T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

aarch64: Use smaller instruction helpers in ISLE (#3618)

Browse Source 
* aarch64: Use smaller instruction helpers in ISLE

This commit moves the aarch64 backend's ISLE to be more similar to the
x64 backend's ISLE where one-liner instruction builders are used for
various forms of instructions instead of always using the
constructor-per-variant-of-`Inst`. Overall I think this change worked
out quite well and sets up some naming idioms as well for various forms
of instructions.

* rebase conflict
This commit is contained in:
Alex Crichton
2021-12-17 17:28:52 -06:00
committed by GitHub
parent e10171b7ea
commit 546e901d32
4 changed files with 2318 additions and 1141 deletions
Download Patch File Download Diff File Expand all files Collapse all files

284
cranelift/codegen/src/isa/aarch64/lower.isle
View File

@@ -28,67 +28,52 @@
;; Base case, simply adding things in registers.
(rule (lower (has_type (fits_in_64 ty) (iadd x y)))
(value_reg (alu_rrr (iadd_op ty) (put_in_reg x) (put_in_reg y))))
(value_reg (add ty (put_in_reg x) (put_in_reg y))))
;; Special cases for when one operand is an immediate that fits in 12 bits.
(rule (lower (has_type (fits_in_64 ty) (iadd x (imm12_from_value y))))
(value_reg (alu_rr_imm12 (iadd_op ty) (put_in_reg x) y)))
(value_reg (add_imm ty (put_in_reg x) y)))
(rule (lower (has_type (fits_in_64 ty) (iadd (imm12_from_value x) y)))
(value_reg (alu_rr_imm12 (iadd_op ty) (put_in_reg y) x)))
(value_reg (add_imm ty (put_in_reg y) x)))
;; Same as the previous special cases, except we can switch the addition to a
;; subtraction if the negated immediate fits in 12 bits.
(rule (lower (has_type (fits_in_64 ty) (iadd x (imm12_from_negated_value y))))
(value_reg (alu_rr_imm12 (isub_op ty) (put_in_reg x) y)))
(value_reg (sub_imm ty (put_in_reg x) y)))
(rule (lower (has_type (fits_in_64 ty) (iadd (imm12_from_negated_value x) y)))
(value_reg (alu_rr_imm12 (isub_op ty) (put_in_reg y) x)))
(value_reg (sub_imm ty (put_in_reg y) x)))
;; Special cases for when we're adding an extended register where the extending
;; operation can get folded into the add itself.
(rule (lower (has_type (fits_in_64 ty) (iadd x (extended_value_from_value y))))
(value_reg (alu_rr_extend_reg (iadd_op ty) (put_in_reg x) y)))
(value_reg (add_extend ty (put_in_reg x) y)))
(rule (lower (has_type (fits_in_64 ty) (iadd (extended_value_from_value x) y)))
(value_reg (alu_rr_extend_reg (iadd_op ty) (put_in_reg y) x)))
(value_reg (add_extend ty (put_in_reg y) x)))
;; Special cases for when we're adding the shift of a different
;; register by a constant amount and the shift can get folded into the add.
(rule (lower (has_type (fits_in_64 ty)
(iadd x (def_inst (ishl y (def_inst (iconst (lshl_from_imm64 <ty amt))))))))
(value_reg (alu_rrr_shift (iadd_op ty) (put_in_reg x) (put_in_reg y) amt)))
(value_reg (add_shift ty (put_in_reg x) (put_in_reg y) amt)))
(rule (lower (has_type (fits_in_64 ty)
(iadd (def_inst (ishl x (def_inst (iconst (lshl_from_imm64 <ty amt))))) y)))
(value_reg (alu_rrr_shift (iadd_op ty) (put_in_reg y) (put_in_reg x) amt)))
(value_reg (add_shift ty (put_in_reg y) (put_in_reg x) amt)))
;; Fold an `iadd` and `imul` combination into a `madd` instruction.
(rule (lower (has_type (fits_in_64 ty) (iadd x (def_inst (imul y z)))))
(value_reg (alu_rrrr (madd_op ty) (put_in_reg y) (put_in_reg z) (put_in_reg x))))
(value_reg (madd ty (put_in_reg y) (put_in_reg z) (put_in_reg x))))
(rule (lower (has_type (fits_in_64 ty) (iadd (def_inst (imul x y)) z)))
(value_reg (alu_rrrr (madd_op ty) (put_in_reg x) (put_in_reg y) (put_in_reg z))))
;; Helper to use either a 32 or 64-bit add depending on the input type.
(decl iadd_op (Type) ALUOp)
(rule (iadd_op (fits_in_32 _ty)) (ALUOp.Add32))
(rule (iadd_op $I64) (ALUOp.Add64))
;; Helper to use either a 32 or 64-bit sub depending on the input type.
(decl isub_op (Type) ALUOp)
(rule (isub_op (fits_in_32 _ty)) (ALUOp.Sub32))
(rule (isub_op $I64) (ALUOp.Sub64))
;; Helper to use either a 32 or 64-bit madd depending on the input type.
