[ARMv8] Add NEON instructions VCVT{A, N, P, M}.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@186574 91177308-0d34-0410-b5e6-96231b3b80d8
stryku · Jul 18, 2013 · 59b3300 · 59b3300
1 parent 722a26d
commit 59b3300
Show file tree

Hide file tree

Showing 7 changed files with 220 additions and 0 deletions.
diff --git a/include/llvm/IR/IntrinsicsARM.td b/include/llvm/IR/IntrinsicsARM.td
@@ -127,6 +127,8 @@ class Neon_CvtFxToFP_Intrinsic
   : Intrinsic<[llvm_anyfloat_ty], [llvm_anyint_ty, llvm_i32_ty], [IntrNoMem]>;
 class Neon_CvtFPToFx_Intrinsic
   : Intrinsic<[llvm_anyint_ty], [llvm_anyfloat_ty, llvm_i32_ty], [IntrNoMem]>;
+class Neon_CvtFPtoInt_1Arg_Intrinsic
+  : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
 
 // The table operands for VTBL and VTBX consist of 1 to 4 v8i8 vectors.
 // Besides the table, VTBL has one other v8i8 argument and VTBX has two.
@@ -319,6 +321,16 @@ def int_arm_neon_vrecpe : Neon_1Arg_Intrinsic;
 // Vector Reciprocal Square Root Estimate.
 def int_arm_neon_vrsqrte : Neon_1Arg_Intrinsic;
 
+// Vector Conversions Between Floating-point and Integer
+def int_arm_neon_vcvtau : Neon_CvtFPtoInt_1Arg_Intrinsic;
+def int_arm_neon_vcvtas : Neon_CvtFPtoInt_1Arg_Intrinsic;
+def int_arm_neon_vcvtnu : Neon_CvtFPtoInt_1Arg_Intrinsic;
+def int_arm_neon_vcvtns : Neon_CvtFPtoInt_1Arg_Intrinsic;
+def int_arm_neon_vcvtpu : Neon_CvtFPtoInt_1Arg_Intrinsic;
+def int_arm_neon_vcvtps : Neon_CvtFPtoInt_1Arg_Intrinsic;
+def int_arm_neon_vcvtmu : Neon_CvtFPtoInt_1Arg_Intrinsic;
+def int_arm_neon_vcvtms : Neon_CvtFPtoInt_1Arg_Intrinsic;
+
 // Vector Conversions Between Floating-point and Fixed-point.
 def int_arm_neon_vcvtfp2fxs : Neon_CvtFPToFx_Intrinsic;
 def int_arm_neon_vcvtfp2fxu : Neon_CvtFPToFx_Intrinsic;

diff --git a/lib/Target/ARM/ARMInstrFormats.td b/lib/Target/ARM/ARMInstrFormats.td
@@ -1964,6 +1964,35 @@ class N2V<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18, bits<2> op17_16,
   let Inst{5}     = Vm{4};
 }
 
+// Same as N2V but not predicated.
+class N2Vnp<bits<2> op17_16, bits<3> op10_8, bit op7, bit op6,
+            dag oops, dag iops, InstrItinClass itin, string OpcodeStr,
+            string Dt, ValueType ResTy, ValueType OpTy, list<dag> pattern>
+   : NeonInp<oops, iops, AddrModeNone, IndexModeNone, N2RegFrm, itin,
+             OpcodeStr, Dt, "$Vd, $Vm", "", pattern> {
+  bits<5> Vd;
+  bits<5> Vm;
+
+  // Encode instruction operands
+  let Inst{22}    = Vd{4};
+  let Inst{15-12} = Vd{3-0};
+  let Inst{5}     = Vm{4};
+  let Inst{3-0}   = Vm{3-0};
+
+  // Encode constant bits
+  let Inst{27-23} = 0b00111;
+  let Inst{21-20} = 0b11;
+  let Inst{19-18} = 0b10;
+  let Inst{17-16} = op17_16;
+  let Inst{11} = 0;
+  let Inst{10-8} = op10_8;
+  let Inst{7} = op7;
+  let Inst{6} = op6;
+  let Inst{4} = 0;
+
+  let DecoderNamespace = "NEON";
+}
+
 // Same as N2V except it doesn't have a datatype suffix.
 class N2VX<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18, bits<2> op17_16,
            bits<5> op11_7, bit op6, bit op4,

diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
@@ -2379,6 +2379,21 @@ class N2VQInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
         (ins QPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
         [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vm))))]>;
 
