Only : Predicate<"Subtarget->isThumb1Only()">;
def IsThumb2 : Predicate<"Subtarget->isThumb2()">;

View File

@ -283,6 +283,13 @@ class N3VQ<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
let isCommutable = Commutable;
}
// Basic 3-register operations, scalar single-precision
class N3VDs<SDNode OpNode, NeonI Inst>
: NEONFPPat<(f32 (OpNode SPR:$a, SPR:$b)),
(EXTRACT_SUBREG (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0),
(INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$b, arm_ssubreg_0)),
arm_ssubreg_0)>;
// Basic 3-register intrinsics, both double- and quad-register.
class N3VDInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
string OpcodeStr, ValueType ResTy, ValueType OpTy,
@ -319,6 +326,15 @@ class N3VQMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
[(set QPR:$dst, (Ty (OpNode QPR:$src1,
(Ty (MulOp QPR:$src2, QPR:$src3)))))]>;
// Multiply-Add/Sub operations, scalar single-precision
class N3VDMulOps<SDNode MulNode, SDNode OpNode, NeonI Inst>
: NEONFPPat<(f32 (OpNode SPR:$acc,
(f32 (MulNode SPR:$a, SPR:$b)))),
(EXTRACT_SUBREG (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$acc, arm_ssubreg_0),
(INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0),
(INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$b, arm_ssubreg_0)),
arm_ssubreg_0)>;
// Neon 3-argument intrinsics, both double- and quad-register.
// The destination register is also used as the first source operand register.
class N3VDInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
@ -886,6 +902,9 @@ defm VADDHN : N3VNInt_HSD<0,1,0b0100,0, "vaddhn.i", int_arm_neon_vaddhn, 1>;
// VRADDHN : Vector Rounding Add and Narrow Returning High Half (D = Q + Q)
defm VRADDHN : N3VNInt_HSD<1,1,0b0100,0, "vraddhn.i", int_arm_neon_vraddhn, 1>;
// Vector Add Operations used for single-precision FP
def : N3VDs<fadd, VADDfd>;
// Vector Multiply Operations.
// VMUL : Vector Multiply (integer, polynomial and floating-point)
@ -908,6 +927,9 @@ def VMULLp : N3VLInt<0, 1, 0b00, 0b1110, 0, "vmull.p8", v8i16, v8i8,
// VQDMULL : Vector Saturating Doubling Multiply Long (Q = D * D)
defm VQDMULL : N3VLInt_HS<0,1,0b1101,0, "vqdmull.s", int_arm_neon_vqdmull, 1>;
// Vector Multiply Operations used for single-precision FP
def : N3VDs<fmul, VMULfd>;
// Vector Multiply-Accumulate and Multiply-Subtract Operations.
// VMLA : Vector Multiply Accumulate (integer and floating-point)
@ -929,6 +951,9 @@ defm VMLSLu : N3VLInt3_QHS<1,1,0b1010,0, "vmlsl.u", int_arm_neon_vmlslu>;
// VQDMLSL : Vector Saturating Doubling Multiply Subtract Long (Q -= D * D)
defm VQDMLSL : N3VLInt3_HS<0, 1, 0b1011, 0, "vqdmlsl.s", int_arm_neon_vqdmlsl>;
// Vector Multiply-Accumulate/Subtract used for single-precision FP
def : N3VDMulOps<fmul, fadd, VMLAfd>;
// Vector Subtract Operations.
// VSUB : Vector Subtract (integer and floating-point)
@ -952,6 +977,9 @@ defm VSUBHN : N3VNInt_HSD<0,1,0b0110,0, "vsubhn.i", int_arm_neon_vsubhn, 0>;
// VRSUBHN : Vector Rounding Subtract and Narrow Returning High Half (D=Q-Q)
defm VRSUBHN : N3VNInt_HSD<1,1,0b0110,0, "vrsubhn.i", int_arm_neon_vrsubhn, 0>;
// Vector Sub Operations used for single-precision FP
def : N3VDs<fsub, VSUBfd>;
// Vector Comparisons.
// VCEQ : Vector Compare Equal