(decl madd_op (Type) ALUOp3)
(rule (madd_op (fits_in_32 _ty)) (ALUOp3.MAdd32))
(rule (madd_op $I64) (ALUOp3.MAdd64))
(value_reg (madd ty (put_in_reg x) (put_in_reg y) (put_in_reg z))))
;; vectors
(rule (lower (has_type ty @ (multi_lane _ _) (iadd x y)))
(value_reg (vec_rrr (VecALUOp.Add) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (add_vec (put_in_reg x) (put_in_reg y) (vector_size ty))))
;; `i128`
(rule (lower (has_type $I128 (iadd x y)))
@@ -115,31 +100,31 @@
;; Base case, simply subtracting things in registers.
(rule (lower (has_type (fits_in_64 ty) (isub x y)))
(value_reg (alu_rrr (isub_op ty) (put_in_reg x) (put_in_reg y))))
(value_reg (sub ty (put_in_reg x) (put_in_reg y))))
;; Special case for when one operand is an immediate that fits in 12 bits.
(rule (lower (has_type (fits_in_64 ty) (isub x (imm12_from_value y))))
(value_reg (alu_rr_imm12 (isub_op ty) (put_in_reg x) y)))
(value_reg (sub_imm ty (put_in_reg x) y)))
;; Same as the previous special case, except we can switch the subtraction to an
;; addition if the negated immediate fits in 12 bits.
(rule (lower (has_type (fits_in_64 ty) (isub x (imm12_from_negated_value y))))
(value_reg (alu_rr_imm12 (iadd_op ty) (put_in_reg x) y)))
(value_reg (add_imm ty (put_in_reg x) y)))
;; Special cases for when we're subtracting an extended register where the
;; extending operation can get folded into the sub itself.
(rule (lower (has_type (fits_in_64 ty) (isub x (extended_value_from_value y))))
(value_reg (alu_rr_extend_reg (isub_op ty) (put_in_reg x) y)))
(value_reg (sub_extend ty (put_in_reg x) y)))
;; Finally a special case for when we're subtracting the shift of a different
;; register by a constant amount and the shift can get folded into the sub.
(rule (lower (has_type (fits_in_64 ty)
(isub x (def_inst (ishl y (def_inst (iconst (lshl_from_imm64 <ty amt))))))))
(value_reg (alu_rrr_shift (isub_op ty) (put_in_reg x) (put_in_reg y) amt)))
(value_reg (sub_shift ty (put_in_reg x) (put_in_reg y) amt)))
;; vectors
(rule (lower (has_type ty @ (multi_lane _ _) (isub x y)))
(value_reg (vec_rrr (VecALUOp.Sub) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (sub_vec (put_in_reg x) (put_in_reg y) (vector_size ty))))
;; `i128`
(rule (lower (has_type $I128 (isub x y)))
@@ -163,38 +148,38 @@
;;;; Rules for `uadd_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type (vec128 ty) (uadd_sat x y)))
(value_reg (vec_rrr (VecALUOp.Uqadd) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (uqadd (put_in_reg x) (put_in_reg y) (vector_size ty))))
;;;; Rules for `sadd_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type (vec128 ty) (sadd_sat x y)))
(value_reg (vec_rrr (VecALUOp.Sqadd) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (sqadd (put_in_reg x) (put_in_reg y) (vector_size ty))))
;;;; Rules for `usub_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type (vec128 ty) (usub_sat x y)))
(value_reg (vec_rrr (VecALUOp.Uqsub) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (uqsub (put_in_reg x) (put_in_reg y) (vector_size ty))))
;;;; Rules for `ssub_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type (vec128 ty) (ssub_sat x y)))
(value_reg (vec_rrr (VecALUOp.Sqsub) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (sqsub (put_in_reg x) (put_in_reg y) (vector_size ty))))
;;;; Rules for `ineg` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; `i64` and smaller.
(rule (lower (has_type (fits_in_64 ty) (ineg x)))
(value_reg (alu_rrr (isub_op ty) (zero_reg) (put_in_reg x))))
(value_reg (sub ty (zero_reg) (put_in_reg x))))
;; vectors.
(rule (lower (has_type (vec128 ty) (ineg x)))
(value_reg (vec_misc (VecMisc2.Neg) (put_in_reg x) (vector_size ty))))
(value_reg (neg (put_in_reg x) (vector_size ty))))
;;;; Rules for `imul` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; `i64` and smaller.
(rule (lower (has_type (fits_in_64 ty) (imul x y)))
(value_reg (alu_rrrr (madd_op ty) (put_in_reg x) (put_in_reg y) (zero_reg))))
(value_reg (madd ty (put_in_reg x) (put_in_reg y) (zero_reg))))
;; `i128`.
(rule (lower (has_type $I128 (imul x y)))
@@ -218,16 +203,16 @@
;; madd dst_hi, x_lo, y_hi, dst_hi
;; madd dst_hi, x_hi, y_lo, dst_hi
;; madd dst_lo, x_lo, y_lo, zero
(dst_hi1 Reg (alu_rrr (ALUOp.UMulH) x_lo y_lo))
(dst_hi2 Reg (alu_rrrr (ALUOp3.MAdd64) x_lo y_hi dst_hi1))
(dst_hi Reg (alu_rrrr (ALUOp3.MAdd64) x_hi y_lo dst_hi2))
(dst_lo Reg (alu_rrrr (ALUOp3.MAdd64) x_lo y_lo (zero_reg)))
(dst_hi1 Reg (umulh x_lo y_lo))
(dst_hi2 Reg (madd64 x_lo y_hi dst_hi1))
(dst_hi Reg (madd64 x_hi y_lo dst_hi2))
(dst_lo Reg (madd64 x_lo y_lo (zero_reg)))
)
(value_regs dst_lo dst_hi)))
;; Case for i8x16, i16x8, and i32x4.