+// Same as above, but not predicated.
+class N2VDIntnp<bits<2> op17_16, bits<3> op10_8, bit op7,
+              InstrItinClass itin, string OpcodeStr, string Dt,
+              ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
+  : N2Vnp<op17_16, op10_8, op7, 0,  (outs DPR:$Vd), (ins DPR:$Vm),
+          itin, OpcodeStr, Dt, ResTy, OpTy,
+          [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vm))))]>;
+
+class N2VQIntnp<bits<2> op17_16, bits<3> op10_8, bit op7,
+              InstrItinClass itin, string OpcodeStr, string Dt,
+              ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
+  : N2Vnp<op17_16, op10_8, op7, 1,  (outs QPR:$Vd), (ins QPR:$Vm),
+          itin, OpcodeStr, Dt, ResTy, OpTy,
+          [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vm))))]>;
+
 // Narrow 2-register operations.
 class N2VN<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
            bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
@@ -5431,6 +5446,26 @@ def  VCVTs2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt", "f32.s32",
 def  VCVTu2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt", "f32.u32",
                      v4f32, v4i32, uint_to_fp>;
 
+// VCVT{A, N, P, M}
+multiclass VCVT_FPI<string op, bits<3> op10_8, SDPatternOperator IntS,
+                    SDPatternOperator IntU> {
+  let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in {
+    def SD : N2VDIntnp<0b11, op10_8, 0, NoItinerary, !strconcat("vcvt", op),
+                       "s32.f32", v2i32, v2f32, IntS>, Requires<[HasV8, HasNEON]>;
+    def SQ : N2VQIntnp<0b11, op10_8, 0, NoItinerary, !strconcat("vcvt", op),
+                       "s32.f32", v4i32, v4f32, IntS>, Requires<[HasV8, HasNEON]>;
+    def UD : N2VDIntnp<0b11, op10_8, 1, NoItinerary, !strconcat("vcvt", op),
+                       "u32.f32", v2i32, v2f32, IntU>, Requires<[HasV8, HasNEON]>;
+    def UQ : N2VQIntnp<0b11, op10_8, 1, NoItinerary, !strconcat("vcvt", op),
+                       "u32.f32", v4i32, v4f32, IntU>, Requires<[HasV8, HasNEON]>;
+  }
+}
+
+defm VCVTAn : VCVT_FPI<"a", 0b000, int_arm_neon_vcvtas, int_arm_neon_vcvtau>;
+defm VCVTNn : VCVT_FPI<"n", 0b001, int_arm_neon_vcvtns, int_arm_neon_vcvtnu>;
+defm VCVTPn : VCVT_FPI<"p", 0b010, int_arm_neon_vcvtps, int_arm_neon_vcvtpu>;
+defm VCVTMn : VCVT_FPI<"m", 0b011, int_arm_neon_vcvtms, int_arm_neon_vcvtmu>;
+
 //   VCVT     : Vector Convert Between Floating-Point and Fixed-Point.
 let DecoderMethod = "DecodeVCVTD" in {
 def VCVTf2xsd : N2VCvtD<0, 1, 0b1111, 0, 1, "vcvt", "s32.f32",

