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

x64: Lower widening and narrowing operations in ISLE (#4722)

Browse Source 
Lower uwiden_high, uwiden_low, swiden_high, swiden_low, snarrow, and unarrow in ISLE.
This commit is contained in:
Trevor Elliott
2022-08-18 11:53:24 -07:00
committed by GitHub
parent 7d9a359f51
commit 8b6019909b
6 changed files with 387 additions and 207 deletions
Download Patch File Download Diff File Expand all files Collapse all files

86
cranelift/codegen/src/isa/x64/lower.isle
View File

@@ -3195,3 +3195,89 @@
(addd_const Xmm (x64_xmm_load_const $I16X8 (iadd_pairwise_addd_const_32))))
(x64_paddd dst addd_const)))
;; Rules for `swiden_low` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $I16X8 (swiden_low val @ (value_type $I8X16))))
(x64_pmovsxbw val))
(rule (lower (has_type $I32X4 (swiden_low val @ (value_type $I16X8))))
(x64_pmovsxwd val))
(rule (lower (has_type $I64X2 (swiden_low val @ (value_type $I32X4))))
(x64_pmovsxdq val))
;; Rules for `swiden_high` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $I16X8 (swiden_high val @ (value_type $I8X16))))
(x64_pmovsxbw (x64_palignr val val 8 (OperandSize.Size32))))
(rule (lower (has_type $I32X4 (swiden_high val @ (value_type $I16X8))))
(x64_pmovsxwd (x64_palignr val val 8 (OperandSize.Size32))))
(rule (lower (has_type $I64X2 (swiden_high val @ (value_type $I32X4))))
(x64_pmovsxdq (x64_pshufd val 0xEE (OperandSize.Size32))))
;; Rules for `uwiden_low` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $I16X8 (uwiden_low val @ (value_type $I8X16))))
(x64_pmovzxbw val))
(rule (lower (has_type $I32X4 (uwiden_low val @ (value_type $I16X8))))
(x64_pmovzxwd val))
(rule (lower (has_type $I64X2 (uwiden_low val @ (value_type $I32X4))))
(x64_pmovzxdq val))
;; Rules for `uwiden_high` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $I16X8 (uwiden_high val @ (value_type $I8X16))))
(x64_pmovzxbw (x64_palignr val val 8 (OperandSize.Size32))))
(rule (lower (has_type $I32X4 (uwiden_high val @ (value_type $I16X8))))
(x64_pmovzxwd (x64_palignr val val 8 (OperandSize.Size32))))
(rule (lower (has_type $I64X2 (uwiden_high val @ (value_type $I32X4))))
(x64_pmovzxdq (x64_pshufd val 0xEE (OperandSize.Size32))))
;; Rules for `snarrow` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $I8X16 (snarrow a @ (value_type $I16X8) b)))
(x64_packsswb a b))
(rule (lower (has_type $I16X8 (snarrow a @ (value_type $I32X4) b)))
(x64_packssdw a b))
;; We're missing a `snarrow` case for $I64X2
;; https://github.com/bytecodealliance/wasmtime/issues/4734
;; This rule is a special case for handling the translation of the wasm op
;; `i32x4.trunc_sat_f64x2_s_zero`. It can be removed once we have an
;; implementation of `snarrow` for `I64X2`.
(rule (lower (has_type $I32X4 (snarrow (has_type $I64X2 (fcvt_to_sint_sat a))
(vconst (u128_from_constant 0)))))
(let (;; y = i32x4.trunc_sat_f64x2_s_zero(x) is lowered to:
;; MOVE xmm_tmp, xmm_x
;; CMPEQPD xmm_tmp, xmm_x
;; MOVE xmm_y, xmm_x
;; ANDPS xmm_tmp, [wasm_f64x2_splat(2147483647.0)]
;; MINPD xmm_y, xmm_tmp
;; CVTTPD2DQ xmm_y, xmm_y
(tmp1 Xmm (x64_cmppd a a (FcmpImm.Equal)))
(umax_mask Xmm (x64_xmm_load_const $F64X2 (snarrow_umax_mask)))
;; ANDPD xmm_y, [wasm_f64x2_splat(2147483647.0)]
(tmp1 Xmm (x64_andps tmp1 umax_mask))
(dst Xmm (x64_minpd a tmp1)))
(x64_cvttpd2dq dst)))
;; Rules for `unarrow` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $I8X16 (unarrow a @ (value_type $I16X8) b)))
(x64_packuswb a b))
(rule (lower (has_type $I16X8 (unarrow a @ (value_type $I32X4) b)))
(x64_packusdw a b))
;; We're missing a `unarrow` case for $I64X2
;; https://github.com/bytecodealliance/wasmtime/issues/4734