X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86InstrSSE.td;h=52da79aab026f0a8eef8acbce4eaa883fafed2cb;hb=d90219463154cafb5d626b7964bf20e572a186df;hp=edb7f6ccac4f3eb72f18ae2a08eb80eaa22e6338;hpb=cc915951eb6f10002665879fae62d05c41d01d33;p=oota-llvm.git diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td index edb7f6ccac4..52da79aab02 100644 --- a/lib/Target/X86/X86InstrSSE.td +++ b/lib/Target/X86/X86InstrSSE.td @@ -382,11 +382,11 @@ let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1, // We set canFoldAsLoad because this can be converted to a constant-pool // load of an all-zeros value if folding it would be beneficial. let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1, - isPseudo = 1, neverHasSideEffects = 1 in { -def V_SET0 : I<0, Pseudo, (outs VR128:$dst), (ins), "", []>; + isPseudo = 1 in { +def V_SET0 : I<0, Pseudo, (outs VR128:$dst), (ins), "", + [(set VR128:$dst, (v4f32 immAllZerosV))]>; } -def : Pat<(v4f32 immAllZerosV), (V_SET0)>; def : Pat<(v2f64 immAllZerosV), (V_SET0)>; def : Pat<(v4i32 immAllZerosV), (V_SET0)>; def : Pat<(v2i64 immAllZerosV), (V_SET0)>; @@ -394,35 +394,29 @@ def : Pat<(v8i16 immAllZerosV), (V_SET0)>; def : Pat<(v16i8 immAllZerosV), (V_SET0)>; -// The same as done above but for AVX. The 256-bit ISA does not support PI, +// The same as done above but for AVX. The 256-bit AVX1 ISA doesn't support PI, // and doesn't need it because on sandy bridge the register is set to zero // at the rename stage without using any execution unit, so SET0PSY // and SET0PDY can be used for vector int instructions without penalty -// FIXME: Change encoding to pseudo! This is blocked right now by the x86 -// JIT implementatioan, it does not expand the instructions below like -// X86MCInstLower does. let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1, - isCodeGenOnly = 1 in { -let Predicates = [HasAVX] in { -def AVX_SET0PSY : PSI<0x57, MRMInitReg, (outs VR256:$dst), (ins), "", - [(set VR256:$dst, (v8f32 immAllZerosV))]>, VEX_4V; -def AVX_SET0PDY : PDI<0x57, MRMInitReg, (outs VR256:$dst), (ins), "", - [(set VR256:$dst, (v4f64 immAllZerosV))]>, VEX_4V; -} -let Predicates = [HasAVX2], neverHasSideEffects = 1 in -def AVX2_SET0 : PDI<0xef, MRMInitReg, (outs VR256:$dst), (ins), "", - []>, VEX_4V; + isPseudo = 1, Predicates = [HasAVX] in { +def AVX_SET0 : I<0, Pseudo, (outs VR256:$dst), (ins), "", + [(set VR256:$dst, (v8f32 immAllZerosV))]>; } -let Predicates = [HasAVX2], AddedComplexity = 5 in { - def : Pat<(v4i64 immAllZerosV), (AVX2_SET0)>; - def : Pat<(v8i32 immAllZerosV), (AVX2_SET0)>; - def : Pat<(v16i16 immAllZerosV), (AVX2_SET0)>; - def : Pat<(v32i8 immAllZerosV), (AVX2_SET0)>; +let Predicates = [HasAVX] in + def : Pat<(v4f64 immAllZerosV), (AVX_SET0)>; + +let Predicates = [HasAVX2] in { + def : Pat<(v4i64 immAllZerosV), (AVX_SET0)>; + def : Pat<(v8i32 immAllZerosV), (AVX_SET0)>; + def : Pat<(v16i16 immAllZerosV), (AVX_SET0)>; + def : Pat<(v32i8 immAllZerosV), (AVX_SET0)>; } -// AVX has no support for 256-bit integer instructions, but since the 128-bit +// AVX1 has no support for 256-bit integer instructions, but since the 128-bit // VPXOR instruction writes zero to its upper part, it's safe build zeros. +let Predicates = [HasAVX1Only] in { def : Pat<(v32i8 immAllZerosV), (SUBREG_TO_REG (i8 0), (V_SET0), sub_xmm)>; def : Pat<(bc_v32i8 (v8f32 immAllZerosV)), (SUBREG_TO_REG (i8 0), (V_SET0), sub_xmm)>; @@ -438,22 +432,17 @@ def : Pat<(bc_v8i32 (v8f32 immAllZerosV)), def : Pat<(v4i64 immAllZerosV), (SUBREG_TO_REG (i64 0), (V_SET0), sub_xmm)>; def : Pat<(bc_v4i64 (v8f32 immAllZerosV)), (SUBREG_TO_REG (i64 0), (V_SET0), sub_xmm)>; +} // We set canFoldAsLoad because this can be converted to a constant-pool // load of an all-ones value if folding it would be beneficial. -// FIXME: Change encoding to pseudo! This is blocked right now by the x86 -// JIT implementation, it does not expand the instructions below like -// X86MCInstLower does. let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1, - isCodeGenOnly = 1, ExeDomain = SSEPackedInt in { - let Predicates = [HasAVX] in - def AVX_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "", - [(set VR128:$dst, (v4i32 immAllOnesV))]>, VEX_4V; - def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "", - [(set VR128:$dst, (v4i32 immAllOnesV))]>; + isPseudo = 1 in { + def V_SETALLONES : I<0, Pseudo, (outs VR128:$dst), (ins), "", + [(set VR128:$dst, (v4i32 immAllOnesV))]>; let Predicates = [HasAVX2] in - def AVX2_SETALLONES : PDI<0x76, MRMInitReg, (outs VR256:$dst), (ins), "", - [(set VR256:$dst, (v8i32 immAllOnesV))]>, VEX_4V; + def AVX2_SETALLONES : I<0, Pseudo, (outs VR256:$dst), (ins), "", + [(set VR256:$dst, (v8i32 immAllOnesV))]>; } @@ -2101,12 +2090,20 @@ let Predicates = [HasAVX] in { def : Pat<(v4f32 (fround (loadv4f64 addr:$src))), (VCVTPD2PSYrm addr:$src)>; + def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))), + (VCVTPS2PDrr VR128:$src)>; def : Pat<(v4f64 (fextend (v4f32 VR128:$src))), (VCVTPS2PDYrr VR128:$src)>; def : Pat<(v4f64 (fextend (loadv4f32 addr:$src))), (VCVTPS2PDYrm addr:$src)>; } +let Predicates = [HasSSE2] in { + // Match fextend for 128 conversions + def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))), + (CVTPS2PDrr VR128:$src)>; +} + //===----------------------------------------------------------------------===// // SSE 1 & 2 - Compare Instructions //===----------------------------------------------------------------------===// @@ -2492,6 +2489,26 @@ let Constraints = "$src1 = $dst" in { SSEPackedDouble>, TB, OpSize; } // Constraints = "$src1 = $dst" +let Predicates = [HasAVX1Only] in { + def : Pat<(v8i32 (X86Unpckl VR256:$src1, (bc_v8i32 (memopv4i64 addr:$src2)))), + (VUNPCKLPSYrm VR256:$src1, addr:$src2)>; + def : Pat<(v8i32 (X86Unpckl VR256:$src1, VR256:$src2)), + (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>; + def : Pat<(v8i32 (X86Unpckh VR256:$src1, (bc_v8i32 (memopv4i64 addr:$src2)))), + (VUNPCKHPSYrm VR256:$src1, addr:$src2)>; + def : Pat<(v8i32 (X86Unpckh VR256:$src1, VR256:$src2)), + (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>; + + def : Pat<(v4i64 (X86Unpckl VR256:$src1, (memopv4i64 addr:$src2))), + (VUNPCKLPDYrm VR256:$src1, addr:$src2)>; + def : Pat<(v4i64 (X86Unpckl VR256:$src1, VR256:$src2)), + (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>; + def : Pat<(v4i64 (X86Unpckh VR256:$src1, (memopv4i64 addr:$src2))), + (VUNPCKHPDYrm VR256:$src1, addr:$src2)>; + def : Pat<(v4i64 (X86Unpckh VR256:$src1, VR256:$src2)), + (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>; +} + let Predicates = [HasAVX], AddedComplexity = 1 in { // FIXME: Instead of X86Movddup, there