(rule (lower (has_type (vec128 ty @ (not_i64x2)) (imul x y)))
(value_reg (vec_rrr (VecALUOp.Mul) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (mul (put_in_reg x) (put_in_reg y) (vector_size ty))))
;; Special lowering for i64x2.
;;
@@ -264,7 +249,7 @@
(rm Reg (put_in_reg y))
;; Reverse the 32-bit elements in the 64-bit words.
;; rd = |g|h|e|f|
(rev Reg (vec_misc (VecMisc2.Rev64) rm (VectorSize.Size32x4)))
(rev Reg (rev64 rm (VectorSize.Size32x4)))
;; Calculate the high half components.
;; rd = |dg|ch|be|af|
@@ -273,28 +258,28 @@
;; discards the bits that would overflow, same as
;; if 64-bit operations were used. Also the Shll
;; below would shift out the overflow bits anyway.
(mul Reg (vec_rrr (VecALUOp.Mul) rev rn (VectorSize.Size32x4)))
(mul Reg (mul rev rn (VectorSize.Size32x4)))
;; Extract the low half components of rn.
;; tmp1 = |c|a|
(tmp1 Reg (vec_rr_narrow (VecRRNarrowOp.Xtn64) rn $false))
(tmp1 Reg (xtn64 rn $false))
;; Sum the respective high half components.
;; rd = |dg+ch|be+af||dg+ch|be+af|
(sum Reg (vec_rrr (VecALUOp.Addp) mul mul (VectorSize.Size32x4)))
(sum Reg (addp mul mul (VectorSize.Size32x4)))
;; Extract the low half components of rm.
;; tmp2 = |g|e|
(tmp2 Reg (vec_rr_narrow (VecRRNarrowOp.Xtn64) rm $false))
(tmp2 Reg (xtn64 rm $false))
;; Shift the high half components, into the high half.
;; rd = |dg+ch << 32|be+af << 32|
(shift Reg (vec_rr_long (VecRRLongOp.Shll32) sum $false))
(shift Reg (shll32 sum $false))
;; Multiply the low components together, and accumulate with the high
;; half.
;; rd = |rd[1] + cg|rd[0] + ae|
(result Reg (vec_rrrr_long (VecRRRLongOp.Umlal32) shift tmp2 tmp1 $false))
(result Reg (umlal32 shift tmp2 tmp1 $false))
)
(value_reg result)))
@@ -302,99 +287,99 @@
(rule (lower (has_type $I16X8
(imul (def_inst (swiden_low x @ (value_type $I8X16)))
(def_inst (swiden_low y @ (value_type $I8X16))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Smull8) (put_in_reg x) (put_in_reg y) $false)))
(value_reg (smull8 (put_in_reg x) (put_in_reg y) $false)))
;; Special case for `i16x8.extmul_high_i8x16_s`.
(rule (lower (has_type $I16X8
(imul (def_inst (swiden_high x @ (value_type $I8X16)))
(def_inst (swiden_high y @ (value_type $I8X16))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Smull8) (put_in_reg x) (put_in_reg y) $true)))
(value_reg (smull8 (put_in_reg x) (put_in_reg y) $true)))
;; Special case for `i16x8.extmul_low_i8x16_u`.
(rule (lower (has_type $I16X8
(imul (def_inst (uwiden_low x @ (value_type $I8X16)))
(def_inst (uwiden_low y @ (value_type $I8X16))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Umull8) (put_in_reg x) (put_in_reg y) $false)))
(value_reg (umull8 (put_in_reg x) (put_in_reg y) $false)))
;; Special case for `i16x8.extmul_high_i8x16_u`.
(rule (lower (has_type $I16X8
(imul (def_inst (uwiden_high x @ (value_type $I8X16)))
(def_inst (uwiden_high y @ (value_type $I8X16))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Umull8) (put_in_reg x) (put_in_reg y) $true)))
(value_reg (umull8 (put_in_reg x) (put_in_reg y) $true)))
;; Special case for `i32x4.extmul_low_i16x8_s`.
(rule (lower (has_type $I32X4
(imul (def_inst (swiden_low x @ (value_type $I16X8)))
(def_inst (swiden_low y @ (value_type $I16X8))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Smull16) (put_in_reg x) (put_in_reg y) $false)))
(value_reg (smull16 (put_in_reg x) (put_in_reg y) $false)))
;; Special case for `i32x4.extmul_high_i16x8_s`.
(rule (lower (has_type $I32X4
(imul (def_inst (swiden_high x @ (value_type $I16X8)))
(def_inst (swiden_high y @ (value_type $I16X8))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Smull16) (put_in_reg x) (put_in_reg y) $true)))
(value_reg (smull16 (put_in_reg x) (put_in_reg y) $true)))
;; Special case for `i32x4.extmul_low_i16x8_u`.
