From fa774e647fc203bf13650e65bedddab687ec28f9 Mon Sep 17 00:00:00 2001 From: Nigel Tao Date: Thu, 18 Feb 2021 22:26:49 +1100 Subject: [PATCH] Add x86_sse42_utility.make_m128i_slice128 --- internal/cgen/builtin.go | 30 +++++-------------------- lang/builtin/builtin.go | 5 ++--- lang/check/bounds.go | 32 +++++++++------------------ lang/check/type.go | 5 ----- lang/token/list.go | 28 +++++++++-------------- std/adler32/common_up_x86_sse42.wuffs | 4 ++-- 6 files changed, 31 insertions(+), 73 deletions(-) diff --git a/internal/cgen/builtin.go b/internal/cgen/builtin.go index cbbbb1ae..dd8a322e 100644 --- a/internal/cgen/builtin.go +++ b/internal/cgen/builtin.go @@ -456,29 +456,6 @@ func (g *gen) writeBuiltinCPUArch(b *buffer, recv *a.Expr, method t.ID, args []* armNeon := recv.MType().Eq(typeExprARMNeon64) || recv.MType().Eq(typeExprARMNeon128) switch method { - case t.IDLoadSlice128: - if !sideEffectsOnly { - // Generate a two part expression using the comma operator: "(etc, - // return_empty_struct call)". The final part is a function call - // (to a static inline function) instead of a struct literal, to - // avoid a "expression result unused" compiler error. - b.writes("(") - } - if err := g.writeExpr(b, recv, false, depth); err != nil { - return err - } - - b.writes(" = _mm_lddqu_si128((const __m128i*)(const void*)(") - if err := g.writeExpr(b, args[0].AsArg().Value(), false, depth); err != nil { - return err - } - b.writes(".ptr))") - - if !sideEffectsOnly { - b.writes(", wuffs_base__make_empty_struct())") - } - return nil - case t.IDTruncateU32, t.IDTruncateU64, t.IDStoreSlice128: switch method { case t.IDTruncateU32: @@ -830,7 +807,7 @@ func (g *gen) writeBuiltinCPUArch(b *buffer, recv *a.Expr, method t.ID, args []* func (g *gen) writeBuiltinCPUArchX86(b *buffer, recv *a.Expr, method t.ID, args []*a.Node, sideEffectsOnly bool, depth uint32) error { methodStr := method.Str(g.tm) if strings.HasPrefix(methodStr, "make_") { - fName, tName := "", "" + fName, tName, ptr := "", "", false switch methodStr { case "make_m128i_multiple_u8": fName, tName = "_mm_set_epi8", "int8_t" @@ -852,6 +829,8 @@ func (g *gen) writeBuiltinCPUArchX86(b *buffer, recv *a.Expr, method t.ID, args fName, tName = "_mm_cvtsi32_si128", "int32_t" case "make_m128i_single_u64": fName, tName = "_mm_cvtsi64x_si128", "int64_t" + case "make_m128i_slice128": + fName, tName, ptr = "_mm_lddqu_si128", "const __m128i*)(const void*", true case "make_m128i_zeroes": fName, tName = "_mm_setzero_si128", "" default: @@ -869,6 +848,9 @@ func (g *gen) writeBuiltinCPUArchX86(b *buffer, recv *a.Expr, method t.ID, args if err := g.writeExpr(b, o.Value(), false, depth); err != nil { return err } + if ptr { + b.writes(".ptr") + } b.writes(")") } b.writes(")") diff --git a/lang/builtin/builtin.go b/lang/builtin/builtin.go index e9eb743b..d73ca44c 100644 --- a/lang/builtin/builtin.go +++ b/lang/builtin/builtin.go @@ -680,13 +680,12 @@ var Funcs = []string{ "x86_sse42_utility.make_m128i_single_u32(a: u32) x86_m128i", "x86_sse42_utility.make_m128i_single_u64(a: u64) x86_m128i", + "x86_sse42_utility.make_m128i_slice128(a: slice base.u8) x86_m128i", + "x86_sse42_utility.make_m128i_zeroes() x86_m128i", // ---- x86_m128i - "x86_m128i.load_u32!(a: u32)", - "x86_m128i.load_u64!(a: u64)", - "x86_m128i.load_slice128!(a: slice base.u8)", "x86_m128i.truncate_u32() u32", "x86_m128i.truncate_u64() u64", "x86_m128i.store_slice128!(a: slice base.u8)", diff --git a/lang/check/bounds.go b/lang/check/bounds.go index 553facd2..24df237a 100644 --- a/lang/check/bounds.go +++ b/lang/check/bounds.go @@ -19,6 +19,7 @@ import ( "fmt" "math/big" "strconv" + "strings" "github.com/google/wuffs/lib/interval" @@ -1220,9 +1221,16 @@ func (q *checker) bcheckExprCallSpecialCases(n *a.Expr, depth uint32) (bounds, e } } else if recvTyp.IsCPUArchType() { - if method >= t.IDLoadSlice128 { - if m := method - t.IDLoadSlice128; m < t.ID(len(lsMethodAdvances)) { - advance, update = lsMethodAdvances[m], false + if s := method.Str(q.tm); strings.HasPrefix(s, "make_") || strings.HasPrefix(s, "store_") { + switch { + case strings.HasSuffix(s, "_slice64"): // 64 bits is 8 bytes. + advance = eight + case strings.HasSuffix(s, "_slice128"): // 128 bits is 16 bytes. + advance = sixteen + case strings.HasSuffix(s, "_slice256"): // 256 bits is 32 bytes. + advance = thirtyTwo + case strings.HasSuffix(s, "_slice512"): // 512 bits is 64 bytes. + advance = sixtyFour } } } @@ -1439,24 +1447,6 @@ var ioMethodAdvances = [...]struct { t.IDWriteExtendedTokenFast - t.IDPeekU8: {one, true}, } -var lsMethodAdvances = [...]