View File

@ -98,7 +98,7 @@ def FADDD : ADbI<0b11100011, (outs DPR:$dst), (ins DPR:$a, DPR:$b),
"faddd", " $dst, $a, $b",
[(set DPR:$dst, (fadd DPR:$a, DPR:$b))]>;
def FADDS : ASbI<0b11100011, (outs SPR:$dst), (ins SPR:$a, SPR:$b),
def FADDS : ASbIn<0b11100011, (outs SPR:$dst), (ins SPR:$a, SPR:$b),
"fadds", " $dst, $a, $b",
[(set SPR:$dst, (fadd SPR:$a, SPR:$b))]>;
@ -125,7 +125,7 @@ def FMULD : ADbI<0b11100010, (outs DPR:$dst), (ins DPR:$a, DPR:$b),
"fmuld", " $dst, $a, $b",
[(set DPR:$dst, (fmul DPR:$a, DPR:$b))]>;
def FMULS : ASbI<0b11100010, (outs SPR:$dst), (ins SPR:$a, SPR:$b),
def FMULS : ASbIn<0b11100010, (outs SPR:$dst), (ins SPR:$a, SPR:$b),
"fmuls", " $dst, $a, $b",
[(set SPR:$dst, (fmul SPR:$a, SPR:$b))]>;
@ -154,7 +154,7 @@ def FSUBD : ADbI<0b11100011, (outs DPR:$dst), (ins DPR:$a, DPR:$b),
let Inst{6} = 1;
}
def FSUBS : ASbI<0b11100011, (outs SPR:$dst), (ins SPR:$a, SPR:$b),
def FSUBS : ASbIn<0b11100011, (outs SPR:$dst), (ins SPR:$a, SPR:$b),
"fsubs", " $dst, $a, $b",
[(set SPR:$dst, (fsub SPR:$a, SPR:$b))]> {
let Inst{6} = 1;
@ -317,7 +317,7 @@ def FMACD : ADbI<0b11100000, (outs DPR:$dst), (ins DPR:$dstin, DPR:$a, DPR:$b),
[(set DPR:$dst, (fadd (fmul DPR:$a, DPR:$b), DPR:$dstin))]>,
RegConstraint<"$dstin = $dst">;
def FMACS : ASbI<0b11100000, (outs SPR:$dst), (ins SPR:$dstin, SPR:$a, SPR:$b),
def FMACS : ASbIn<0b11100000, (outs SPR:$dst), (ins SPR:$dstin, SPR:$a, SPR:$b),
"fmacs", " $dst, $a, $b",
[(set SPR:$dst, (fadd (fmul SPR:$a, SPR:$b), SPR:$dstin))]>,
RegConstraint<"$dstin = $dst">;
@ -339,7 +339,7 @@ def FNMACD : ADbI<0b11100000, (outs DPR:$dst), (ins DPR:$dstin, DPR:$a, DPR:$b),
let Inst{6} = 1;
}
def FNMACS : ASbI<0b11100000, (outs SPR:$dst), (ins SPR:$dstin, SPR:$a, SPR:$b),
def FNMACS : ASbIn<0b11100000, (outs SPR:$dst), (ins SPR:$dstin, SPR:$a, SPR:$b),
"fnmacs", " $dst, $a, $b",
[(set SPR:$dst, (fadd (fneg (fmul SPR:$a, SPR:$b)), SPR:$dstin))]>,
RegConstraint<"$dstin = $dst"> {

View File

@ -25,6 +25,7 @@ ARMSubtarget::ARMSubtarget(const std::string &TT, const std::string &FS,
bool isThumb)
: ARMArchVersion(V4T)
, ARMFPUType(None)
, UseNEONForSinglePrecisionFP(false)
, IsThumb(isThumb)
, ThumbMode(Thumb1)
, IsR9Reserved(ReserveR9)