should be a X86Unpckl here, the // problem is during lowering, where it's not possible to recognize the load @@ -2675,14 +2692,12 @@ multiclass sse12_fp_alias_pack_logical opc, string OpcodeStr, } // Alias bitwise logical operations using SSE logical ops on packed FP values. -let mayLoad = 0 in { - defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand, - SSE_BIT_ITINS_P>; - defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for, - SSE_BIT_ITINS_P>; - defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor, - SSE_BIT_ITINS_P>; -} +defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand, + SSE_BIT_ITINS_P>; +defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for, + SSE_BIT_ITINS_P>; +defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor, + SSE_BIT_ITINS_P>; let neverHasSideEffects = 1, Pattern = [], isCommutable = 0 in defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", undef, @@ -2786,27 +2801,23 @@ multiclass basic_sse12_fp_binop_s opc, string OpcodeStr, SDNode OpNode, multiclass basic_sse12_fp_binop_p opc, string OpcodeStr, SDNode OpNode, SizeItins itins, bit Is2Addr = 1> { - let mayLoad = 0 in { defm PS : sse12_fp_packed, TB; defm PD : sse12_fp_packed, TB, OpSize; - } } multiclass basic_sse12_fp_binop_p_y opc, string OpcodeStr, SDNode OpNode, SizeItins itins> { - let mayLoad = 0 in { - defm PSY : sse12_fp_packed, TB; - defm PDY : sse12_fp_packed, TB, OpSize; - } } multiclass basic_sse12_fp_binop_s_int opc, string OpcodeStr, @@ -2864,7 +2875,8 @@ let isCommutable = 0 in { basic_sse12_fp_binop_s_int<0x5C, "sub", SSE_ALU_ITINS_S, 0>, VEX_4V, VEX_LIG; defm VSUB : basic_sse12_fp_binop_p<0x5C, "sub", fsub, SSE_ALU_ITINS_P, 0>, - basic_sse12_fp_binop_p_y<0x5C, "sub", fsub, SSE_ALU_ITINS_P>, VEX_4V; + basic_sse12_fp_binop_p_y<0x5C, "sub", fsub, SSE_ALU_ITINS_P>, + VEX_4V; defm VDIV : basic_sse12_fp_binop_s<0x5E, "div", fdiv, SSE_DIV_ITINS_S, 0>, basic_sse12_fp_binop_s_int<0x5E, "div", SSE_DIV_ITINS_S, 0>, VEX_4V, VEX_LIG; @@ -2915,6 +2927,23 @@ let Constraints = "$src1 = $dst" in { } } +let isCommutable = 1, isCodeGenOnly = 1 in { + defm VMAXC: basic_sse12_fp_binop_s<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_S, 0>, + VEX_4V, VEX_LIG; + defm VMAXC: basic_sse12_fp_binop_p<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_P, 0>, + basic_sse12_fp_binop_p_y<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_P>, VEX_4V; + defm VMINC: basic_sse12_fp_binop_s<0x5D, "min", X86fminc, SSE_ALU_ITINS_S, 0>, + VEX_4V, VEX_LIG; + defm VMINC: basic_sse12_fp_binop_p<0x5D, "min", X86fminc, SSE_ALU_ITINS_P, 0>, + basic_sse12_fp_binop_p_y<0x5D, "min", X86fminc, SSE_ALU_ITINS_P>, VEX_4V; + let Constraints = "$src1 = $dst" in { + defm MAXC: basic_sse12_fp_binop_s<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_S>, + basic_sse12_fp_binop_p<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_P>; + defm MINC: basic_sse12_fp_binop_s<0x5D, "min", X86fminc, SSE_ALU_ITINS_S>, + basic_sse12_fp_binop_p<0x5D, "min", X86fminc, SSE_ALU_ITINS_P>; + } +} + /// Unop Arithmetic /// In addition, we also have a special variant of the scalar form here to /// represent the associated intrinsic operation. This form is unlike the @@ -2966,7 +2995,7 @@ multiclass sse1_fp_unop_s_avx opc, string OpcodeStr> { def SSr : SSI; - let mayLoad = 1 in + let mayLoad = 1 in { def SSm : SSI; @@ -2974,6 +3003,7 @@ multiclass sse1_fp_unop_s_avx opc, string OpcodeStr> { (ins VR128:$src1, ssmem:$src2), !