diff --git a/test/MC/ARM/neon-v8.s b/test/MC/ARM/neon-v8.s
@@ -8,3 +8,39 @@ vminnm.f32 d5, d4, d30
 @ CHECK: vminnm.f32 d5, d4, d30 @ encoding: [0x3e,0x5f,0x24,0xf3]
 vminnm.f32 q0, q13, q2
 @ CHECK: vminnm.f32 q0, q13, q2 @ encoding: [0xd4,0x0f,0x2a,0xf3]
+
+vcvta.s32.f32	d4, d6
+@ CHECK: vcvta.s32.f32	d4, d6 @ encoding: [0x06,0x40,0xbb,0xf3]
+vcvta.u32.f32	d12, d10
+@ CHECK: vcvta.u32.f32	d12, d10 @ encoding: [0x8a,0xc0,0xbb,0xf3]
+vcvta.s32.f32	q4, q6
+@ CHECK: vcvta.s32.f32	q4, q6 @ encoding: [0x4c,0x80,0xbb,0xf3]
+vcvta.u32.f32	q4, q10
+@ CHECK: vcvta.u32.f32	q4, q10 @ encoding: [0xe4,0x80,0xbb,0xf3]
+
+vcvtm.s32.f32	d1, d30
+@ CHECK: vcvtm.s32.f32	d1, d30 @ encoding: [0x2e,0x13,0xbb,0xf3]
+vcvtm.u32.f32	d12, d10
+@ CHECK: vcvtm.u32.f32	d12, d10 @ encoding: [0x8a,0xc3,0xbb,0xf3]
+vcvtm.s32.f32	q1, q10
+@ CHECK: vcvtm.s32.f32	q1, q10 @ encoding: [0x64,0x23,0xbb,0xf3]
+vcvtm.u32.f32	q13, q1
+@ CHECK: vcvtm.u32.f32	q13, q1 @ encoding: [0xc2,0xa3,0xfb,0xf3]
+
+vcvtn.s32.f32	d15, d17
+@ CHECK: vcvtn.s32.f32	d15, d17 @ encoding: [0x21,0xf1,0xbb,0xf3]
+vcvtn.u32.f32	d5, d3
+@ CHECK: vcvtn.u32.f32	d5, d3 @ encoding: [0x83,0x51,0xbb,0xf3]
+vcvtn.s32.f32	q3, q8
+@ CHECK: vcvtn.s32.f32	q3, q8 @ encoding: [0x60,0x61,0xbb,0xf3]
+vcvtn.u32.f32	q5, q3
+@ CHECK: vcvtn.u32.f32	q5, q3 @ encoding: [0xc6,0xa1,0xbb,0xf3]
+
+vcvtp.s32.f32	d11, d21
+@ CHECK: vcvtp.s32.f32	d11, d21 @ encoding: [0x25,0xb2,0xbb,0xf3]
+vcvtp.u32.f32	d14, d23
+@ CHECK: vcvtp.u32.f32	d14, d23 @ encoding: [0xa7,0xe2,0xbb,0xf3]
+vcvtp.s32.f32	q4, q15
+@ CHECK: vcvtp.s32.f32	q4, q15 @ encoding: [0x6e,0x82,0xbb,0xf3]
+vcvtp.u32.f32	q9, q8
+@ CHECK: vcvtp.u32.f32	q9, q8 @ encoding: [0xe0,0x22,0xfb,0xf3]
diff --git a/test/MC/ARM/thumb-neon-v8.s b/test/MC/ARM/thumb-neon-v8.s
@@ -8,3 +8,39 @@ vminnm.f32 d5, d4, d30
 @ CHECK: vminnm.f32 d5, d4, d30 @ encoding: [0x24,0xff,0x3e,0x5f]
 vminnm.f32 q0, q13, q2
 @ CHECK: vminnm.f32 q0, q13, q2 @ encoding: [0x2a,0xff,0xd4,0x0f]
+
+vcvta.s32.f32	d4, d6
+@ CHECK: vcvta.s32.f32	d4, d6 @ encoding: [0xbb,0xff,0x06,0x40]
+vcvta.u32.f32	d12, d10
+@ CHECK: vcvta.u32.f32	d12, d10 @ encoding: [0xbb,0xff,0x8a,0xc0]
+vcvta.s32.f32	q4, q6
+@ CHECK: vcvta.s32.f32	q4, q6 @ encoding: [0xbb,0xff,0x4c,0x80]
+vcvta.u32.f32	q4, q10
+@ CHECK: vcvta.u32.f32	q4, q10 @ encoding: [0xbb,0xff,0xe4,0x80]
+
+vcvtm.s32.f32	d1, d30
+@ CHECK: vcvtm.s32.f32	d1, d30 @ encoding: [0xbb,0xff,0x2e,0x13]
+vcvtm.u32.f32	d12, d10
+@ CHECK: vcvtm.u32.f32	d12, d10 @ encoding: [0xbb,0xff,0x8a,0xc3]
+vcvtm.s32.f32	q1, q10
+@ CHECK: vcvtm.s32.f32	q1, q10 @ encoding: [0xbb,0xff,0x64,0x23]
+vcvtm.u32.f32	q13, q1
+@ CHECK: vcvtm.u32.f32	q13, q1 @ encoding: [0xfb,0xff,0xc2,0xa3]
+
+vcvtn.s32.f32	d15, d17
+@ CHECK: vcvtn.s32.f32	d15, d17 @ encoding: [0xbb,0xff,0x21,0xf1]
