From b2d24e23cec654746ee38846b1daae05bfa91c4c Mon Sep 17 00:00:00 2001 From: Dmitry Preobrazhensky Date: Fri, 7 Jul 2017 14:29:06 +0000 Subject: [PATCH] [AMDGPU][mc][gfx9] Added support of op_sel/op_sel_hi for V_MAD_MIX* See https://bugs.llvm.org//show_bug.cgi?id=33595 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D35021 llvm-svn: 307402 --- .../Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp | 2 +- llvm/lib/Target/AMDGPU/SIInstrInfo.td | 2 + llvm/lib/Target/AMDGPU/VOP3PInstructions.td | 28 +- llvm/test/MC/AMDGPU/gfx9_asm_all.s | 459 +++++++++++++++++++++ llvm/test/MC/AMDGPU/vop3p-err.s | 41 -- llvm/test/MC/AMDGPU/vop3p.s | 63 ++- 6 files changed, 530 insertions(+), 65 deletions(-) diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp index 7b87560..cacc10e 100644 --- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp +++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp @@ -4284,7 +4284,7 @@ void AMDGPUAsmParser::cvtVOP3P(MCInst &Inst, const OperandVector &Operands) { int ModIdx = AMDGPU::getNamedOperandIdx(Opc, ModOps[J]); - Inst.getOperand(ModIdx).setImm(ModVal); + Inst.getOperand(ModIdx).setImm(Inst.getOperand(ModIdx).getImm() | ModVal); } } diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td index 4a81fb3..ffb0136 100644 --- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td +++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td @@ -1502,6 +1502,8 @@ def VOP_B32_F16_F16 : VOPProfile <[i32, f16, f16, untyped]>; def VOP_V2F16_V2F16_V2F16_V2F16 : VOPProfile <[v2f16, v2f16, v2f16, v2f16]>; def VOP_V2I16_V2I16_V2I16_V2I16 : VOPProfile <[v2i16, v2i16, v2i16, v2i16]>; +def VOP_F32_V2F16_V2F16_V2F16 : VOPProfile <[f32, v2f16, v2f16, v2f16]>; + def VOP_NONE : VOPProfile <[untyped, untyped, untyped, untyped]>; def VOP_F32_F32 : VOPProfile <[f32, f32, untyped, untyped]>; diff --git a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td index 96d3430..f2de1f9 100644 --- a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td +++ b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td @@ -16,12 +16,21 @@ class VOP3PInst !if(P.HasModifiers, getVOP3PModPat.ret, getVOP3Pat.ret) >; -// Non-packed instructions that use the VOP3P encoding. i.e. where -// omod/abs are used. +// Non-packed instructions that use the VOP3P encoding. +// VOP3 neg/abs and VOP3P opsel/opsel_hi modifiers are allowed. class VOP3_VOP3PInst : - VOP3P_Pseudo.ret, getVOP3Pat.ret) ->; + VOP3P_Pseudo { + let InOperandList = + (ins + FP32InputMods:$src0_modifiers, VCSrc_f32:$src0, + FP32InputMods:$src1_modifiers, VCSrc_f32:$src1, + FP32InputMods:$src2_modifiers, VCSrc_f32:$src2, + clampmod:$clamp, + op_sel:$op_sel, + op_sel_hi:$op_sel_hi); + let AsmOperands = + " $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$op_sel$op_sel_hi$clamp"; +} let isCommutable = 1 in { def V_PK_FMA_F16 : VOP3PInst<"v_pk_fma_f16", VOP3_Profile, fma>; @@ -46,9 +55,12 @@ def V_PK_ASHRREV_I16 : VOP3PInst<"v_pk_ashrrev_i16", VOP3_Profile, lshr_rev>; // XXX - Commutable? -def V_MAD_MIX_F32 : VOP3_VOP3PInst<"v_mad_mix_f32", VOP3_Profile>; -def V_MAD_MIXLO_F16 : VOP3_VOP3PInst<"v_mad_mixlo_f16", VOP3_Profile>; -def V_MAD_MIXHI_F16 : VOP3_VOP3PInst<"v_mad_mixhi_f16", VOP3_Profile>; +// These are VOP3a-like opcodes which accept no omod. +// Size of src arguments (16/32) is controlled by op_sel. +// For 16-bit src arguments their location (hi/lo) are controlled by op_sel_hi. +def V_MAD_MIX_F32 : VOP3_VOP3PInst<"v_mad_mix_f32", VOP3_Profile>; +def V_MAD_MIXLO_F16 : VOP3_VOP3PInst<"v_mad_mixlo_f16", VOP3_Profile>; +def V_MAD_MIXHI_F16 : VOP3_VOP3PInst<"v_mad_mixhi_f16", VOP3_Profile>; multiclass VOP3P_Real_vi op> { diff --git a/llvm/test/MC/AMDGPU/gfx9_asm_all.s b/llvm/test/MC/AMDGPU/gfx9_asm_all.s index 0c3dbd2..56484a3 100644 --- a/llvm/test/MC/AMDGPU/gfx9_asm_all.s +++ b/llvm/test/MC/AMDGPU/gfx9_asm_all.s @@ -104933,3 +104933,462 @@ v_cmpx_t_u32_sdwa s[6:7], v1, v2 src0_sel:DWORD src1_sel:WORD_1 v_cmpx_t_u32_sdwa s[6:7], v1, sext(v2) src0_sel:DWORD src1_sel:DWORD // CHECK: [0xf9,0x04,0xbe,0x7d,0x01,0x86,0x06,0x0e] + +v_mad_mix_f32 v5, v1, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v255, v1, v2, v3 +// CHECK: [0xff,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v255, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0xff,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, s1, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, s101, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x65,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, flat_scratch_lo, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x66,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, flat_scratch_hi, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x67,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, vcc_lo, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x6a,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, vcc_hi, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x6b,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, m0, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x7c,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, exec_lo, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x7e,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, exec_hi, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x7f,0x04,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v255, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xff,0x0f,0x1c] + +v_mad_mix_f32 v5, v1, s2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, s101, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xcb,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, flat_scratch_lo, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xcd,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, flat_scratch_hi, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xcf,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, vcc_lo, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xd5,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, vcc_hi, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xd7,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, m0, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xf9,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, exec_lo, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xfd,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, exec_hi, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0xff,0x0c,0x1c] + +v_mad_mix_f32 v5, v1, v2, v255 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0xfe,0x1f] + +v_mad_mix_f32 v5, v1, v2, s3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x18] + +v_mad_mix_f32 v5, v1, v2, s101 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x96,0x19] + +v_mad_mix_f32 v5, v1, v2, flat_scratch_lo +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x9a,0x19] + +v_mad_mix_f32 v5, v1, v2, flat_scratch_hi +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x9e,0x19] + +v_mad_mix_f32 v5, v1, v2, vcc_lo +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0xaa,0x19] + +v_mad_mix_f32 v5, v1, v2, vcc_hi +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0xae,0x19] + +v_mad_mix_f32 v5, v1, v2, m0 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0xf2,0x19] + +v_mad_mix_f32 v5, v1, v2, exec_lo +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0xfa,0x19] + +v_mad_mix_f32 v5, v1, v2, exec_hi +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0xfe,0x19] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel:[0,0,0] +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel:[1,0,0] +// CHECK: [0x05,0x48,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel:[0,1,0] +// CHECK: [0x05,0x50,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel:[0,0,1] +// CHECK: [0x05,0x60,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel:[1,1,1] +// CHECK: [0x05,0x78,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel_hi:[1,1,1] +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel_hi:[0,0,0] +// CHECK: [0x05,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x04] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel_hi:[1,0,0] +// CHECK: [0x05,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x0c] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel_hi:[0,1,0] +// CHECK: [0x05,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x14] + +v_mad_mix_f32 v5, v1, v2, v3 op_sel_hi:[0,0,1] +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x04] + +v_mad_mix_f32 v5, -v1, v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x3c] + +v_mad_mix_f32 v5, v1, -v2, v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x5c] + +v_mad_mix_f32 v5, v1, v2, -v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x9c] + +v_mad_mix_f32 