View File

@ -42,6 +42,9 @@ protected:
/// ARMFPUType - Floating Point Unit type.
ARMFPEnum ARMFPUType;
/// UseNEONForSinglePrecisionFP - if NEON is available use for FP
bool UseNEONForSinglePrecisionFP;
/// IsThumb - True if we are in thumb mode, false if in ARM mode.
bool IsThumb;
@ -98,6 +101,8 @@ protected:
bool hasVFP2() const { return ARMFPUType >= VFPv2; }
bool hasVFP3() const { return ARMFPUType >= VFPv3; }
bool hasNEON() const { return ARMFPUType >= NEON; }
bool useNEONForSinglePrecisionFP() const {
return hasNEON() && UseNEONForSinglePrecisionFP; }
bool isTargetDarwin() const { return TargetType == isDarwin; }
bool isTargetELF() const { return TargetType == isELF; }

10
test/CodeGen/ARM/fadds.ll Normal file
View File

@ -0,0 +1,10 @@
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fadds\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {vadd.f32\\W*d\[0-9\]+,\\W*d\[0-9\]+,\\W*d\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fadds\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %a, float %b) {
entry:
%0 = fadd float %a, %b
ret float %0
}

10
test/CodeGen/ARM/fdivs.ll Normal file
View File

@ -0,0 +1,10 @@
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fdivs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {fdivs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fdivs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %a, float %b) {
entry:
%0 = fdiv float %a, %b
ret float %0
}

11
test/CodeGen/ARM/fmacs.ll Normal file
View File

@ -0,0 +1,11 @@
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fmacs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {vmla.f32\\W*d\[0-9\]+,\\W*d\[0-9\]+,\\W*d\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fmacs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %acc, float %a, float %b) {
entry:
%0 = fmul float %a, %b
%1 = fadd float %acc, %0
ret float %1
}

11
test/CodeGen/ARM/fmscs.ll Normal file
View File

@ -0,0 +1,11 @@
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fmscs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {fmscs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fmscs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %acc, float %a, float %b) {
entry:
%0 = fmul float %a, %b
%1 = fsub float %0, %acc
ret float %1
}

10
test/CodeGen/ARM/fmuls.ll Normal file
View File

@ -0,0 +1,10 @@
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fmuls\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {vmul.f32\\W*d\[0-9\]+,\\W*d\[0-9\]+,\\W*d\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fmuls\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %a, float %b) {
entry:
%0 = fmul float %a, %b
ret float %0
}

View File

@ -0,0 +1,12 @@
; XFAIL: *
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fnmacs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {vmls.f32\\W*d\[0-9\]+,\\W*d\[0-9\]+,\\W*d\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fnmacs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %acc, float %a, float %b) {
entry:
%0 = fmul float %a, %b
%1 = fsub float %acc, %0
ret float %1
}

View File

@ -0,0 +1,13 @@
; XFAIL: *
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fnmscs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {fnmscs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fnmscs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %acc, float %a, float %b) {
entry:
%0 = fmul float %a, %b
%1 = fsub float 0.0, %0
%2 = fsub float %1, %acc
ret float %2
}

View File

@ -0,0 +1,12 @@
; XFAIL: *
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fnmuls\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {fnmuls\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fnmuls\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %a, float %b) {
entry:
%0 = fmul float %a, %b
%1 = fsub float 0.0, %0
ret float %1
}

10
test/CodeGen/ARM/fsubs.ll Normal file
View File

@ -0,0 +1,10 @@
; RUN: llvm-as < %s | llc -march=arm -mattr=+vfp2 | grep -E {fsubs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,+neonfp | grep -E {vsub.f32\\W*d\[0-9\]+,\\W*d\[0-9\]+,\\W*d\[0-9\]+} | count 1
; RUN: llvm-as < %s | llc -march=arm -mattr=+neon,-neonfp | grep -E {fsubs\\W*s\[0-9\]+,\\W*s\[0-9\]+,\\W*s\[0-9\]+} | count 1
define float @test(float %a, float %b) {
entry:
%0 = fsub float %a, %b
ret float %0
}