+vcvtn.u32.f32	d5, d3
+@ CHECK: vcvtn.u32.f32	d5, d3 @ encoding: [0xbb,0xff,0x83,0x51]
+vcvtn.s32.f32	q3, q8
+@ CHECK: vcvtn.s32.f32	q3, q8 @ encoding: [0xbb,0xff,0x60,0x61]
+vcvtn.u32.f32	q5, q3
+@ CHECK: vcvtn.u32.f32	q5, q3 @ encoding: [0xbb,0xff,0xc6,0xa1]
+
+vcvtp.s32.f32	d11, d21
+@ CHECK: vcvtp.s32.f32	d11, d21 @ encoding: [0xbb,0xff,0x25,0xb2]
+vcvtp.u32.f32	d14, d23
+@ CHECK: vcvtp.u32.f32	d14, d23 @ encoding: [0xbb,0xff,0xa7,0xe2]
+vcvtp.s32.f32	q4, q15
+@ CHECK: vcvtp.s32.f32	q4, q15 @ encoding: [0xbb,0xff,0x6e,0x82]
+vcvtp.u32.f32	q9, q8
+@ CHECK: vcvtp.u32.f32	q9, q8 @ encoding: [0xfb,0xff,0xe0,0x22]
diff --git a/test/MC/Disassembler/ARM/neon-v8.txt b/test/MC/Disassembler/ARM/neon-v8.txt
@@ -8,3 +8,39 @@
 # CHECK: vminnm.f32 d5, d4, d30
 0xd4 0x0f 0x2a 0xf3
 # CHECK: vminnm.f32 q0, q13, q2
+
+0x06 0x40 0xbb 0xf3
+# CHECK: vcvta.s32.f32	d4, d6
+0x8a 0xc0 0xbb 0xf3
+# CHECK: vcvta.u32.f32	d12, d10
+0x4c 0x80 0xbb 0xf3
+# CHECK: vcvta.s32.f32	q4, q6
+0xe4 0x80 0xbb 0xf3
+# CHECK: vcvta.u32.f32	q4, q10
+
+0x2e 0x13 0xbb 0xf3
+# CHECK: vcvtm.s32.f32	d1, d30
+0x8a 0xc3 0xbb 0xf3
+# CHECK: vcvtm.u32.f32	d12, d10
+0x64 0x23 0xbb 0xf3
+# CHECK: vcvtm.s32.f32	q1, q10
+0xc2 0xa3 0xfb 0xf3
+# CHECK: vcvtm.u32.f32	q13, q1
+
+0x21 0xf1 0xbb 0xf3
+# CHECK: vcvtn.s32.f32	d15, d17
+0x83 0x51 0xbb 0xf3
+# CHECK: vcvtn.u32.f32	d5, d3
+0x60 0x61 0xbb 0xf3
+# CHECK: vcvtn.s32.f32	q3, q8
+0xc6 0xa1 0xbb 0xf3
+# CHECK: vcvtn.u32.f32	q5, q3
+
+0x25 0xb2 0xbb 0xf3
+# CHECK: vcvtp.s32.f32	d11, d21
+0xa7 0xe2 0xbb 0xf3
+# CHECK: vcvtp.u32.f32	d14, d23
+0x6e 0x82 0xbb 0xf3
+# CHECK: vcvtp.s32.f32	q4, q15
+0xe0 0x22 0xfb 0xf3
+# CHECK: vcvtp.u32.f32	q9, q8
diff --git a/test/MC/Disassembler/ARM/thumb-neon-v8.txt b/test/MC/Disassembler/ARM/thumb-neon-v8.txt
@@ -8,3 +8,39 @@
 # CHECK: vminnm.f32 d5, d4, d30
 0x2a 0xff 0xd4 0x0f
 # CHECK: vminnm.f32 q0, q13, q2
+
+0xbb 0xff 0x06 0x40
+# CHECK: vcvta.s32.f32	d4, d6
+0xbb 0xff 0x8a 0xc0
+# CHECK: vcvta.u32.f32	d12, d10
+0xbb 0xff 0x4c 0x80
+# CHECK: vcvta.s32.f32	q4, q6
+0xbb 0xff 0xe4 0x80
+# CHECK: vcvta.u32.f32	q4, q10
+
+0xbb 0xff 0x2e 0x13
+# CHECK: vcvtm.s32.f32	d1, d30
+0xbb 0xff 0x8a 0xc3
+# CHECK: vcvtm.u32.f32	d12, d10
+0xbb 0xff 0x64 0x23
+# CHECK: vcvtm.s32.f32	q1, q10
+0xfb 0xff 0xc2 0xa3
+# CHECK: vcvtm.u32.f32	q13, q1
+
+0xbb 0xff 0x21 0xf1
+# CHECK: vcvtn.s32.f32	d15, d17
+0xbb 0xff 0x83 0x51
+# CHECK: vcvtn.u32.f32	d5, d3
+0xbb 0xff 0x60 0x61
+# CHECK: vcvtn.s32.f32	q3, q8
+0xbb 0xff 0xc6 0xa1
+# CHECK: vcvtn.u32.f32	q5, q3
+
+0xbb 0xff 0x25 0xb2
+# CHECK: vcvtp.s32.f32	d11, d21
+0xbb 0xff 0xa7 0xe2
+# CHECK: vcvtp.u32.f32	d14, d23
+0xbb 0xff 0x6e 0x82
+# CHECK: vcvtp.s32.f32	q4, q15
+0xfb 0xff 0xe0 0x22
+# CHECK: vcvtp.u32.f32	q9, q8