*big.Int{ - // 64 bits is 8 bytes. - // 128 bits is 16 bytes. - // 256 bits is 32 bytes. - // 512 bits is 64 bytes. - - t.IDLoadSlice128 - t.IDLoadSlice128: sixteen, - t.IDLoadSlice256 - t.IDLoadSlice128: thirtyTwo, - t.IDLoadSlice512 - t.IDLoadSlice128: sixtyFour, - - t.IDStoreSlice128 - t.IDLoadSlice128: sixteen, - t.IDStoreSlice256 - t.IDLoadSlice128: thirtyTwo, - t.IDStoreSlice512 - t.IDLoadSlice128: sixtyFour, - - t.IDCreateSlice64 - t.IDLoadSlice128: eight, - t.IDCreateSlice128 - t.IDLoadSlice128: sixteen, -} - func makeConstValueExpr(tm *t.Map, cv *big.Int) (*a.Expr, error) { id, err := tm.Insert(cv.String()) if err != nil { diff --git a/lang/check/type.go b/lang/check/type.go index b9fcacf6..8b53343f 100644 --- a/lang/check/type.go +++ b/lang/check/type.go @@ -607,11 +607,6 @@ func (q *checker) tcheckExprCall(n *a.Expr, depth uint32) error { } genericType1 = a.NewTypeExpr(t.IDSlice, 0, 0, nil, nil, genericType2.Inner()) } - - if f.FuncName().IsBuiltInLoad() && (lhs.LHS().AsExpr().Operator() != 0) { - return fmt.Errorf(`check: %q receiver %q must be a local variable`, - f.QQID().Str(q.tm), lhs.LHS().AsExpr().Str(q.tm)) - } } // Check that the func's in type matches the arguments. diff --git a/lang/token/list.go b/lang/token/list.go index 4ffd216a..6773914e 100644 --- a/lang/token/list.go +++ b/lang/token/list.go @@ -115,7 +115,6 @@ func (x ID) IsTightRight() bool { return x < ID(len(isTightRight)) && isTightRig func (x ID) IsAssign() bool { return minAssign <= x && x <= maxAssign } func (x ID) IsBuiltInCPUArch() bool { return minBuiltInCPUArch <= x && x <= maxBuiltInCPUArch } -func (x ID) IsBuiltInLoad() bool { return minBuiltInLoad <= x && x <= maxBuiltInLoad } func (x ID) IsCannotAssignTo() bool { return minCannotAssignTo <= x && x <= maxCannotAssignTo } func (x ID) IsClose() bool { return minClose <= x && x <= maxClose } func (x ID) IsKeyword() bool { return minKeyword <= x && x <= maxKeyword } @@ -699,18 +698,13 @@ const ( // -------- - minBuiltInLoad = 0x380 - maxBuiltInLoad = 0x387 + IDStoreSlice64 = ID(0x380) + IDStoreSlice128 = ID(0x381) + IDStoreSlice256 = ID(0x382) + IDStoreSlice512 = ID(0x383) - IDLoadSlice128 = ID(0x382) - IDLoadSlice256 = ID(0x383) - IDLoadSlice512 = ID(0x384) - - IDTruncateU32 = ID(0x388) - IDTruncateU64 = ID(0x389) - IDStoreSlice128 = ID(0x38A) - IDStoreSlice256 = ID(0x38B) - IDStoreSlice512 = ID(0x38C) + IDTruncateU32 = ID(0x388) + IDTruncateU64 = ID(0x389) IDCreateSlice64 = ID(0x390) IDCreateSlice128 = ID(0x391) @@ -1117,16 +1111,14 @@ var builtInsByID = [nBuiltInIDs]string{ IDX86M128I: "x86_m128i", - IDLoadSlice128: "load_slice128", - IDLoadSlice256: "load_slice256", - IDLoadSlice512: "load_slice512", - - IDTruncateU32: "truncate_u32", - IDTruncateU64: "truncate_u64", + IDStoreSlice64: "store_slice64", IDStoreSlice128: "store_slice128", IDStoreSlice256: "store_slice256", IDStoreSlice512: "store_slice512", + IDTruncateU32: "truncate_u32", + IDTruncateU64: "truncate_u64", + IDCreateSlice64: "create_slice64", IDCreateSlice128: "create_slice128", } diff --git a/std/adler32/common_up_x86_sse42.wuffs b/std/adler32/common_up_x86_sse42.wuffs index 239790e8..93c5c1b3 100644 --- a/std/adler32/common_up_x86_sse42.wuffs +++ b/std/adler32/common_up_x86_sse42.wuffs @@ -97,8 +97,8 @@ pri func hasher.up_x86_sse42!(x: slice base.u8), // // Let p__left = [u8×16: p00, p01, p02, ..., p15] // Let p_right = [u8×16: p16, p17, p18, ..., p31] - p__left.load_slice128!(a: p[.. 16]) - p_right.load_slice128!(a: p[16 .. 32]) + p__left = util.make_m128i_slice128(a: p[.. 16]) + p_right = util.make_m128i_slice128(a: p[16 .. 32]) // For v2j, we need to calculate the sums of the s1j terms for each // of p's 32 elements. This is simply 32 times the same number,