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>; + } } /// sse1_fp_unop_p - SSE1 unops in packed form. @@ -3046,20 +3076,20 @@ multiclass sse2_fp_unop_s opc, string OpcodeStr, } /// sse2_fp_unop_s_avx - AVX SSE2 unops in scalar form. +let hasSideEffects = 0 in multiclass sse2_fp_unop_s_avx opc, string OpcodeStr> { - let neverHasSideEffects = 1 in { def SDr : SDI; - let mayLoad = 1 in + let mayLoad = 1 in { def SDm : SDI; - } def SDm_Int : SDI; + } } /// sse2_fp_unop_p - SSE2 unops in vector forms. @@ -4299,28 +4329,6 @@ let Constraints = "$src1 = $dst" in { } } // ExeDomain = SSEPackedInt -// Patterns for using AVX1 instructions with integer vectors -// Here to give AVX2 priority -let Predicates = [HasAVX] in { - def : Pat<(v8i32 (X86Unpckl VR256:$src1, (bc_v8i32 (memopv4i64 addr:$src2)))), - (VUNPCKLPSYrm VR256:$src1, addr:$src2)>; - def : Pat<(v8i32 (X86Unpckl VR256:$src1, VR256:$src2)), - (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>; - def : Pat<(v8i32 (X86Unpckh VR256:$src1, (bc_v8i32 (memopv4i64 addr:$src2)))), - (VUNPCKHPSYrm VR256:$src1, addr:$src2)>; - def : Pat<(v8i32 (X86Unpckh VR256:$src1, VR256:$src2)), - (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>; - - def : Pat<(v4i64 (X86Unpckl VR256:$src1, (memopv4i64 addr:$src2))), - (VUNPCKLPDYrm VR256:$src1, addr:$src2)>; - def : Pat<(v4i64 (X86Unpckl VR256:$src1, VR256:$src2)), - (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>; - def : Pat<(v4i64 (X86Unpckh VR256:$src1, (memopv4i64 addr:$src2))), - (VUNPCKHPDYrm VR256:$src1, addr:$src2)>; - def : Pat<(v4i64 (X86Unpckh VR256:$src1, VR256:$src2)), - (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>; -} - //===---------------------------------------------------------------------===// // SSE2 - Packed Integer Extract and Insert //===---------------------------------------------------------------------===// @@ -6841,81 +6849,42 @@ let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX], neverHasSideEffects = 1 in { } // Packed Compare Implicit Length Strings, Return Index -let Defs = [ECX, EFLAGS] in { - multiclass SS42AI_pcmpistri { +let Defs = [ECX, EFLAGS], neverHasSideEffects = 1 in { + multiclass SS42AI_pcmpistri { def rr : SS42AI<0x63, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2, i8imm:$src3), !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"), - [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)), - (implicit EFLAGS)]>, OpSize; + []>, OpSize; + let mayLoad = 1 in def rm : SS42AI<0x63, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2, i8imm:$src3), !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"), - [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)), - (implicit EFLAGS)]>, OpSize; + []>, OpSize; } } -let Predicates = [HasAVX] in { -defm VPCMPISTRI : SS42AI_pcmpistri, - VEX; -defm VPCMPISTRIA : SS42AI_pcmpistri, - VEX; -defm VPCMPISTRIC : SS42AI_pcmpistri, - VEX; -defm VPCMPISTRIO : SS42AI_pcmpistri, - VEX; -defm VPCMPISTRIS : SS42AI_pcmpistri, - VEX; -defm VPCMPISTRIZ : SS42AI_pcmpistri, - VEX; -} - -defm PCMPISTRI : SS42AI_pcmpistri; -defm PCMPISTRIA : SS42AI_pcmpistri; -defm PCMPISTRIC : SS42AI_pcmpistri; -defm PCMPISTRIO : SS42AI_pcmpistri; -defm