X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86InstrFMA.td;h=fd800cf077f7bb08acee5475fb73fec3701fd3f5;hb=199a1320b792c866ee6546dcbd157bebf54edaca;hp=206f7b600f3c6474fc09cbc03b061acfaf34c695;hpb=d8f4348cab0f6ac58d62746aceb822b7d615807e;p=oota-llvm.git diff --git a/lib/Target/X86/X86InstrFMA.td b/lib/Target/X86/X86InstrFMA.td index 206f7b600f3..fd800cf077f 100644 --- a/lib/Target/X86/X86InstrFMA.td +++ b/lib/Target/X86/X86InstrFMA.td @@ -15,12 +15,31 @@ // FMA3 - Intel 3 operand Fused Multiply-Add instructions //===----------------------------------------------------------------------===// -let Constraints = "$src1 = $dst" in { +// For all FMA opcodes declared in fma3p_rm and fma3s_rm milticlasses defined +// below, both the register and memory variants are commutable. +// For the register form the commutable operands are 1, 2 and 3. +// For the memory variant the folded operand must be in 3. Thus, +// in that case, only the operands 1 and 2 can be swapped. +// Commuting some of operands may require the opcode change. +// FMA*213*: +// operands 1 and 2 (memory & register forms): *213* --> *213*(no changes); +// operands 1 and 3 (register forms only): *213* --> *231*; +// operands 2 and 3 (register forms only): *213* --> *132*. +// FMA*132*: +// operands 1 and 2 (memory & register forms): *132* --> *231*; +// operands 1 and 3 (register forms only): *132* --> *132*(no changes); +// operands 2 and 3 (register forms only): *132* --> *213*. +// FMA*231*: +// operands 1 and 2 (memory & register forms): *231* --> *132*; +// operands 1 and 3 (register forms only): *231* --> *213*; +// operands 2 and 3 (register forms only): *231* --> *231*(no changes). + +let Constraints = "$src1 = $dst", hasSideEffects = 0, isCommutable = 1 in multiclass fma3p_rm opc, string OpcodeStr, PatFrag MemFrag128, PatFrag MemFrag256, ValueType OpVT128, ValueType OpVT256, SDPatternOperator Op = null_frag> { - let isCommutable = 1, usesCustomInserter = 1 in + let usesCustomInserter = 1 in def r : FMA3 opc, string OpcodeStr, [(set VR128:$dst, (OpVT128 (Op VR128:$src2, VR128:$src1, (MemFrag128 addr:$src3))))]>; - let isCommutable = 1, usesCustomInserter = 1 in + let usesCustomInserter = 1 in def rY : FMA3 opc, string OpcodeStr, (OpVT256 (Op VR256:$src2, VR256:$src1, (MemFrag256 addr:$src3))))]>, VEX_L; } -} // Constraints = "$src1 = $dst" multiclass fma3p_forms opc132, bits<8> opc213, bits<8> opc231, string OpcodeStr, string PackTy, @@ -62,14 +80,12 @@ multiclass fma3p_forms opc132, bits<8> opc213, bits<8> opc231, defm r213 : fma3p_rm; -let neverHasSideEffects = 1 in { defm r132 : fma3p_rm; defm r231 : fma3p_rm; -} // neverHasSideEffects = 1 } // Fused Multiply-Add @@ -114,73 +130,122 @@ let ExeDomain = SSEPackedDouble in { v4f64>, VEX_W; } -let Constraints = "$src1 = $dst" in { -multiclass fma3s_rm opc, string OpcodeStr, X86MemOperand x86memop, - RegisterClass RC, ValueType OpVT, PatFrag mem_frag, +// All source register operands of FMA opcodes defined in fma3s_rm multiclass +// can be commuted. In many cases such commute transformation requres an opcode +// adjustment, for example, commuting the operands 1 and 2 in FMA*132 form +// would require an opcode change to FMA*231: +// FMA*132* reg1, reg2, reg3; // reg1 * reg3 + reg2; +// --> +// FMA*231* reg2, reg1, reg3; // reg1 * reg3 + reg2; +// Please see more detailed comment at the very beginning of the section +// defining FMA3 opcodes above. +let Constraints = "$src1 = $dst", isCommutable = 1, hasSideEffects = 0 in +multiclass fma3s_rm opc, string OpcodeStr, + X86MemOperand x86memop, RegisterClass RC, SDPatternOperator