(rule (lower (has_type $I32X4
(imul (def_inst (uwiden_low x @ (value_type $I16X8)))
(def_inst (uwiden_low y @ (value_type $I16X8))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Umull16) (put_in_reg x) (put_in_reg y) $false)))
(value_reg (umull16 (put_in_reg x) (put_in_reg y) $false)))
;; Special case for `i32x4.extmul_high_i16x8_u`.
(rule (lower (has_type $I32X4
(imul (def_inst (uwiden_high x @ (value_type $I16X8)))
(def_inst (uwiden_high y @ (value_type $I16X8))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Umull16) (put_in_reg x) (put_in_reg y) $true)))
(value_reg (umull16 (put_in_reg x) (put_in_reg y) $true)))
;; Special case for `i64x2.extmul_low_i32x4_s`.
(rule (lower (has_type $I64X2
(imul (def_inst (swiden_low x @ (value_type $I32X4)))
(def_inst (swiden_low y @ (value_type $I32X4))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Smull32) (put_in_reg x) (put_in_reg y) $false)))
(value_reg (smull32 (put_in_reg x) (put_in_reg y) $false)))
;; Special case for `i64x2.extmul_high_i32x4_s`.
(rule (lower (has_type $I64X2
(imul (def_inst (swiden_high x @ (value_type $I32X4)))
(def_inst (swiden_high y @ (value_type $I32X4))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Smull32) (put_in_reg x) (put_in_reg y) $true)))
(value_reg (smull32 (put_in_reg x) (put_in_reg y) $true)))
;; Special case for `i64x2.extmul_low_i32x4_u`.
(rule (lower (has_type $I64X2
(imul (def_inst (uwiden_low x @ (value_type $I32X4)))
(def_inst (uwiden_low y @ (value_type $I32X4))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Umull32) (put_in_reg x) (put_in_reg y) $false)))
(value_reg (umull32 (put_in_reg x) (put_in_reg y) $false)))
;; Special case for `i64x2.extmul_high_i32x4_u`.
(rule (lower (has_type $I64X2
(imul (def_inst (uwiden_high x @ (value_type $I32X4)))
(def_inst (uwiden_high y @ (value_type $I32X4))))))
(value_reg (vec_rrr_long (VecRRRLongOp.Umull32) (put_in_reg x) (put_in_reg y) $true)))
(value_reg (umull32 (put_in_reg x) (put_in_reg y) $true)))
;;;; Rules for `smulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $I64 (smulhi x y)))
(value_reg (alu_rrr (ALUOp.SMulH) (put_in_reg x) (put_in_reg y))))
(value_reg (smulh (put_in_reg x) (put_in_reg y))))
(rule (lower (has_type (fits_in_32 ty) (smulhi x y)))
(let (
(x64 Reg (put_in_reg_sext64 x))
(y64 Reg (put_in_reg_sext64 y))
(mul Reg (alu_rrrr (ALUOp3.MAdd64) x64 y64 (zero_reg)))
(result Reg (alu_rr_imm_shift (ALUOp.Asr64) mul (imm_shift_from_u8 (ty_bits ty))))
(mul Reg (madd64 x64 y64 (zero_reg)))
(result Reg (asr64_imm mul (imm_shift_from_u8 (ty_bits ty))))
)
(value_reg result)))
;;;; Rules for `umulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $I64 (umulhi x y)))
(value_reg (alu_rrr (ALUOp.UMulH) (put_in_reg x) (put_in_reg y))))
(value_reg (umulh (put_in_reg x) (put_in_reg y))))
(rule (lower (has_type (fits_in_32 ty) (umulhi x y)))
(let (
(x64 Reg (put_in_reg_zext64 x))
(y64 Reg (put_in_reg_zext64 y))
(mul Reg (alu_rrrr (ALUOp3.MAdd64) x64 y64 (zero_reg)))
(result Reg (alu_rr_imm_shift (ALUOp.Lsr64) mul (imm_shift_from_u8 (ty_bits ty))))
(mul Reg (madd64 x64 y64 (zero_reg)))
(result Reg (lsr64_imm mul (imm_shift_from_u8 (ty_bits ty))))
)
(value_reg result)))
@@ -406,9 +391,7 @@
;; Note that aarch64's `udiv` doesn't trap so to respect the semantics of
;; CLIF's `udiv` the check for zero needs to be manually performed.
(rule (lower (has_type (fits_in_64 ty) (udiv x y)))
(value_reg (alu_rrr (ALUOp.UDiv64)
(put_in_reg_zext64 x)
(put_nonzero_in_reg_zext64 y))))
(value_reg (udiv64 (put_in_reg_zext64 x) (put_nonzero_in_reg_zext64 y))))
;; Helper for placing a `Value` into a `Reg` and validating that it's nonzero.
(decl put_nonzero_in_reg_zext64 (Value) Reg)
@@ -445,7 +428,7 @@
(x64 Reg (put_in_reg_sext64 x))
(y64 Reg (put_nonzero_in_reg_sext64 y))
(valid_x64 Reg (trap_if_div_overflow ty x64 y64))
(result Reg (alu_rrr (ALUOp.SDiv64) valid_x64 y64))
(result Reg (sdiv64 valid_x64 y64))
)
(value_reg result)))
@@ -456,9 +439,7 @@
;; Special case for `sdiv` where no checks are needed due to division by a
;; constant meaning the checks are always passed.