v5, -v1, -v2, -v3 +// CHECK: [0x05,0x40,0xa0,0xd3,0x01,0x05,0x0e,0xfc] + +v_mad_mix_f32 v5, |v1|, v2, v3 +// CHECK: [0x05,0x41,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, |v2|, v3 +// CHECK: [0x05,0x42,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v2, |v3| +// CHECK: [0x05,0x44,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, |v1|, |v2|, |v3| +// CHECK: [0x05,0x47,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v5, v1, v2, v3 clamp +// CHECK: [0x05,0xc0,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v255, v1, v2, v3 +// CHECK: [0xff,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v255, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0xff,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, s1, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, s101, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x65,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, flat_scratch_lo, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x66,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, flat_scratch_hi, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x67,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, vcc_lo, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x6a,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, vcc_hi, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x6b,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, m0, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x7c,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, exec_lo, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x7e,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, exec_hi, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x7f,0x04,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v255, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xff,0x0f,0x1c] + +v_mad_mixhi_f16 v5, v1, s2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, s101, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xcb,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, flat_scratch_lo, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xcd,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, flat_scratch_hi, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xcf,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, vcc_lo, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xd5,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, vcc_hi, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xd7,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, m0, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xf9,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, exec_lo, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xfd,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, exec_hi, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0xff,0x0c,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v255 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0xfe,0x1f] + +v_mad_mixhi_f16 v5, v1, v2, s3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x18] + +v_mad_mixhi_f16 v5, v1, v2, s101 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x96,0x19] + +v_mad_mixhi_f16 v5, v1, v2, flat_scratch_lo +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x9a,0x19] + +v_mad_mixhi_f16 v5, v1, v2, flat_scratch_hi +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x9e,0x19] + +v_mad_mixhi_f16 v5, v1, v2, vcc_lo +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0xaa,0x19] + +v_mad_mixhi_f16 v5, v1, v2, vcc_hi +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0xae,0x19] + +v_mad_mixhi_f16 v5, v1, v2, m0 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0xf2,0x19] + +v_mad_mixhi_f16 v5, v1, v2, exec_lo +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0xfa,0x19] + +v_mad_mixhi_f16 v5, v1, v2, exec_hi +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0xfe,0x19] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel:[0,0,0] +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel:[1,0,0] +// CHECK: [0x05,0x48,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel:[0,1,0] +// CHECK: [0x05,0x50,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel:[0,0,1] +// CHECK: [0x05,0x60,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel:[1,1,1] +// CHECK: [0x05,0x78,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel_hi:[1,1,1] +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel_hi:[0,0,0] +// CHECK: [0x05,0x00,0xa2,0xd3,0x01,0x05,0x0e,0x04] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel_hi:[1,0,0] +// CHECK: [0x05,0x00,0xa2,0xd3,0x01,0x05,0x0e,0x0c] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel_hi:[0,1,0] +// CHECK: [0x05,0x00,0xa2,0xd3,0x01,0x05,0x0e,0x14] + +v_mad_mixhi_f16 v5, v1, v2, v3 op_sel_hi:[0,0,1] +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x04] + +v_mad_mixhi_f16 v5, -v1, v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x3c] + +v_mad_mixhi_f16 v5, v1, -v2, v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x5c] + +v_mad_mixhi_f16 v5, v1, v2, -v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x9c] + +v_mad_mixhi_f16 v5, -v1, -v2, -v3 +// CHECK: [0x05,0x40,0xa2,0xd3,0x01,0x05,0x0e,0xfc] + +v_mad_mixhi_f16 v5, |v1|, v2, v3 +// CHECK: [0x05,0x41,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, |v2|, v3 +// CHECK: [0x05,0x42,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, |v3| +// CHECK: [0x05,0x44,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, |v1|, |v2|, |v3| +// CHECK: [0x05,0x47,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixhi_f16 v5, v1, v2, v3 clamp +// CHECK: [0x05,0xc0,0xa2,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v255, v1, v2, v3 +// CHECK: [0xff,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v255, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0xff,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, s1, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, s101, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x65,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, flat_scratch_lo, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x66,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, flat_scratch_hi, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x67,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, vcc_lo, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x6a,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, vcc_hi, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x6b,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, m0, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x7c,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, exec_lo, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x7e,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, exec_hi, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x7f,0x04,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v255, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xff,0x0f,0x1c] + +v_mad_mixlo_f16 v5, v1, s2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, s101, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xcb,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, flat_scratch_lo, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xcd,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, flat_scratch_hi, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xcf,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, vcc_lo, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xd5,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, vcc_hi, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xd7,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, m0, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xf9,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, exec_lo, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xfd,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, exec_hi, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0xff,0x0c,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v255 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0xfe,0x1f] + +v_mad_mixlo_f16 v5, v1, v2, s3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x18] + +v_mad_mixlo_f16 v5, v1, v2, s101 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x96,0x19] + +v_mad_mixlo_f16 v5, v1, v2, flat_scratch_lo +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x9a,0x19] + +v_mad_mixlo_f16 v5, v1, v2, flat_scratch_hi +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x9e,0x19] + +v_mad_mixlo_f16 v5, v1, v2, vcc_lo +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0xaa,0x19] + +v_mad_mixlo_f16 v5, v1, v2, vcc_hi +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0xae,0x19] + +v_mad_mixlo_f16 v5, v1, v2, m0 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0xf2,0x19] + +v_mad_mixlo_f16 v5, v1, v2, exec_lo +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0xfa,0x19] + +v_mad_mixlo_f16 v5, v1, v2, exec_hi +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0xfe,0x19] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel:[0,0,0] +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel:[1,0,0] +// CHECK: [0x05,0x48,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel:[0,1,0] +// CHECK: [0x05,0x50,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel:[0,0,1] +// CHECK: [0x05,0x60,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel:[1,1,1] +// CHECK: [0x05,0x78,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel_hi:[1,1,1] +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel_hi:[0,0,0] +// CHECK: [0x05,0x00,0xa1,0xd3,0x01,0x05,0x0e,0x04] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel_hi:[1,0,0] +// CHECK: [0x05,0x00,0xa1,0xd3,0x01,0x05,0x0e,0x0c] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel_hi:[0,1,0] +// CHECK: [0x05,0x00,0xa1,0xd3,0x01,0x05,0x0e,0x14] + +v_mad_mixlo_f16 v5, v1, v2, v3 op_sel_hi:[0,0,1] +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x04] + +v_mad_mixlo_f16 v5, -v1, v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x3c] + +v_mad_mixlo_f16 v5, v1, -v2, v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x5c] + +v_mad_mixlo_f16 v5, v1, v2, -v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x9c] + +v_mad_mixlo_f16 v5, -v1, -v2, -v3 +// CHECK: [0x05,0x40,0xa1,0xd3,0x01,0x05,0x0e,0xfc] + +v_mad_mixlo_f16 v5, |v1|, v2, v3 +// CHECK: [0x05,0x41,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, |v2|, v3 +// CHECK: [0x05,0x42,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, |v3| +// CHECK: [0x05,0x44,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, |v1|, |v2|, |v3| +// CHECK: [0x05,0x47,0xa1,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mixlo_f16 v5, v1, v2, v3 clamp +// CHECK: [0x05,0xc0,0xa1,0xd3,0x01,0x05,0x0e,0x1c] diff --git a/llvm/test/MC/AMDGPU/vop3p-err.s b/llvm/test/MC/AMDGPU/vop3p-err.s index f4b1a3d..bc6f610 100644 --- a/llvm/test/MC/AMDGPU/vop3p-err.s +++ b/llvm/test/MC/AMDGPU/vop3p-err.s @@ -71,47 +71,6 @@ v_pk_add_u16 v1, abs(v2), v3 // GFX9: :19: error: invalid operand for instruction v_pk_add_u16 v1, -v2, v3 - -// -// Packed operands on the non-packed VOP3P instructions -// - -// GFX9: invalid operand for instruction -v_mad_mix_f32 v1, v2, v3, v4 op_sel:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mix_f32 v1, v2, v3, v4 op_sel_hi:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mix_f32 v1, v2, v3, v4 neg_lo:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mix_f32 v1, v2, v3, v4 neg_hi:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mixlo_f16 v1, v2, v3, v4 op_sel:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mixlo_f16 v1, v2, v3, v4 op_sel_hi:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mixlo_f16 v1, v2, v3, v4 neg_lo:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mixlo_f16 v1, v2, v3, v4 neg_hi:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mixhi_f16 v1, v2, v3, v4 op_sel:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mixhi_f16 v1, v2, v3, v4 op_sel_hi:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mixhi_f16 v1, v2, v3, v4 neg_lo:[0,0,0] - -// GFX9: invalid operand for instruction -v_mad_mixhi_f16 v1, v2, v3, v4 neg_hi:[0,0,0] - // // Constant bus restrictions // diff --git a/llvm/test/MC/AMDGPU/vop3p.s b/llvm/test/MC/AMDGPU/vop3p.s index c9eda69..97c3650 100644 --- a/llvm/test/MC/AMDGPU/vop3p.s +++ b/llvm/test/MC/AMDGPU/vop3p.s @@ -169,48 +169,81 @@ v_pk_max_f16 v0, v1, v2 // GFX9: v_pk_max_f16 v0, v1, v2 ; encoding: [0x00,0x00,0x92,0xd3,0x01,0x05,0x02,0x18] v_mad_mix_f32 v0, v1, v2, v3 -// GFX9: v_mad_mix_f32 v0, v1, v2, v3 ; encoding: [0x00,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x04] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x1c] v_mad_mixlo_f16 v0, v1, v2, v3 -// GFX9: v_mad_mixlo_f16 v0, v1, v2, v3 ; encoding: [0x00,0x00,0xa1,0xd3,0x01,0x05,0x0e,0x04] +// GFX9: v_mad_mixlo_f16 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa1,0xd3,0x01,0x05,0x0e,0x1c] v_mad_mixhi_f16 v0, v1, v2, v3 -// GFX9: v_mad_mixhi_f16 