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

Remove the Cranelift vselect instruction (#5918)

Browse Source 
* Remove the Cranelift `vselect` instruction

This instruction is documented as selecting lanes based on the "truthy"
value of the condition lane, but the current status of the
implementation of this instruction is:

* x64 - uses the high bit for `f32x4` and `f64x2` and otherwise uses the
  high bit of each byte doing a byte-wise lane select rather than
  whatever the controlling type is.

* AArch64 - this is the same as `bitselect` which is a bit-wise
  selection rather than a lane-wise selection.

* s390x - this is the same as AArch64, a bit-wise selection rather than
  lane-wise.

* interpreter - the interpreter implements the documented semantics of
  selecting based on "truthy" values.

Coupled with the status of the implementation is the fact that this
instruction is not used by WebAssembly SIMD today either. The only use
of this instruction in Cranelift is the nan-canonicalization pass. By
moving nan-canonicalization to `bitselect`, since that has the desired
semantics, there's no longer any need for `vselect`.

Given this situation this commit subsqeuently removes `vselect` and all
usage of it throughout Cranelift.

Closes #5917

* Review comments

* Bring back vselect opts as bitselect opts

* Clean up vselect usage in the interpreter

* Move bitcast in nan canonicalization

* Add a comment about float optimization
This commit is contained in:
Alex Crichton
2023-03-07 18:42:05 -06:00
committed by GitHub
parent fc45ccc125
commit 07518dfd36
14 changed files with 163 additions and 333 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
cranelift/codegen/meta/src/shared/instructions.rs
View File

@@ -1409,8 +1409,8 @@ pub(crate) fn define(
r#"
Conditional select.
This instruction selects whole values. Use `vselect` for
lane-wise selection.
This instruction selects whole values. Use `bitselect` to choose each
bit according to a mask.
"#,
&formats.ternary,
)
@@ -1458,7 +1458,7 @@ pub(crate) fn define(
For each bit in `c`, this instruction selects the corresponding bit from `x` if the bit
in `x` is 1 and the corresponding bit from `y` if the bit in `c` is 0. See also:
`select`, `vselect`.
`select`.
"#,
&formats.ternary,
)
@@ -1484,26 +1484,7 @@ pub(crate) fn define(
.operands_out(vec![a]),
);
let c = &Operand::new("c", &TxN.as_bool()).with_doc("Controlling vector");
let x = &Operand::new("x", TxN).with_doc("Value to use where `c` is true");
let y = &Operand::new("y", TxN).with_doc("Value to use where `c` is false");
let a = &Operand::new("a", TxN);
ig.push(
Inst::new(
"vselect",
r#"
Vector lane select.
Select lanes from ``x`` or ``y`` controlled by the lanes of the truthy
vector ``c``.
"#,
&formats.ternary,
)
.operands_in(vec![c, x, y])
.operands_out(vec![a]),
);
let s = &Operand::new("s", i8);
ig.push(

5
cranelift/codegen/src/isa/aarch64/lower.isle
View File

@@ -1659,11 +1659,6 @@
(rule 1 (lower (has_type (ty_vec128 ty) (bitselect c x y)))
(bsl ty c x y))
;;;; Rules for `vselect` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type (ty_vec128 ty) (vselect c x y)))
(bsl ty c x y))
;;;; Rules for `ireduce` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; T -> I{64,32,16,8}: We can simply pass through the value: values

7
cranelift/codegen/src/isa/s390x/lower.isle
View File

@@ -1170,13 +1170,6 @@
(vec_select ty y z x))
;;;; Rules for `vselect` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Vector select.
(rule (lower (has_type (ty_vec128 ty) (vselect x y z)))
(vec_select ty y z x))
;;;; Rules for `bmask` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type ty (bmask x)))