(rule (lower (has_type (fits_in_64 ty) (sdiv x (def_inst (iconst (safe_divisor_from_imm64 y))))))
(value_reg (alu_rrr (ALUOp.SDiv64)
(put_in_reg_sext64 x)
(imm ty y))))
(value_reg (sdiv64 (put_in_reg_sext64 x) (imm ty y))))
;; Helper for placing a `Value` into a `Reg` and validating that it's nonzero.
(decl put_nonzero_in_reg_sext64 (Value) Reg)
@@ -489,8 +470,8 @@
(let (
(x64 Reg (put_in_reg_zext64 x))
(y64 Reg (put_nonzero_in_reg_zext64 y))
(div Reg (alu_rrr (ALUOp.UDiv64) x64 y64))
(result Reg (alu_rrrr (ALUOp3.MSub64) div y64 x64))
(div Reg (udiv64 x64 y64))
(result Reg (msub64 div y64 x64))
)
(value_reg result)))
@@ -498,8 +479,8 @@
(let (
(x64 Reg (put_in_reg_sext64 x))
(y64 Reg (put_nonzero_in_reg_sext64 y))
(div Reg (alu_rrr (ALUOp.SDiv64) x64 y64))
(result Reg (alu_rrrr (ALUOp3.MSub64) div y64 x64))
(div Reg (sdiv64 x64 y64))
(result Reg (msub64 div y64 x64))
)
(value_reg result)))
@@ -556,7 +537,7 @@
(rule (lower (has_type $I128 (sextend x)))
(let (
(lo Reg (put_in_reg_sext64 x))
(hi Reg (alu_rr_imm_shift (ALUOp.Asr64) lo (imm_shift_from_u8 63)))
(hi Reg (asr64_imm lo (imm_shift_from_u8 63)))
)
(value_regs lo hi)))
@@ -573,7 +554,7 @@
lane
(vector_size in)
(size_from_ty $I64)))
(hi Reg (alu_rr_imm_shift (ALUOp.Asr64) lo (imm_shift_from_u8 63)))
(hi Reg (asr64_imm lo (imm_shift_from_u8 63)))
)
(value_regs lo hi)))
@@ -585,29 +566,25 @@
(lo Reg (mov_from_vec (put_in_reg vec)
lane
(VectorSize.Size64x2)))
(hi Reg (alu_rr_imm_shift (ALUOp.Asr64) lo (imm_shift_from_u8 63)))
(hi Reg (asr64_imm lo (imm_shift_from_u8 63)))
)
(value_regs lo hi)))
;;;; Rules for `bnot` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(decl orr_not_op (Type) ALUOp)
(rule (orr_not_op (fits_in_32 _ty)) (ALUOp.OrrNot32))
(rule (orr_not_op $I64) (ALUOp.OrrNot64))
;; Base case using `orn` between two registers.
;;
;; Note that bitwise negation is implemented here as
;;
;; NOT rd, rm ==> ORR_NOT rd, zero, rm
(rule (lower (has_type (fits_in_64 ty) (bnot x)))
(value_reg (alu_rrr (orr_not_op ty) (zero_reg) (put_in_reg x))))
(value_reg (orr_not ty (zero_reg) (put_in_reg x))))
;; Special case to use `AluRRRShift` if it's a `bnot` of a const-left-shifted
;; Special case to use `orr_not_shift` if it's a `bnot` of a const-left-shifted
;; value.
(rule (lower (has_type (fits_in_64 ty)
(bnot (def_inst (ishl x (def_inst (iconst (lshl_from_imm64 <ty amt))))))))
(value_reg (alu_rrr_shift (orr_not_op ty) (zero_reg) (put_in_reg x) amt)))
(value_reg (orr_not_shift ty (zero_reg) (put_in_reg x) amt)))
;; Implementation of `bnot` for `i128`.
(rule (lower (has_type $I128 (bnot x)))
@@ -615,14 +592,14 @@
(x_regs ValueRegs (put_in_regs x))
(x_lo Reg (value_regs_get x_regs 0))
(x_hi Reg (value_regs_get x_regs 1))
(new_lo Reg (alu_rrr (ALUOp.OrrNot64) (zero_reg) x_lo))
(new_hi Reg (alu_rrr (ALUOp.OrrNot64) (zero_reg) x_hi))
(new_lo Reg (orr_not64 (zero_reg) x_lo))
(new_hi Reg (orr_not64 (zero_reg) x_hi))
)
(value_regs new_lo new_hi)))
;; Implementation of `bnot` for vector types.
