diff options
Diffstat (limited to 'lib/Target/X86/X86InstrSSE.td')
| -rw-r--r-- | lib/Target/X86/X86InstrSSE.td | 427 |
1 files changed, 235 insertions, 192 deletions
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td index e26c979..2743dba 100644 --- a/lib/Target/X86/X86InstrSSE.td +++ b/lib/Target/X86/X86InstrSSE.td @@ -160,6 +160,32 @@ def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>; def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>; def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>; +// MOVNT Support +// Like 'store', but requires the non-temporal bit to be set +def nontemporalstore : PatFrag<(ops node:$val, node:$ptr), + (st node:$val, node:$ptr), [{ + if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) + return ST->isNonTemporal(); + return false; +}]>; + +def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr), + (st node:$val, node:$ptr), [{ + if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) + return ST->isNonTemporal() && !ST->isTruncatingStore() && + ST->getAddressingMode() == ISD::UNINDEXED && + ST->getAlignment() >= 16; + return false; +}]>; + +def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr), + (st node:$val, node:$ptr), [{ + if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) + return ST->isNonTemporal() && + ST->getAlignment() < 16; + return false; +}]>; + def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>; def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>; def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>; @@ -344,18 +370,56 @@ let Uses = [EFLAGS], usesCustomInserter = 1 in { // SSE1 Instructions //===----------------------------------------------------------------------===// -// Move Instructions -let neverHasSideEffects = 1 in -def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src), - "movss\t{$src, $dst|$dst, $src}", []>; -let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in +// Move Instructions. Register-to-register movss is not used for FR32 +// register copies because it's a partial register update; FsMOVAPSrr is +// used instead. Register-to-register movss is not modeled as an INSERT_SUBREG +// because INSERT_SUBREG requires that the insert be implementable in terms of +// a copy, and just mentioned, we don't use movss for copies. +let Constraints = "$src1 = $dst" in +def MOVSSrr : SSI<0x10, MRMSrcReg, + (outs VR128:$dst), (ins VR128:$src1, FR32:$src2), + "movss\t{$src2, $dst|$dst, $src2}", + [(set VR128:$dst, + (movl VR128:$src1, (scalar_to_vector FR32:$src2)))]>; + +// Extract the low 32-bit value from one vector and insert it into another. +let AddedComplexity = 15 in +def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)), + (MOVSSrr VR128:$src1, + (EXTRACT_SUBREG (v4f32 VR128:$src2), x86_subreg_ss))>; + +// Implicitly promote a 32-bit scalar to a vector. +def : Pat<(v4f32 (scalar_to_vector FR32:$src)), + (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, x86_subreg_ss)>; + +// Loading from memory automatically zeroing upper bits. +let canFoldAsLoad = 1, isReMaterializable = 1 in def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src), "movss\t{$src, $dst|$dst, $src}", [(set FR32:$dst, (loadf32 addr:$src))]>; + +// MOVSSrm zeros the high parts of the register; represent this +// with SUBREG_TO_REG. +let AddedComplexity = 20 in { +def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))), + (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), x86_subreg_ss)>; +def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))), + (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), x86_subreg_ss)>; +def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))), + (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), x86_subreg_ss)>; +} + +// Store scalar value to memory. def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src), "movss\t{$src, $dst|$dst, $src}", [(store FR32:$src, addr:$dst)]>; +// Extract and store. +def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))), + addr:$dst), + (MOVSSmr addr:$dst, + (EXTRACT_SUBREG (v4f32 VR128:$src), x86_subreg_ss))>; + // Conversion instructions def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src), "cvttss2si\t{$src, $dst|$dst, $src}", @@ -518,7 +582,7 @@ def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src), // Alias instruction to load FR32 from f128mem using movaps. Upper bits are // disregarded. -let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in +let canFoldAsLoad = 1, isReMaterializable = 1 in def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src), "movaps\t{$src, $dst|$dst, $src}", [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>; @@ -715,7 +779,7 @@ defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin, let neverHasSideEffects = 1 in def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src), "movaps\t{$src, $dst|$dst, $src}", []>; -let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in +let canFoldAsLoad = 