OpNode = null_frag> { - let isCommutable = 1, usesCustomInserter = 1 in + let usesCustomInserter = 1 in def r : FMA3; + [(set RC:$dst, (OpNode RC:$src2, RC:$src1, RC:$src3))]>; + let mayLoad = 1 in def m : FMA3; + (OpNode RC:$src2, RC:$src1, (load addr:$src3)))]>; } -multiclass fma3s_rm_int opc, string OpcodeStr, Operand memop, - ComplexPattern mem_cpat, Intrinsic IntId, - RegisterClass RC> { - let isCodeGenOnly = 1 in { - let isCommutable = 1 in - def r_Int : FMA3; - def m_Int : FMA3; - } // isCodeGenOnly +// These FMA*_Int instructions are defined specially for being used when +// the scalar FMA intrinsics are lowered to machine instructions, and in that +// sense, they are similar to existing ADD*_Int, SUB*_Int, MUL*_Int, etc. +// instructions. +// +// All of the FMA*_Int opcodes are defined as commutable here. +// Commuting the 2nd and 3rd source register operands of FMAs is quite trivial +// and the corresponding optimizations have been developed. +// Commuting the 1st operand of FMA*_Int requires some additional analysis, +// the commute optimization is legal only if all users of FMA*_Int use only +// the lowest element of the FMA*_Int instruction. Even though such analysis +// may be not implemented yet we allow the routines doing the actual commute +// transformation to decide if one or another instruction is commutable or not. +let Constraints = "$src1 = $dst", isCommutable = 1, isCodeGenOnly = 1, + hasSideEffects = 0 in +multiclass fma3s_rm_int opc, string OpcodeStr, + Operand memopr, RegisterClass RC> { + def r_Int : FMA3; + + let mayLoad = 1 in + def m_Int : FMA3; } -} // Constraints = "$src1 = $dst" multiclass fma3s_forms opc132, bits<8> opc213, bits<8> opc231, - string OpStr, string PackTy, Intrinsic Int, - SDNode OpNode, RegisterClass RC, ValueType OpVT, - X86MemOperand x86memop, Operand memop, PatFrag mem_frag, - ComplexPattern mem_cpat> { -let neverHasSideEffects = 1 in { - defm r132 : fma3s_rm; - defm r231 : fma3s_rm; + string OpStr, string PackTy, + SDNode OpNode, RegisterClass RC, + X86MemOperand x86memop> { + defm r132 : fma3s_rm; + defm r213 : fma3s_rm; + defm r231 : fma3s_rm; } -defm r213 : fma3s_rm, - fma3s_rm_int; +// The FMA 213 form is created for lowering of scalar FMA intrinscis +// to machine instructions. +// The FMA 132 form can trivially be get by commuting the 2nd and 3rd operands +// of FMA 213 form. +// The FMA 231 form can be get only by commuting the 1st operand of 213 or 132 +// forms and is possible only after special analysis of all uses of the initial +// instruction. Such analysis do not exist yet and thus introducing the 231 +// form of FMA*_Int instructions is done using an optimistic assumption that +// such analysis will be implemented eventually. +multiclass fma3s_int_forms opc132, bits<8> opc213, bits<8> opc231, + string OpStr, string PackTy, + RegisterClass RC, Operand memop> { + defm r132 : fma3s_rm_int; + defm r213 : fma3s_rm_int; + defm r231 : fma3s_rm_int; } multiclass fma3s opc132, bits<8> opc213, bits<8> opc231, string OpStr, Intrinsic IntF32, Intrinsic IntF64, SDNode OpNode> { - defm SS : fma3s_forms; - defm SD : fma3s_forms, VEX_W; + let ExeDomain = SSEPackedSingle in + defm SS : fma3s_forms, + fma3s_int_forms; + + let ExeDomain = SSEPackedDouble in + defm SD : fma3s_forms, + fma3s_int_forms, + VEX_W; + + // These patterns use the 123 ordering, instead of 213, even though + // they match the intrinsic to the 213 version of the instruction. + // This is because src1 is tied to dest, and the scalar intrinsics + // require the pass-through values to come from the first source + // operand, not