(rule (lower (has_type (vec128 ty) (bnot x)))
(value_reg (vec_misc (VecMisc2.Not) (put_in_reg x) (vector_size ty))))
(value_reg (not (put_in_reg x) (vector_size ty))))
;;;; Rules for `band` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -635,7 +612,7 @@
(rule (lower (has_type $I128 (band x y))) (i128_alu_bitop (ALUOp.And64) x y))
(rule (lower (has_type (vec128 ty) (band x y)))
(value_reg (vec_rrr (VecALUOp.And) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (and_vec (put_in_reg x) (put_in_reg y) (vector_size ty))))
;;;; Rules for `bor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -648,7 +625,7 @@
(rule (lower (has_type $I128 (bor x y))) (i128_alu_bitop (ALUOp.Orr64) x y))
(rule (lower (has_type (vec128 ty) (bor x y)))
(value_reg (vec_rrr (VecALUOp.Orr) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (orr_vec (put_in_reg x) (put_in_reg y) (vector_size ty))))
;;;; Rules for `bxor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -661,7 +638,7 @@
(rule (lower (has_type $I128 (bxor x y))) (i128_alu_bitop (ALUOp.Eor64) x y))
(rule (lower (has_type (vec128 ty) (bxor x y)))
(value_reg (vec_rrr (VecALUOp.Eor) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (eor_vec (put_in_reg x) (put_in_reg y) (vector_size ty))))
;;;; Rules for `band_not` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -674,7 +651,7 @@
(rule (lower (has_type $I128 (band_not x y))) (i128_alu_bitop (ALUOp.AndNot64) x y))
(rule (lower (has_type (vec128 ty) (band_not x y)))
(value_reg (vec_rrr (VecALUOp.Bic) (put_in_reg x) (put_in_reg y) (vector_size ty))))
(value_reg (bic_vec (put_in_reg x) (put_in_reg y) (vector_size ty))))
;;;; Rules for `bor_not` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -724,15 +701,12 @@
(let (
(src_lo Reg (value_regs_get src 0))
(src_hi Reg (value_regs_get src 1))
(lo_lshift Reg (alu_rrr (ALUOp.Lsl64) src_lo amt))
(hi_lshift Reg (alu_rrr (ALUOp.Lsl64) src_hi amt))
(inv_amt Reg (alu_rrr (ALUOp.OrrNot32) (zero_reg) amt))
(lo_rshift Reg (alu_rrr (ALUOp.Lsr64)
(alu_rr_imm_shift (ALUOp.Lsr64)
src_lo
(imm_shift_from_u8 1))
inv_amt))
(maybe_hi Reg (alu_rrr (ALUOp.Orr64) hi_lshift lo_rshift))
(lo_lshift Reg (lsl64 src_lo amt))
(hi_lshift Reg (lsl64 src_hi amt))
(inv_amt Reg (orr_not32 (zero_reg) amt))
(lo_rshift Reg (lsr64 (lsr64_imm src_lo (imm_shift_from_u8 1))
inv_amt))
(maybe_hi Reg (orr64 hi_lshift lo_rshift))
)
(with_flags_2
(tst64_imm amt (u64_into_imm_logic $I64 64))
@@ -745,7 +719,7 @@
(size VectorSize (vector_size ty))
(shift Reg (vec_dup (put_in_reg y) size))
)
(value_reg (vec_rrr (VecALUOp.Sshl) (put_in_reg x) shift size))))
(value_reg (sshl (put_in_reg x) shift size))))
;; Helper function to emit a shift operation with the opcode specified and
;; the output type specified. The `Reg` provided is shifted by the `Value`
@@ -767,7 +741,7 @@
(rule (do_shift op (fits_in_16 ty) x y)
(let (
(shift_amt Reg (value_regs_get (put_in_regs y) 0))
(masked_shift_amt Reg (alu_rr_imm_logic (ALUOp.And32) shift_amt (shift_mask ty)))
(masked_shift_amt Reg (and32_imm shift_amt (shift_mask ty)))
)
(alu_rrr op x masked_shift_amt)))
@@ -805,9 +779,9 @@
(rule (lower (has_type (vec128 ty) (ushr x y)))
(let (
(size VectorSize (vector_size ty))
(shift Reg (vec_dup (alu_rrr (ALUOp.Sub32) (zero_reg) (put_in_reg y)) size))
(shift Reg (vec_dup (sub32 (zero_reg) (put_in_reg y)) size))
)
(value_reg (vec_rrr (VecALUOp.Ushl) (put_in_reg x) shift size))))
(value_reg (ushl (put_in_reg x) shift size))))
;; lsr lo_rshift, src_lo, amt
;; lsr hi_rshift, src_hi, amt
@@ -823,16 +797,13 @@
(let (
(src_lo Reg (value_regs_get src 0))
(src_hi Reg (value_regs_get src 1))
(lo_rshift Reg (alu_rrr (ALUOp.Lsr64) src_lo amt))
(hi_rshift Reg (alu_rrr (ALUOp.Lsr64) src_hi amt))
(lo_rshift Reg (lsr64 src_lo amt))
(hi_rshift Reg (lsr64 src_hi amt))
(inv_amt Reg (alu_rrr (ALUOp.OrrNot32) (zero_reg) amt))
(hi_lshift Reg (alu_rrr (ALUOp.Lsl64)
(alu_rr_imm_shift (ALUOp.Lsl64)
src_hi
(imm_shift_from_u8 1))
inv_amt))
(maybe_lo Reg (alu_rrr (ALUOp.Orr64) lo_rshift hi_lshift))