1, isReMaterializable = 1 in def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src), "movaps\t{$src, $dst|$dst, $src}", [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>; @@ -727,7 +791,7 @@ def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src), let neverHasSideEffects = 1 in def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src), "movups\t{$src, $dst|$dst, $src}", []>; -let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in +let canFoldAsLoad = 1, isReMaterializable = 1 in def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src), "movups\t{$src, $dst|$dst, $src}", [(set VR128:$dst, (loadv4f32 addr:$src))]>; @@ -736,7 +800,7 @@ def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src), [(store (v4f32 VR128:$src), addr:$dst)]>; // Intrinsic forms of MOVUPS load and store -let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in +let canFoldAsLoad = 1, isReMaterializable = 1 in def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src), "movups\t{$src, $dst|$dst, $src}", [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>; @@ -762,6 +826,9 @@ let Constraints = "$src1 = $dst" in { } // Constraints = "$src1 = $dst" +def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))), + (MOVHPSrm VR128:$src1, addr:$src2)>; + def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src), "movlps\t{$src, $dst|$dst, $src}", [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)), @@ -793,9 +860,9 @@ def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), let AddedComplexity = 20 in { def : Pat<(v4f32 (movddup VR128:$src, (undef))), - (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>; + (MOVLHPSrr VR128:$src, VR128:$src)>; def : Pat<(v2i64 (movddup VR128:$src, (undef))), - (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>; + (MOVLHPSrr VR128:$src, VR128:$src)>; } @@ -1010,10 +1077,33 @@ def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src), "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>; // Non-temporal stores -def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src), +def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src), "movntps\t{$src, $dst|$dst, $src}", [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>; +let AddedComplexity = 400 in { // Prefer non-temporal versions +def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src), + "movntps\t{$src, $dst|$dst, $src}", + [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>; + +def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src), + "movntdq\t{$src, $dst|$dst, $src}", + [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>; + +def : Pat<(alignednontemporalstore (v2i64 VR128:$src), addr:$dst), + (MOVNTDQ_64mr VR128:$src, addr:$dst)>; + +def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src), + "movnti\t{$src, $dst|$dst, $src}", + [(nontemporalstore (i32 GR32:$src), addr:$dst)]>, + TB, Requires<[HasSSE2]>; + +def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src), + "movnti\t{$src, $dst|$dst, $src}", + [(nontemporalstore (i64 GR64:$src), addr:$dst)]>, + TB, Requires<[HasSSE2]>; +} + // Load, store, and memory fence def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>; @@ -1032,84 +1122,73 @@ let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1, def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "", [(set VR128:$dst, (v4i32 immAllZerosV))]>; -let Predicates = [HasSSE1] in { - def : Pat<(v2i64 immAllZerosV), (V_SET0)>; - def : Pat<(v8i16 immAllZerosV), (V_SET0)>; - def : Pat<(v16i8 immAllZerosV), (V_SET0)>; - def : Pat<(v2f64 immAllZerosV), (V_SET0)>; - def : Pat<(v4f32 immAllZerosV), (V_SET0)>; -} - -// FR32 to 128-bit vector conversion. -let isAsCheapAsAMove = 1 in -def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src), - "movss\t{$src, $dst|$dst, $src}", - [(set VR128:$dst, - (v4f32 (scalar_to_vector FR32:$src)))]>; -def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src), - "movss\t{$src, $dst|$dst, $src}", - [(set VR128:$dst, - (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>; - -// FIXME: may not be able to eliminate this movss with coalescing the src and -// dest register classes are different. We really want to write this pattern -// like this: -// def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))), -// (f32 FR32:$src)>; -let isAsCheapAsAMove = 1 in -def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src), - "movss\t{$src, $dst|$dst, $src}", - [(set FR32:$dst, (vector_extract (v4f32 VR128:$src), - (iPTR 0)))]>; -def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src), - "movss\t{$src, $dst|$dst, $src}", - [(store (f32 (vector_extract (v4f32 VR128:$src), - (iPTR 0))), addr:$dst)]>; - - -// Move to lower bits of a VR128, leaving upper bits alone. -// Three operand (but two address) aliases. -let Constraints = "$src1 = $dst" in { -let neverHasSideEffects = 1 in - def MOVLSS2PSrr : SSI<0x10, MRMSrcReg, - (outs VR128:$dst), (ins VR128:$src1, FR32:$src2), - "movss\t{$src2, $dst|$dst, $src2}", []>; - - let