PCMPISTRIS : SS42AI_pcmpistri; -defm PCMPISTRIZ : SS42AI_pcmpistri; +let Predicates = [HasAVX] in +defm VPCMPISTRI : SS42AI_pcmpistri<"vpcmpistri">, VEX; +defm PCMPISTRI : SS42AI_pcmpistri<"pcmpistri">; // Packed Compare Explicit Length Strings, Return Index -let Defs = [ECX, EFLAGS], Uses = [EAX, EDX] in { - multiclass SS42AI_pcmpestri { +let Defs = [ECX, EFLAGS], Uses = [EAX, EDX], neverHasSideEffects = 1 in { + multiclass SS42AI_pcmpestri { def rr : SS42AI<0x61, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src3, i8imm:$src5), !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"), - [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)), - (implicit EFLAGS)]>, OpSize; + []>, OpSize; + let mayLoad = 1 in def rm : SS42AI<0x61, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src3, i8imm:$src5), !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"), - [(set ECX, - (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)), - (implicit EFLAGS)]>, OpSize; + []>, OpSize; } } -let Predicates = [HasAVX] in { -defm VPCMPESTRI : SS42AI_pcmpestri, - VEX; -defm VPCMPESTRIA : SS42AI_pcmpestri, - VEX; -defm VPCMPESTRIC : SS42AI_pcmpestri, - VEX; -defm VPCMPESTRIO : SS42AI_pcmpestri, - VEX; -defm VPCMPESTRIS : SS42AI_pcmpestri, - VEX; -defm VPCMPESTRIZ : SS42AI_pcmpestri, - VEX; -} - -defm PCMPESTRI : SS42AI_pcmpestri; -defm PCMPESTRIA : SS42AI_pcmpestri; -defm PCMPESTRIC : SS42AI_pcmpestri; -defm PCMPESTRIO : SS42AI_pcmpestri; -defm PCMPESTRIS : SS42AI_pcmpestri; -defm PCMPESTRIZ : SS42AI_pcmpestri; +let Predicates = [HasAVX] in +defm VPCMPESTRI : SS42AI_pcmpestri<"vpcmpestri">, VEX; +defm PCMPESTRI : SS42AI_pcmpestri<"pcmpestri">; //===----------------------------------------------------------------------===// // SSE4.2 - CRC Instructions @@ -7257,6 +7226,18 @@ def : Pat<(vinsertf128_insert:$ins (v4f64 VR256:$src1), (v2f64 VR128:$src2), (i32 imm)), (VINSERTF128rr VR256:$src1, VR128:$src2, (INSERT_get_vinsertf128_imm VR256:$ins))>; + +def : Pat<(vinsertf128_insert:$ins (v8f32 VR256:$src1), (loadv4f32 addr:$src2), + (i32 imm)), + (VINSERTF128rm VR256:$src1, addr:$src2, + (INSERT_get_vinsertf128_imm VR256:$ins))>; +def : Pat<(vinsertf128_insert:$ins (v4f64 VR256:$src1), (loadv2f64 addr:$src2), + (i32 imm)), + (VINSERTF128rm VR256:$src1, addr:$src2, + (INSERT_get_vinsertf128_imm VR256:$ins))>; +} + +let Predicates = [HasAVX1Only] in { def : Pat<(vinsertf128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2), (i32 imm)), (VINSERTF128rr VR256:$src1, VR128:$src2, @@ -7274,14 +7255,6 @@ def : Pat<(vinsertf128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2), (VINSERTF128rr VR256:$src1, VR128:$src2, (INSERT_get_vinsertf128_imm VR256:$ins))>; -def : Pat<(vinsertf128_insert:$ins (v8f32 VR256:$src1), (loadv4f32 addr:$src2), - (i32 imm)), - (VINSERTF128rm VR256:$src1, addr:$src2, - (INSERT_get_vinsertf128_imm VR256:$ins))>; -def : Pat<(vinsertf128_insert:$ins (v4f64 VR256:$src1), (loadv2f64 addr:$src2), - (i32 imm)), - (VINSERTF128rm VR256:$src1, addr:$src2, - (INSERT_get_vinsertf128_imm VR256:$ins))>; def : Pat<(vinsertf128_insert:$ins (v4i64 VR256:$src1), (loadv2i64 addr:$src2), (i32 imm)), (VINSERTF128rm VR256:$src1, addr:$src2, @@ -7305,19 +7278,15 @@ def VEXTRACTF128mr : AVXAIi8<0x19, MRMDestMem, (outs), // Extract and store. let Predicates = [HasAVX] in { - def : Pat<(alignedstore (int_x86_avx_vextractf128_ps_256 VR256:$src1, imm:$src2), addr:$dst), - (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; - def : Pat<(alignedstore (int_x86_avx_vextractf128_pd_256 VR256:$src1, imm:$src2), addr:$dst), - (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; - def : Pat<(alignedstore (int_x86_avx_vextractf128_si_256 VR256:$src1, imm:$src2), addr:$dst), - (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; - - def : Pat<(int_x86_sse_storeu_ps addr:$dst, (int_x86_avx_vextractf128_ps_256 VR256:$src1, imm:$src2)), - (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; - def : Pat<(int_x86_sse2_storeu_pd addr:$dst, (int_x86_avx_vextractf128_pd_256 VR256:$src1, imm:$src2)), - (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; - def : Pat<(int_x86_sse2_storeu_dq addr:$dst, (bc_v16i8 (int_x86_avx_vextractf128_si_256 VR256:$src1, imm:$src2))), - (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; + def : Pat<(alignedstore (int_x86_avx_vextractf128_ps_256 VR256:$src1, + imm:$src2), addr:$dst), + (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; + def : Pat<(alignedstore (int_x86_avx_vextractf128_pd_256 VR256:$src1, + imm:$src2), addr:$dst), + (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; + def : Pat<(alignedstore (int_x86_avx_vextractf128_si_256 VR256:$src1, + imm:$src2), addr:$dst), + (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>; } // AVX1 patterns @@ -7337,6 +7306,9 @@ def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)), (v2f64 (VEXTRACTF128rr (v4f64 VR256:$src1), (EXTRACT_get_vextractf128_imm VR128:$ext)))>; +} + +let Predicates = [HasAVX1Only] in { def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)), (v2i64 (VEXTRACTF128rr (v4i64 VR256:$src1), @@ -7469,29 +7441,29 @@ def VPERM2F128rm : AVXAIi8<0x06, MRMSrcMem, (outs VR256:$dst), } let Predicates = [HasAVX] in { +def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))), + (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>; +def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, + (memopv4f64 addr:$src2), (i8 imm:$imm))), + (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>; +} + +let Predicates = [HasAVX1Only] in { def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))), (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>; def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))), (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>; -def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))), - (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>; def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))), (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>; def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))), (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>; -def : Pat<(v8f32 (X86VPerm2x128 VR256:$src1, - (memopv8f32 addr:$src2), (i8 imm:$imm))), - (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>; def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, (bc_v8i32 (memopv4i64 addr:$src2)), (i8 imm:$imm))), (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>; def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1, (memopv4i64 addr:$src2), (i8 imm:$imm))), (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>; -def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, - (memopv4f64 