(inv_amt Reg (orr_not32 (zero_reg) amt))
(hi_lshift Reg (lsl64 (lsl64_imm src_hi (imm_shift_from_u8 1))
inv_amt))
(maybe_lo Reg (orr64 lo_rshift hi_lshift))
)
(with_flags_2
(tst64_imm amt (u64_into_imm_logic $I64 64))
@@ -859,9 +830,9 @@
(rule (lower (has_type (vec128 ty) (sshr x y)))
(let (
(size VectorSize (vector_size ty))
(shift Reg (vec_dup (alu_rrr (ALUOp.Sub32) (zero_reg) (put_in_reg y)) size))
(shift Reg (vec_dup (sub32 (zero_reg) (put_in_reg y)) size))
)
(value_reg (vec_rrr (VecALUOp.Sshl) (put_in_reg x) shift size))))
(value_reg (sshl (put_in_reg x) shift size))))
;; lsr lo_rshift, src_lo, amt
;; asr hi_rshift, src_hi, amt
@@ -878,17 +849,14 @@
(let (
(src_lo Reg (value_regs_get src 0))
(src_hi Reg (value_regs_get src 1))
(lo_rshift Reg (alu_rrr (ALUOp.Lsr64) src_lo amt))
(hi_rshift Reg (alu_rrr (ALUOp.Asr64) src_hi amt))
(lo_rshift Reg (lsr64 src_lo amt))
(hi_rshift Reg (asr64 src_hi amt))
(inv_amt Reg (alu_rrr (ALUOp.OrrNot32) (zero_reg) amt))
(hi_lshift Reg (alu_rrr (ALUOp.Lsl64)
(alu_rr_imm_shift (ALUOp.Lsl64)
src_hi
(imm_shift_from_u8 1))
inv_amt))
(hi_sign Reg (alu_rr_imm_shift (ALUOp.Asr64) src_hi (imm_shift_from_u8 63)))
(maybe_lo Reg (alu_rrr (ALUOp.Orr64) lo_rshift hi_lshift))
(inv_amt Reg (orr_not32 (zero_reg) amt))
(hi_lshift Reg (lsl64 (lsl64_imm src_hi (imm_shift_from_u8 1))
inv_amt))
(hi_sign Reg (asr64_imm src_hi (imm_shift_from_u8 63)))
(maybe_lo Reg (orr64 lo_rshift hi_lshift))
)
(with_flags_2
(tst64_imm amt (u64_into_imm_logic $I64 64))
@@ -899,7 +867,7 @@
;; General 8/16-bit case.
(rule (lower (has_type (fits_in_16 ty) (rotl x y)))
(let ((neg_shift Reg (alu_rrr (ALUOp.Sub32) (zero_reg) (put_in_reg y))))
(let ((neg_shift Reg (sub32 (zero_reg) (put_in_reg y))))
(value_reg (small_rotr ty (put_in_reg_zext32 x) neg_shift))))
;; Specialization for the 8/16-bit case when the rotation amount is an immediate.
@@ -916,21 +884,21 @@
;; General 32-bit case.
(rule (lower (has_type $I32 (rotl x y)))
(let ((neg_shift Reg (alu_rrr (ALUOp.Sub32) (zero_reg) (put_in_reg y))))
(value_reg (alu_rrr (ALUOp.RotR32) (put_in_reg x) neg_shift))))
(let ((neg_shift Reg (sub32 (zero_reg) (put_in_reg y))))
(value_reg (rotr32 (put_in_reg x) neg_shift))))
;; General 64-bit case.
(rule (lower (has_type $I64 (rotl x y)))
(let ((neg_shift Reg (alu_rrr (ALUOp.Sub64) (zero_reg) (put_in_reg y))))
(value_reg (alu_rrr (ALUOp.RotR64) (put_in_reg x) neg_shift))))
(let ((neg_shift Reg (sub64 (zero_reg) (put_in_reg y))))
(value_reg (rotr64 (put_in_reg x) neg_shift))))
;; Specialization for the 32-bit case when the rotation amount is an immediate.
(rule (lower (has_type $I32 (rotl x (def_inst (iconst (imm_shift_from_imm64 <$I32 n))))))
(value_reg (alu_rr_imm_shift (ALUOp.RotR32) (put_in_reg x) (negate_imm_shift $I32 n))))
(value_reg (rotr32_imm (put_in_reg x) (negate_imm_shift $I32 n))))
;; Specialization for the 64-bit case when the rotation amount is an immediate.
(rule (lower (has_type $I64 (rotl x (def_inst (iconst (imm_shift_from_imm64 <$I64 n))))))
(value_reg (alu_rr_imm_shift (ALUOp.RotR64) (put_in_reg x) (negate_imm_shift $I64 n))))
(value_reg (rotr64_imm (put_in_reg x) (negate_imm_shift $I64 n))))
(decl negate_imm_shift (Type ImmShift) ImmShift)
(extern constructor negate_imm_shift negate_imm_shift)
@@ -942,13 +910,13 @@
(let (
(val ValueRegs (put_in_regs x))
(amt Reg (value_regs_get (put_in_regs y) 0))
(neg_amt Reg (alu_rrr (ALUOp.Sub64) (imm $I64 128) amt))
(neg_amt Reg (sub64 (imm $I64 128) amt))
(lshift ValueRegs (lower_shl128 val amt))
(rshift ValueRegs (lower_ushr128 val neg_amt))
)
(value_regs
(alu_rrr (ALUOp.Orr64) (value_regs_get lshift 0) (value_regs_get rshift 0))
(alu_rrr (ALUOp.Orr64) (value_regs_get lshift 1) (value_regs_get rshift 1)))))
(orr64 (value_regs_get lshift 0) (value_regs_get rshift 0))
(orr64 (value_regs_get lshift 1) (value_regs_get rshift 1)))))
;;;; Rules for `rotr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -958,11 +926,11 @@
;; General 32-bit case.