AddedComplexity = 15 in - def MOVLPSrr : SSI<0x10, MRMSrcReg, - (outs VR128:$dst), (ins VR128:$src1, VR128:$src2), - "movss\t{$src2, $dst|$dst, $src2}", - [(set VR128:$dst, - (v4f32 (movl VR128:$src1, VR128:$src2)))]>; -} +def : Pat<(v2i64 immAllZerosV), (V_SET0)>; +def : Pat<(v8i16 immAllZerosV), (V_SET0)>; +def : Pat<(v16i8 immAllZerosV), (V_SET0)>; +def : Pat<(v2f64 immAllZerosV), (V_SET0)>; +def : Pat<(v4f32 immAllZerosV), (V_SET0)>; -// Move to lower bits of a VR128 and zeroing upper bits. -// Loading from memory automatically zeroing upper bits. -let AddedComplexity = 20 in -def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src), - "movss\t{$src, $dst|$dst, $src}", - [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector - (loadf32 addr:$src))))))]>; - -def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))), - (MOVZSS2PSrm addr:$src)>; +def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))), + (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), x86_subreg_ss))>; //===---------------------------------------------------------------------===// // SSE2 Instructions //===---------------------------------------------------------------------===// -// Move Instructions -let neverHasSideEffects = 1 in -def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src), - "movsd\t{$src, $dst|$dst, $src}", []>; -let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in +// Move Instructions. Register-to-register movsd is not used for FR64 +// register copies because it's a partial register update; FsMOVAPDrr is +// used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG +// because INSERT_SUBREG requires that the insert be implementable in terms of +// a copy, and just mentioned, we don't use movsd for copies. +let Constraints = "$src1 = $dst" in +def MOVSDrr : SDI<0x10, MRMSrcReg, + (outs VR128:$dst), (ins VR128:$src1, FR64:$src2), + "movsd\t{$src2, $dst|$dst, $src2}", + [(set VR128:$dst, + (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>; + +// Extract the low 64-bit value from one vector and insert it into another. +let AddedComplexity = 15 in +def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)), + (MOVSDrr VR128:$src1, + (EXTRACT_SUBREG (v2f64 VR128:$src2), x86_subreg_sd))>; + +// Implicitly promote a 64-bit scalar to a vector. +def : Pat<(v2f64 (scalar_to_vector FR64:$src)), + (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, x86_subreg_sd)>; + +// Loading from memory automatically zeroing upper bits. +let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src), "movsd\t{$src, $dst|$dst, $src}", [(set FR64:$dst, (loadf64 addr:$src))]>; + +// MOVSDrm zeros the high parts of the register; represent this +// with SUBREG_TO_REG. +let AddedComplexity = 20 in { +def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))), + (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>; +def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))), + (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>; +def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))), + (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>; +def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))), + (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>; +def : Pat<(v2f64 (X86vzload addr:$src)), + (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>; +} + +// Store scalar value to memory. def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src), "movsd\t{$src, $dst|$dst, $src}", [(store FR64:$src, addr:$dst)]>; +// Extract and store. +def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))), + addr:$dst), + (MOVSDmr addr:$dst, + (EXTRACT_SUBREG (v2f64 VR128:$src), x86_subreg_sd))>; + // Conversion instructions def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src), "cvttsd2si\t{$src, $dst|$dst, $src}", @@ -1163,7 +1242,8 @@ def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src), Requires<[HasSSE2, OptForSize]>; def : Pat<(extloadf32 addr:$src), - (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[HasSSE2, OptForSpeed]>; + (CVTSS2SDrr (MOVSSrm addr:$src))>, + Requires<[HasSSE2, OptForSpeed]>; // Match intrinsics which expect XMM operand(s). def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src), @@ -1282,7 +1362,7 @@ def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src), // Alias instruction to load FR64 from f128mem using movapd. Upper bits are // disregarded. -let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in +let canFoldAsLoad = 1, isReMaterializable = 1 in def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src), "movapd\t{$src, $dst|$dst, $src}", [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>; @@ -1480,7 +1560,7 @@ defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin, let neverHasSideEffects = 1 in def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src), "movapd\t{$src, $dst|$dst, $src}", []>; -let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in +let canFoldAsLoad = 1, isReMaterializable = 1 in def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src), "movapd\t{$src, $dst|$dst, $src}", [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>; @@ -2295,34 +2375,47 @@ def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask), [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>; // Non-temporal stores -def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src), - "movntpd\t{$src, $dst|$dst, $src}", - [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>; -def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src), - "movntdq\t{$src, $dst|$dst, $src}", - [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>; -def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src), +def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src), + "movntpd\t{$src, $dst|$dst, $src}", + [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>; +def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src), + "movntdq\t{$src, $dst|$dst, $src}", + [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>; +def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src), "movnti\t{$src, $dst|$dst, $src}", [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>, TB, Requires<[HasSSE2]>; +let AddedComplexity = 400 in { // Prefer non-temporal versions +def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src), + "movntpd\t{$src, $dst|$dst, $src}", + [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>; + +def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src), + "movntdq\t{$src, $dst|$dst, $src}", + [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>; + +def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst), + (MOVNTDQmr VR128:$src, addr:$dst)>; +} + // Flush cache def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src), "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>, TB, Requires<[HasSSE2]>; // Load, store, and memory fence -def LFENCE : I<0xAE, MRM5r, (outs), (ins), +def LFENCE : I<0xAE, MRM_E8, (outs), (ins), "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>; -def MFENCE : I<0xAE, MRM6r, (outs), (ins), +def MFENCE : I<0xAE, MRM_F0, (outs), (ins), "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>; //TODO: custom lower this so as to never even generate the noop -def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss), +def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 0)), (NOOP)>; def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>; def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>; -def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss), +def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)), (MFENCE)>; // Alias instructions that map zero vector to pxor / xorp* for sse. @@ -2334,17 +2427,6 @@ let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1, def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "", [(set VR128:$dst, (v4i32 immAllOnesV))]>; -// FR64 to 128-bit vector conversion. -let isAsCheapAsAMove = 1 in -def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src), - "movsd\t{$src, $dst|$dst, $src}", - [(set VR128:$dst, - (v2f64 (scalar_to_vector FR64:$src)))]>; -def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src), - "movsd\t{$src, $dst|$dst, $src}", - [(set VR128:$dst, - (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>; - def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src), "movd\t{$src, $dst|$dst, $src}", [(set VR128:$dst, @@ -2373,20 +2455,9 @@ def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src), [(store (i64 (vector_extract (v2i64 VR128:$src), (iPTR 0))), addr:$dst)]>; -// FIXME: may not be able to eliminate this movss with coalescing the src and -// dest register classes are different. We really want to write this pattern -// like this: -// def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))), -// (f32 FR32:$src)>; -let isAsCheapAsAMove = 1 in -def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src), - "movsd\t{$src, $dst|$dst, $src}", - [(set FR64:$dst, (vector_extract (v2f64 VR128:$src), - (iPTR 0)))]>; -def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src), - "movsd\t{$src, $dst|$dst, $src}", - [(store (f64 (vector_extract (v2f64 VR128:$src), - (iPTR 0))), addr:$dst)]>; +def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))), + (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), x86_subreg_sd))>; + def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src), "movd\t{$src, $dst|$dst, $src}", [(set GR32:$dst, (vector_extract (v4i32 VR128:$src), @@ -2403,44 +2474,11 @@ def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src), "movd\t{$src, $dst|$dst, $src}", [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>; - -// Move to lower bits of a VR128, leaving upper bits alone. -// Three operand (but two address) aliases. -let Constraints = "$src1 = $dst" in { - let neverHasSideEffects = 1 in - def MOVLSD2PDrr : SDI<0x10, MRMSrcReg, - (outs