v0, v1, v2, v3 ; encoding: [0x00,0x00,0xa2,0xd3,0x01,0x05,0x0e,0x04] - +// GFX9: v_mad_mixhi_f16 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa2,0xd3,0x01,0x05,0x0e,0x1c] // // Regular source modifiers on non-packed instructions // v_mad_mix_f32 v0, abs(v1), v2, v3 -// GFX9: v_mad_mix_f32 v0, |v1|, v2, v3 ; encoding: [0x00,0x01,0xa0,0xd3,0x01,0x05,0x0e,0x04] +// GFX9: v_mad_mix_f32 v0, |v1|, v2, v3 ; encoding: [0x00,0x41,0xa0,0xd3,0x01,0x05,0x0e,0x1c] v_mad_mix_f32 v0, v1, abs(v2), v3 -// GFX9: v_mad_mix_f32 v0, v1, |v2|, v3 ; encoding: [0x00,0x02,0xa0,0xd3,0x01,0x05,0x0e,0x04] +// GFX9: v_mad_mix_f32 v0, v1, |v2|, v3 ; encoding: [0x00,0x42,0xa0,0xd3,0x01,0x05,0x0e,0x1c] v_mad_mix_f32 v0, v1, v2, abs(v3) -// GFX9: v_mad_mix_f32 v0, v1, v2, |v3| ; encoding: [0x00,0x04,0xa0,0xd3,0x01,0x05,0x0e,0x04] +// GFX9: v_mad_mix_f32 v0, v1, v2, |v3| ; encoding: [0x00,0x44,0xa0,0xd3,0x01,0x05,0x0e,0x1c] v_mad_mix_f32 v0, -v1, v2, v3 -// GFX9: v_mad_mix_f32 v0, -v1, v2, v3 ; encoding: [0x00,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x24] +// GFX9: v_mad_mix_f32 v0, -v1, v2, v3 ; encoding: [0x00,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x3c] v_mad_mix_f32 v0, v1, -v2, v3 -// GFX9: v_mad_mix_f32 v0, v1, -v2, v3 ; encoding: [0x00,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x44] +// GFX9: v_mad_mix_f32 v0, v1, -v2, v3 ; encoding: [0x00,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x5c] v_mad_mix_f32 v0, v1, v2, -v3 -// GFX9: v_mad_mix_f32 v0, v1, v2, -v3 ; encoding: [0x00,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x84] +// GFX9: v_mad_mix_f32 v0, v1, v2, -v3 ; encoding: [0x00,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x9c] v_mad_mix_f32 v0, -abs(v1), v2, v3 -// GFX9: v_mad_mix_f32 v0, -|v1|, v2, v3 ; encoding: [0x00,0x01,0xa0,0xd3,0x01,0x05,0x0e,0x24] +// GFX9: v_mad_mix_f32 v0, -|v1|, v2, v3 ; encoding: [0x00,0x41,0xa0,0xd3,0x01,0x05,0x0e,0x3c] v_mad_mix_f32 v0, v1, -abs(v2), v3 -// GFX9: v_mad_mix_f32 v0, v1, -|v2|, v3 ; encoding: [0x00,0x02,0xa0,0xd3,0x01,0x05,0x0e,0x44] +// GFX9: v_mad_mix_f32 v0, v1, -|v2|, v3 ; encoding: [0x00,0x42,0xa0,0xd3,0x01,0x05,0x0e,0x5c] v_mad_mix_f32 v0, v1, v2, -abs(v3) -// GFX9: v_mad_mix_f32 v0, v1, v2, -|v3| ; encoding: [0x00,0x04,0xa0,0xd3,0x01,0x05,0x0e,0x84] +// GFX9: v_mad_mix_f32 v0, v1, v2, -|v3| ; encoding: [0x00,0x44,0xa0,0xd3,0x01,0x05,0x0e,0x9c] v_mad_mixlo_f16 v0, abs(v1), -v2, abs(v3) -// GFX9: v_mad_mixlo_f16 v0, |v1|, -v2, |v3| ; encoding: [0x00,0x05,0xa1,0xd3,0x01,0x05,0x0e,0x44] +// GFX9: v_mad_mixlo_f16 v0, |v1|, -v2, |v3| ; encoding: [0x00,0x45,0xa1,0xd3,0x01,0x05,0x0e,0x5c] v_mad_mixhi_f16 v0, -v1, abs(v2), -abs(v3) -// GFX9: v_mad_mixhi_f16 v0, -v1, |v2|, -|v3| ; encoding: [0x00,0x06,0xa2,0xd3,0x01,0x05,0x0e,0xa4] +// GFX9: v_mad_mixhi_f16 v0, -v1, |v2|, -|v3| ; encoding: [0x00,0x46,0xa2,0xd3,0x01,0x05,0x0e,0xbc] + +// +// op_sel with non-packed instructions +// + +v_mad_mix_f32 v0, v1, v2, v3 op_sel:[0,0,0] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel:[1,0,0] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 op_sel:[1,0,0] ; encoding: [0x00,0x48,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel:[0,1,0] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 op_sel:[0,1,0] ; encoding: [0x00,0x50,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel:[0,0,1] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 op_sel:[0,0,1] ; encoding: [0x00,0x60,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel:[1,1,1] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 op_sel:[1,1,1] ; encoding: [0x00,0x78,0xa0,0xd3,0x01,0x05,0x0e,0x1c] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[0,0,0] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x04] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[1,0,0] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x0c] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[0,1,0] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xa0,0xd3,0x01,0x05,0x0e,0x14] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[0,0,1] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[0,0,1] ; encoding: [0x00,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x04] + +v_mad_mix_f32 v0, v1, v2, v3 op_sel_hi:[1,1,1] +// GFX9: v_mad_mix_f32 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa0,0xd3,0x01,0x05,0x0e,0x1c] -- 2.7.4