(rule (lower (has_type $I32 (rotr x y)))
(value_reg (alu_rrr (ALUOp.RotR32) (put_in_reg x) (put_in_reg y))))
(value_reg (rotr32 (put_in_reg x) (put_in_reg y))))
;; General 64-bit case.
(rule (lower (has_type $I64 (rotr x y)))
(value_reg (alu_rrr (ALUOp.RotR64) (put_in_reg x) (put_in_reg y))))
(value_reg (rotr64 (put_in_reg x) (put_in_reg y))))
;; Specialization for the 8/16-bit case when the rotation amount is an immediate.
(rule (lower (has_type (fits_in_16 ty) (rotr x (def_inst (iconst (imm_shift_from_imm64 <ty n))))))
@@ -970,11 +938,11 @@
;; Specialization for the 32-bit case when the rotation amount is an immediate.
(rule (lower (has_type $I32 (rotr x (def_inst (iconst (imm_shift_from_imm64 <$I32 n))))))
(value_reg (alu_rr_imm_shift (ALUOp.RotR32) (put_in_reg x) n)))
(value_reg (rotr32_imm (put_in_reg x) n)))
;; Specialization for the 64-bit case when the rotation amount is an immediate.
(rule (lower (has_type $I64 (rotr x (def_inst (iconst (imm_shift_from_imm64 <$I64 n))))))
(value_reg (alu_rr_imm_shift (ALUOp.RotR64) (put_in_reg x) n)))
(value_reg (rotr64_imm (put_in_reg x) n)))
;; For a < 32-bit rotate-right, we synthesize this as:
;;
@@ -991,13 +959,13 @@
(decl small_rotr (Type Reg Reg) Reg)
(rule (small_rotr ty val amt)
(let (
(masked_amt Reg (alu_rr_imm_logic (ALUOp.And32) amt (rotr_mask ty)))
(tmp_sub Reg (alu_rr_imm12 (ALUOp.Sub32) masked_amt (u8_into_imm12 (ty_bits ty))))
(neg_amt Reg (alu_rrr (ALUOp.Sub32) (zero_reg) tmp_sub))
(val_rshift Reg (alu_rrr (ALUOp.Lsr32) val masked_amt))
(val_lshift Reg (alu_rrr (ALUOp.Lsl32) val neg_amt))
(masked_amt Reg (and32_imm amt (rotr_mask ty)))
(tmp_sub Reg (sub32_imm masked_amt (u8_into_imm12 (ty_bits ty))))
(neg_amt Reg (sub32 (zero_reg) tmp_sub))
(val_rshift Reg (lsr32 val masked_amt))
(val_lshift Reg (lsl32 val neg_amt))
)
(alu_rrr (ALUOp.Orr32) val_lshift val_rshift)))
(orr32 val_lshift val_rshift)))
(decl rotr_mask (Type) ImmLogic)
(extern constructor rotr_mask rotr_mask)
@@ -1014,10 +982,10 @@
(decl small_rotr_imm (Type Reg ImmShift) Reg)
(rule (small_rotr_imm ty val amt)
(let (
(val_rshift Reg (alu_rr_imm_shift (ALUOp.Lsr32) val amt))
(val_lshift Reg (alu_rr_imm_shift (ALUOp.Lsl32) val (rotr_opposite_amount ty amt)))
(val_rshift Reg (lsr32_imm val amt))
(val_lshift Reg (lsl32_imm val (rotr_opposite_amount ty amt)))
)
(alu_rrr (ALUOp.Orr32) val_lshift val_rshift)))
(orr32 val_lshift val_rshift)))
(decl rotr_opposite_amount (Type ImmShift) ImmShift)
(extern constructor rotr_opposite_amount rotr_opposite_amount)
@@ -1029,10 +997,10 @@
(let (
(val ValueRegs (put_in_regs x))
(amt Reg (value_regs_get (put_in_regs y) 0))
(neg_amt Reg (alu_rrr (ALUOp.Sub64) (imm $I64 128) amt))
(neg_amt Reg (sub64 (imm $I64 128) amt))
(rshift ValueRegs (lower_ushr128 val amt))
(lshift ValueRegs (lower_shl128 val neg_amt))
(hi Reg (alu_rrr (ALUOp.Orr64) (value_regs_get rshift 1) (value_regs_get lshift 1)))
(lo Reg (alu_rrr (ALUOp.Orr64) (value_regs_get rshift 0) (value_regs_get lshift 0)))
(hi Reg (orr64 (value_regs_get rshift 1) (value_regs_get lshift 1)))
(lo Reg (orr64 (value_regs_get rshift 0) (value_regs_get lshift 0)))
)
(value_regs lo hi)))