From 8f41dd923e6872a2fe1462a95de3ddcfe6b701c2 Mon Sep 17 00:00:00 2001 From: Akira Hatanaka Date: Fri, 12 Apr 2013 22:14:24 +0000 Subject: [PATCH] [mips] v4i8 and v2i16 add, sub and mul instruction selection patterns. llvm-svn: 179420 --- llvm/lib/Target/Mips/MipsDSPInstrInfo.td | 26 +++++-- llvm/lib/Target/Mips/MipsSEISelLowering.cpp | 5 ++ llvm/test/CodeGen/Mips/dsp-patterns.ll | 102 ++++++++++++++++++++++++++-- 3 files changed, 121 insertions(+), 12 deletions(-) diff --git a/llvm/lib/Target/Mips/MipsDSPInstrInfo.td b/llvm/lib/Target/Mips/MipsDSPInstrInfo.td index 3c116e1..2f8bd5d 100644 --- a/llvm/lib/Target/Mips/MipsDSPInstrInfo.td +++ b/llvm/lib/Target/Mips/MipsDSPInstrInfo.td @@ -515,27 +515,27 @@ class INSV_DESC_BASE, IsCommutable; class ADDU_S_QB_DESC : ADDU_QB_DESC_BASE<"addu_s.qb", int_mips_addu_s_qb, NoItinerary, DSPRegs, DSPRegs>, IsCommutable; -class SUBU_QB_DESC : ADDU_QB_DESC_BASE<"subu.qb", int_mips_subu_qb, NoItinerary, +class SUBU_QB_DESC : ADDU_QB_DESC_BASE<"subu.qb", null_frag, NoItinerary, DSPRegs, DSPRegs>; class SUBU_S_QB_DESC : ADDU_QB_DESC_BASE<"subu_s.qb", int_mips_subu_s_qb, NoItinerary, DSPRegs, DSPRegs>; -class ADDQ_PH_DESC : ADDU_QB_DESC_BASE<"addq.ph", int_mips_addq_ph, NoItinerary, +class ADDQ_PH_DESC : ADDU_QB_DESC_BASE<"addq.ph", null_frag, NoItinerary, DSPRegs, DSPRegs>, IsCommutable; class ADDQ_S_PH_DESC : ADDU_QB_DESC_BASE<"addq_s.ph", int_mips_addq_s_ph, NoItinerary, DSPRegs, DSPRegs>, IsCommutable; -class SUBQ_PH_DESC : ADDU_QB_DESC_BASE<"subq.ph", int_mips_subq_ph, NoItinerary, +class SUBQ_PH_DESC : ADDU_QB_DESC_BASE<"subq.ph", null_frag, NoItinerary, DSPRegs, DSPRegs>; class SUBQ_S_PH_DESC : ADDU_QB_DESC_BASE<"subq_s.ph", int_mips_subq_s_ph, @@ -951,7 +951,7 @@ class ABSQ_S_QB_DESC : ABSQ_S_PH_R2_DESC_BASE<"absq_s.qb", int_mips_absq_s_qb, NoItinerary, DSPRegs>; // Multiplication -class MUL_PH_DESC : ADDUH_QB_DESC_BASE<"mul.ph", int_mips_mul_ph, NoItinerary, +class MUL_PH_DESC : ADDUH_QB_DESC_BASE<"mul.ph", null_frag, NoItinerary, DSPRegs>, IsCommutable; class MUL_S_PH_DESC : ADDUH_QB_DESC_BASE<"mul_s.ph", int_mips_mul_s_ph, @@ -1232,6 +1232,22 @@ def : DSPPat<(store (v2i16 DSPRegs:$val), addr:$a), def : DSPPat<(store (v4i8 DSPRegs:$val), addr:$a), (SW (COPY_TO_REGCLASS DSPRegs:$val, CPURegs), addr:$a)>; +// Binary operations. +class DSPBinPat : + DSPPat<(Node ValTy:$a, ValTy:$b), (Inst ValTy:$a, ValTy:$b), Pred>; + +def : DSPBinPat; +def : DSPBinPat; +def : DSPBinPat; +def : DSPBinPat; +def : DSPBinPat; +def : DSPBinPat; +def : DSPBinPat; +def : DSPBinPat; +def : DSPBinPat; +def : DSPBinPat; + // Extr patterns. class EXTR_W_TY1_R2_Pat : DSPPat<(i32 (OpNode CPURegs:$rs, ACRegsDSP:$ac)), diff --git a/llvm/lib/Target/Mips/MipsSEISelLowering.cpp b/llvm/lib/Target/Mips/MipsSEISelLowering.cpp index 805228b..70ba985 100644 --- a/llvm/lib/Target/Mips/MipsSEISelLowering.cpp +++ b/llvm/lib/Target/Mips/MipsSEISelLowering.cpp @@ -45,12 +45,17 @@ MipsSETargetLowering::MipsSETargetLowering(MipsTargetMachine &TM) for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc) setOperationAction(Opc, VecTys[i], Expand); + setOperationAction(ISD::ADD, VecTys[i], Legal); + setOperationAction(ISD::SUB, VecTys[i], Legal); setOperationAction(ISD::LOAD, VecTys[i], Legal); setOperationAction(ISD::STORE, VecTys[i], Legal); setOperationAction(ISD::BITCAST, VecTys[i], Legal); } } + if (Subtarget->hasDSPR2()) + setOperationAction(ISD::MUL, MVT::v2i16, Legal); + if (!TM.Options.UseSoftFloat) { addRegisterClass(MVT::f32, &Mips::FGR32RegClass); diff --git