the second. + def : Pat<(IntF32 VR128:$src1, VR128:$src2, VR128:$src3), + (COPY_TO_REGCLASS(!cast(NAME#"SSr213r_Int") + $src1, $src2, $src3), VR128)>; + + def : Pat<(IntF64 VR128:$src1, VR128:$src2, VR128:$src3), + (COPY_TO_REGCLASS(!cast(NAME#"SDr213r_Int") + $src1, $src2, $src3), VR128)>; } defm VFMADD : fma3s<0x99, 0xA9, 0xB9, "vfmadd", int_x86_fma_vfmadd_ss, @@ -312,36 +377,23 @@ let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in { } // isCodeGenOnly = 1 } -defm VFMADDSS4 : fma4s<0x6A, "vfmaddss", FR32, f32mem, f32, X86Fmadd, loadf32>, - fma4s_int<0x6A, "vfmaddss", ssmem, sse_load_f32, - int_x86_fma_vfmadd_ss>; -defm VFMADDSD4 : fma4s<0x6B, "vfmaddsd", FR64, f64mem, f64, X86Fmadd, loadf64>, - fma4s_int<0x6B, "vfmaddsd", sdmem, sse_load_f64, - int_x86_fma_vfmadd_sd>; -defm VFMSUBSS4 : fma4s<0x6E, "vfmsubss", FR32, f32mem, f32, X86Fmsub, loadf32>, - fma4s_int<0x6E, "vfmsubss", ssmem, sse_load_f32, - int_x86_fma_vfmsub_ss>; -defm VFMSUBSD4 : fma4s<0x6F, "vfmsubsd", FR64, f64mem, f64, X86Fmsub, loadf64>, - fma4s_int<0x6F, "vfmsubsd", sdmem, sse_load_f64, - int_x86_fma_vfmsub_sd>; -defm VFNMADDSS4 : fma4s<0x7A, "vfnmaddss", FR32, f32mem, f32, - X86Fnmadd, loadf32>, - fma4s_int<0x7A, "vfnmaddss", ssmem, sse_load_f32, - int_x86_fma_vfnmadd_ss>; -defm VFNMADDSD4 : fma4s<0x7B, "vfnmaddsd", FR64, f64mem, f64, - X86Fnmadd, loadf64>, - fma4s_int<0x7B, "vfnmaddsd", sdmem, sse_load_f64, - int_x86_fma_vfnmadd_sd>; -defm VFNMSUBSS4 : fma4s<0x7E, "vfnmsubss", FR32, f32mem, f32, - X86Fnmsub, loadf32>, - fma4s_int<0x7E, "vfnmsubss", ssmem, sse_load_f32, - int_x86_fma_vfnmsub_ss>; -defm VFNMSUBSD4 : fma4s<0x7F, "vfnmsubsd", FR64, f64mem, f64, - X86Fnmsub, loadf64>, - fma4s_int<0x7F, "vfnmsubsd", sdmem, sse_load_f64, - int_x86_fma_vfnmsub_sd>; - let ExeDomain = SSEPackedSingle in { + // Scalar Instructions + defm VFMADDSS4 : fma4s<0x6A, "vfmaddss", FR32, f32mem, f32, X86Fmadd, loadf32>, + fma4s_int<0x6A, "vfmaddss", ssmem, sse_load_f32, + int_x86_fma_vfmadd_ss>; + defm VFMSUBSS4 : fma4s<0x6E, "vfmsubss", FR32, f32mem, f32, X86Fmsub, loadf32>, + fma4s_int<0x6E, "vfmsubss", ssmem, sse_load_f32, + int_x86_fma_vfmsub_ss>; + defm VFNMADDSS4 : fma4s<0x7A, "vfnmaddss", FR32, f32mem, f32, + X86Fnmadd, loadf32>, + fma4s_int<0x7A, "vfnmaddss", ssmem, sse_load_f32, + int_x86_fma_vfnmadd_ss>; + defm VFNMSUBSS4 : fma4s<0x7E, "vfnmsubss", FR32, f32mem, f32, + X86Fnmsub, loadf32>, + fma4s_int<0x7E, "vfnmsubss", ssmem, sse_load_f32, + int_x86_fma_vfnmsub_ss>; + // Packed Instructions defm VFMADDPS4 : fma4p<0x68, "vfmaddps", X86Fmadd, v4f32, v8f32, loadv4f32, loadv8f32>; defm VFMSUBPS4 : fma4p<0x6C, "vfmsubps", X86Fmsub, v4f32, v8f32, @@ -357,6 +409,22 @@ let ExeDomain = SSEPackedSingle in { } let ExeDomain = SSEPackedDouble in { + // Scalar Instructions + defm VFMADDSD4 : fma4s<0x6B, "vfmaddsd", FR64, f64mem, f64, X86Fmadd, loadf64>, + fma4s_int<0x6B, "vfmaddsd", sdmem, sse_load_f64, + int_x86_fma_vfmadd_sd>; + defm VFMSUBSD4 : fma4s<0x6F, "vfmsubsd", FR64, f64mem, f64, X86Fmsub, loadf64>, + fma4s_int<0x6F, "vfmsubsd", sdmem, sse_load_f64, + int_x86_fma_vfmsub_sd>; + defm VFNMADDSD4 : fma4s<0x7B, "vfnmaddsd", FR64, f64mem, f64, + X86Fnmadd, loadf64>, + fma4s_int<0x7B, "vfnmaddsd", sdmem, sse_load_f64, + int_x86_fma_vfnmadd_sd>; + defm VFNMSUBSD4 : fma4s<0x7F, "vfnmsubsd", FR64, f64mem, f64, + X86Fnmsub, loadf64>, + fma4s_int<0x7F, "vfnmsubsd", sdmem, sse_load_f64, + int_x86_fma_vfnmsub_sd>; + // Packed Instructions defm VFMADDPD4 : fma4p<0x69, "vfmaddpd", X86Fmadd, v2f64, v4f64, loadv2f64, loadv4f64>; defm VFMSUBPD4 : fma4p<0x6D, "vfmsubpd", X86Fmsub, v2f64, v4f64,