addr:$src2), (i8 imm:$imm))), - (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>; def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, (bc_v32i8 (memopv4i64 addr:$src2)), (i8 imm:$imm))), (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>; @@ -7678,19 +7650,22 @@ let Predicates = [HasAVX2] in { } // AVX1 broadcast patterns -let Predicates = [HasAVX] in { +let Predicates = [HasAVX1Only] in { def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))), (VBROADCASTSSYrm addr:$src)>; def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))), (VBROADCASTSDYrm addr:$src)>; +def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))), + (VBROADCASTSSrm addr:$src)>; +} + +let Predicates = [HasAVX] in { def : Pat<(v8f32 (X86VBroadcast (loadf32 addr:$src))), (VBROADCASTSSYrm addr:$src)>; def : Pat<(v4f64 (X86VBroadcast (loadf64 addr:$src))), (VBROADCASTSDYrm addr:$src)>; def : Pat<(v4f32 (X86VBroadcast (loadf32 addr:$src))), (VBROADCASTSSrm addr:$src)>; -def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))), - (VBROADCASTSSrm addr:$src)>; // Provide fallback in case the load node that is used in the patterns above // is used by additional users, which prevents the pattern selection. @@ -7770,7 +7745,6 @@ defm VPERMPD : avx2_perm_imm<0x01, "vpermpd", memopv4f64, v4f64>, VEX_W; //===----------------------------------------------------------------------===// // VPERM2I128 - Permute Floating-Point Values in 128-bit chunks // -let AddedComplexity = 1 in { def VPERM2I128rr : AVX2AIi8<0x46, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src1, VR256:$src2, i8imm:$src3), "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", @@ -7781,9 +7755,8 @@ def VPERM2I128rm : AVX2AIi8<0x46, MRMSrcMem, (outs VR256:$dst), "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (memopv4i64 addr:$src2), (i8 imm:$src3)))]>, VEX_4V; -} -let Predicates = [HasAVX2], AddedComplexity = 1 in { +let Predicates = [HasAVX2] in { def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))), (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>; def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))), @@ -7818,7 +7791,7 @@ def VINSERTI128rm : AVX2AIi8<0x38, MRMSrcMem, (outs VR256:$dst), []>, VEX_4V; } -let Predicates = [HasAVX2], AddedComplexity = 1 in { +let Predicates = [HasAVX2] in { def : Pat<(vinsertf128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2), (i32 imm)), (VINSERTI128rr VR256:$src1, VR128:$src2, @@ -7835,6 +7808,11 @@ def : Pat<(vinsertf128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2), (i32 imm)), (VINSERTI128rr VR256:$src1, VR128:$src2, (INSERT_get_vinsertf128_imm VR256:$ins))>; + +def : Pat<(vinsertf128_insert:$ins (v4i64 VR256:$src1), (loadv2i64 addr:$src2), + (i32 imm)), + (VINSERTI128rm VR256:$src1, addr:$src2, + (INSERT_get_vinsertf128_imm VR256:$ins))>; } //===----------------------------------------------------------------------===// @@ -7851,7 +7829,7 @@ def VEXTRACTI128mr : AVX2AIi8<0x39, MRMDestMem, (outs), (ins i128mem:$dst, VR256:$src1, i8imm:$src2), "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>, VEX; -let Predicates = [HasAVX2], AddedComplexity = 1 in { +let Predicates = [HasAVX2] in { def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)), (v2i64 (VEXTRACTI128rr (v4i64 VR256:$src1),