VR128:$dst), (ins VR128:$src1, FR64:$src2), - "movsd\t{$src2, $dst|$dst, $src2}", []>; - - let AddedComplexity = 15 in - def MOVLPDrr : SDI<0x10, MRMSrcReg, - (outs VR128:$dst), (ins VR128:$src1, VR128:$src2), - "movsd\t{$src2, $dst|$dst, $src2}", - [(set VR128:$dst, - (v2f64 (movl VR128:$src1, VR128:$src2)))]>; -} - // Store / copy lower 64-bits of a XMM register. def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src), "movq\t{$src, $dst|$dst, $src}", [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>; -// Move to lower bits of a VR128 and zeroing upper bits. -// Loading from memory automatically zeroing upper bits. -let AddedComplexity = 20 in { -def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src), - "movsd\t{$src, $dst|$dst, $src}", - [(set VR128:$dst, - (v2f64 (X86vzmovl (v2f64 (scalar_to_vector - (loadf64 addr:$src))))))]>; - -def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))), - (MOVZSD2PDrm addr:$src)>; -def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))), - (MOVZSD2PDrm addr:$src)>; -def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>; -} - // movd / movq to XMM register zero-extends let AddedComplexity = 15 in { def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src), @@ -2613,9 +2651,9 @@ let Constraints = "$src1 = $dst" in { } // Thread synchronization -def MONITOR : I<0x01, MRM1r, (outs), (ins), "monitor", +def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor", [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>; -def MWAIT : I<0x01, MRM1r, (outs), (ins), "mwait", +def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait", [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>; // vector_shuffle v1, <undef> <1, 1, 3, 3> @@ -2986,13 +3024,15 @@ let Predicates = [HasSSE2] in { let AddedComplexity = 15 in { // Zeroing a VR128 then do a MOVS{S|D} to the lower bits. def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))), - (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>; + (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>; def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))), - (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>; + (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>; def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))), - (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>; + (MOVSSrr (v4f32 (V_SET0)), + (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), x86_subreg_ss)))>; def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))), - (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>; + (MOVSSrr (v4i32 (V_SET0)), + (EXTRACT_SUBREG (v4i32 VR128:$src), x86_subreg_ss))>; } // Splat v2f64 / v2i64 @@ -3010,8 +3050,7 @@ def : Pat<(unpckh (v2i64 VR128:$src), (undef)), // Special unary SHUFPSrri case. def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))), (SHUFPSrri VR128:$src1, VR128:$src1, - (SHUFFLE_get_shuf_imm VR128:$src3))>, - Requires<[HasSSE1]>; + (SHUFFLE_get_shuf_imm VR128:$src3))>; let AddedComplexity = 5 in def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))), (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>, @@ -3057,13 +3096,13 @@ def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))), } let AddedComplexity = 10 in { def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))), - (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>; + (UNPCKLPSrr VR128:$src, VR128:$src)>; def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))), - (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>; + (PUNPCKLBWrr VR128:$src, VR128:$src)>; def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))), - (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>; + (PUNPCKLWDrr VR128:$src, VR128:$src)>; def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))), - (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>; + (PUNPCKLDQrr VR128:$src, VR128:$src)>; } // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...> @@ -3077,13 +3116,13 @@ def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))), } let AddedComplexity = 10 in { def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))), - (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>; + (UNPCKHPSrr VR128:$src, VR128:$src)>; def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))), - (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>; + (PUNPCKHBWrr VR128:$src, VR128:$src)>; def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))), - (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>; + (PUNPCKHWDrr VR128:$src, VR128:$src)>; def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))), - (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>; + (PUNPCKHDQrr VR128:$src, VR128:$src)>; } let AddedComplexity = 20 in { @@ -3105,45 +3144,49 @@ def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))), let AddedComplexity = 20 in { // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))), - (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>; + (MOVLPSrm