a/llvm/test/CodeGen/Mips/dsp-patterns.ll b/llvm/test/CodeGen/Mips/dsp-patterns.ll index 0752f69..4feba11029 100644 --- a/llvm/test/CodeGen/Mips/dsp-patterns.ll +++ b/llvm/test/CodeGen/Mips/dsp-patterns.ll @@ -1,7 +1,8 @@ -; RUN: llc -march=mips -mattr=dsp < %s | FileCheck %s +; RUN: llc -march=mips -mattr=dsp < %s | FileCheck %s -check-prefix=R1 +; RUN: llc -march=mips -mattr=dspr2 < %s | FileCheck %s -check-prefix=R2 -; CHECK: test_lbux: -; CHECK: lbux ${{[0-9]+}} +; R1: test_lbux: +; R1: lbux ${{[0-9]+}} define zeroext i8 @test_lbux(i8* nocapture %b, i32 %i) { entry: @@ -10,8 +11,8 @@ entry: ret i8 %0 } -; CHECK: test_lhx: -; CHECK: lhx ${{[0-9]+}} +; R1: test_lhx: +; R1: lhx ${{[0-9]+}} define signext i16 @test_lhx(i16* nocapture %b, i32 %i) { entry: @@ -20,8 +21,8 @@ entry: ret i16 %0 } -; CHECK: test_lwx: -; CHECK: lwx ${{[0-9]+}} +; R1: test_lwx: +; R1: lwx ${{[0-9]+}} define i32 @test_lwx(i32* nocapture %b, i32 %i) { entry: @@ -29,3 +30,90 @@ entry: %0 = load i32* %add.ptr, align 4 ret i32 %0 } + +; R1: test_add_v2q15_: +; R1: addq.ph ${{[0-9]+}} + +define { i32 } @test_add_v2q15_(i32 %a.coerce, i32 %b.coerce) { +entry: + %0 = bitcast i32 %a.coerce to <2 x i16> + %1 = bitcast i32 %b.coerce to <2 x i16> + %add = add <2 x i16> %0, %1 + %2 = bitcast <2 x i16> %add to i32 + %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0 + ret { i32 } %.fca.0.insert +} + +; R1: test_sub_v2q15_: +; R1: subq.ph ${{[0-9]+}} + +define { i32 } @test_sub_v2q15_(i32 %a.coerce, i32 %b.coerce) { +entry: + %0 = bitcast i32 %a.coerce to <2 x i16> + %1 = bitcast i32 %b.coerce to <2 x i16> + %sub = sub <2 x i16> %0, %1 + %2 = bitcast <2 x i16> %sub to i32 + %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0 + ret { i32 } %.fca.0.insert +} + +; R2: test_mul_v2q15_: +; R2: mul.ph ${{[0-9]+}} + +; mul.ph is an R2 instruction. Check that multiply node gets expanded. +; R1: test_mul_v2q15_: +; R1: mul ${{[0-9]+}} +; R1: mul ${{[0-9]+}} + +define { i32 } @test_mul_v2q15_(i32 %a.coerce, i32 %b.coerce) { +entry: + %0 = bitcast i32 %a.coerce to <2 x i16> + %1 = bitcast i32 %b.coerce to <2 x i16> + %mul = mul <2 x i16> %0, %1 + %2 = bitcast <2 x i16> %mul to i32 + %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0 + ret { i32 } %.fca.0.insert +} + +; R1: test_add_v4i8_: +; R1: addu.qb ${{[0-9]+}} + +define { i32 } @test_add_v4i8_(i32 %a.coerce, i32 %b.coerce) { +entry: + %0 = bitcast i32 %a.coerce to <4 x i8> + %1 = bitcast i32 %b.coerce to <4 x i8> + %add = add <4 x i8> %0, %1 + %2 = bitcast <4 x i8> %add to i32 + %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0 + ret { i32 } %.fca.0.insert +} + +; R1: test_sub_v4i8_: +; R1: subu.qb ${{[0-9]+}} + +define { i32 } @test_sub_v4i8_(i32 %a.coerce, i32 %b.coerce) { +entry: + %0 = bitcast i32 %a.coerce to <4 x i8> + %1 = bitcast i32 %b.coerce to <4 x i8> + %sub = sub <4 x i8> %0, %1 + %2 = bitcast <4 x i8> %sub to i32 + %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0 + ret { i32 } %.fca.0.insert +} + +; DSP-ASE doesn't have a v4i8 multiply instruction. Check that multiply node gets expanded. +; R2: test_mul_v4i8_: +; R2: mul ${{[0-9]+}} +; R2: mul ${{[0-9]+}} +; R2: mul ${{[0-9]+}} +; R2: mul ${{[0-9]+}} + +define { i32 } @test_mul_v4i8_(i32 %a.coerce, i32 %b.coerce) { +entry: + %0 = bitcast i32 %a.coerce to <4 x i8> + %1 = bitcast i32 %b.coerce to <4 x i8> + %mul = mul <4 x i8> %0, %1 + %2 = bitcast <4 x i8> %mul to i32 + %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0 + ret { i32 } %.fca.0.insert +} -- 2.7.4