VR128:$src1, addr:$src2)>; def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))), - (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>; + (MOVLPDrm VR128:$src1, addr:$src2)>; def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))), - (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>; + (MOVLPSrm VR128:$src1, addr:$src2)>; def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))), - (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>; + (MOVLPDrm VR128:$src1, addr:$src2)>; } // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1), - (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>; + (MOVLPSmr addr:$src1, VR128:$src2)>; def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1), - (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>; + (MOVLPDmr addr:$src1, VR128:$src2)>; def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)), addr:$src1), - (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>; + (MOVLPSmr addr:$src1, VR128:$src2)>; def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1), - (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>; + (MOVLPDmr addr:$src1, VR128:$src2)>; let AddedComplexity = 15 in { // Setting the lowest element in the vector. def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)), - (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>; + (MOVSSrr (v4i32 VR128:$src1), + (EXTRACT_SUBREG (v4i32 VR128:$src2), x86_subreg_ss))>; def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)), - (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>; + (MOVSDrr (v2i64 VR128:$src1), + (EXTRACT_SUBREG (v2i64 VR128:$src2), x86_subreg_sd))>; -// vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd) +// vector_shuffle v1, v2 <4, 5, 2, 3> using movsd def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)), - (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>; + (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, x86_subreg_sd))>, + Requires<[HasSSE2]>; def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)), - (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>; + (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, x86_subreg_sd))>, + Requires<[HasSSE2]>; } // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but // fall back to this for SSE1) def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))), (SHUFPSrri VR128:$src2, VR128:$src1, - (SHUFFLE_get_shuf_imm VR128:$src3))>, Requires<[HasSSE1]>; + (SHUFFLE_get_shuf_imm VR128:$src3))>; // Set lowest element and zero upper elements. let AddedComplexity = 15 in @@ -3185,30 +3228,30 @@ def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))), // Use movaps / movups for SSE integer load / store (one byte shorter). def : Pat<(alignedloadv4i32 addr:$src), - (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>; + (MOVAPSrm addr:$src)>; def : Pat<(loadv4i32 addr:$src), - (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>; + (MOVUPSrm addr:$src)>; def : Pat<(alignedloadv2i64 addr:$src), - (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>; + (MOVAPSrm addr:$src)>; def : Pat<(loadv2i64 addr:$src), - (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>; + (MOVUPSrm addr:$src)>; def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst), - (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>; + (MOVAPSmr addr:$dst, VR128:$src)>; def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst), - (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>; + (MOVAPSmr addr:$dst, VR128:$src)>; def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst), - (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>; + (MOVAPSmr addr:$dst, VR128:$src)>; def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst), - (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>; + (MOVAPSmr addr:$dst, VR128:$src)>; def : Pat<(store (v2i64 VR128:$src), addr:$dst), - (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>; + (MOVUPSmr addr:$dst, VR128:$src)>; def : Pat<(store (v4i32 VR128:$src), addr:$dst), - (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>; + (MOVUPSmr addr:$dst, VR128:$src)>; def : Pat<(store (v8i16 VR128:$src), addr:$dst), - (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>; + (MOVUPSmr addr:$dst, VR128:$src)>; def : Pat<(store (v16i8 VR128:$src), addr:$dst), - (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>; + (MOVUPSmr addr:$dst, VR128:$src)>; //===----------------------------------------------------------------------===// // SSE4.1 Instructions @@ -3397,7 +3440,7 @@ let Constraints = "$src1 = $dst" in { (ins VR128:$src1, i128mem:$src2), !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), [(set VR128:$dst, - (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize; + (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize; def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2), !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), |