11
cranelift/codegen/src/isa/x64/lower.isle
View File

@@ -1193,7 +1193,7 @@
(sse_or ty b a)))
;; If every byte of the condition is guaranteed to be all ones or all zeroes,
;; we can use x86_blend like vselect does.
;; we can use x64_blend.
(rule 1 (lower (has_type ty @ (multi_lane _bits _lanes)
(bitselect condition
if_true
@@ -1226,15 +1226,6 @@
(x86_blendv condition if_true if_false)))
(x64_blendvpd if_false if_true condition))
;;;; Rules for `vselect` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type ty @ (multi_lane _bits _lanes)
(vselect condition if_true if_false)))
(x64_blend ty
condition
if_true
if_false))
;;;; Rules for `insertlane` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (insertlane vec @ (value_type ty) val (u8_from_uimm8 idx)))

4
cranelift/codegen/src/nan_canonicalization.rs
View File

@@ -7,6 +7,7 @@ use crate::ir::condcodes::FloatCC;
use crate::ir::immediates::{Ieee32, Ieee64};
use crate::ir::types;
use crate::ir::{Function, Inst, InstBuilder, InstructionData, Opcode, Value};
use crate::opts::MemFlags;
use crate::timing;
// Canonical 32-bit and 64-bit NaN values.
@@ -70,9 +71,10 @@ fn add_nan_canon_seq(pos: &mut FuncCursor, inst: Inst) {
.select(is_nan, canon_nan, new_res);
};
let vector_select = |pos: &mut FuncCursor, canon_nan: Value| {
let is_nan = pos.ins().bitcast(val_type, MemFlags::new(), is_nan);
pos.ins()
.with_result(val)
.vselect(is_nan, canon_nan, new_res);
.bitselect(is_nan, canon_nan, new_res);
};
match val_type {

44
cranelift/codegen/src/opts/algebraic.isle
View File

@@ -454,56 +454,56 @@
(select ty (icmp _ (IntCC.UnsignedGreaterThanOrEqual) x y) y x))
(umin ty x y))
;; Transform vselect-of-icmp into {u,s}{min,max} instructions where possible.
;; Transform bitselect-of-icmp into {u,s}{min,max} instructions where possible.
(rule (simplify
(vselect ty (icmp _ (IntCC.SignedGreaterThan) x y) x y))
(bitselect ty (icmp _ (IntCC.SignedGreaterThan) x y) x y))
(smax ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.SignedGreaterThanOrEqual) x y) x y))
(bitselect ty (icmp _ (IntCC.SignedGreaterThanOrEqual) x y) x y))
(smax ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.UnsignedGreaterThan) x y) x y))
(bitselect ty (icmp _ (IntCC.UnsignedGreaterThan) x y) x y))
(umax ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.UnsignedGreaterThanOrEqual) x y) x y))
(bitselect ty (icmp _ (IntCC.UnsignedGreaterThanOrEqual) x y) x y))
(umax ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.SignedLessThan) x y) x y))
(bitselect ty (icmp _ (IntCC.SignedLessThan) x y) x y))
(smin ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.SignedLessThanOrEqual) x y) x y))
(bitselect ty (icmp _ (IntCC.SignedLessThanOrEqual) x y) x y))
(smin ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.UnsignedLessThan) x y) x y))
(bitselect ty (icmp _ (IntCC.UnsignedLessThan) x y) x y))
(umin ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.UnsignedLessThanOrEqual) x y) x y))
(bitselect ty (icmp _ (IntCC.UnsignedLessThanOrEqual) x y) x y))
(umin ty x y))
;; These are the same rules as above, but when the operands for select are swapped
(rule (simplify
(vselect ty (icmp _ (IntCC.SignedLessThan) x y) y x))
(bitselect ty (icmp _ (IntCC.SignedLessThan) x y) y x))
(smax ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.SignedLessThanOrEqual) x y) y x))
(bitselect ty (icmp _ (IntCC.SignedLessThanOrEqual) x y) y x))
(smax ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.UnsignedLessThan) x y) y x))
(bitselect ty (icmp _ (IntCC.UnsignedLessThan) x y) y x))
(umax ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.UnsignedLessThanOrEqual) x y) y x))
(bitselect ty (icmp _ (IntCC.UnsignedLessThanOrEqual) x y) y x))
(umax ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.SignedGreaterThan) x y) y x))
(bitselect ty (icmp _ (IntCC.SignedGreaterThan) x y) y x))
(smin ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.SignedGreaterThanOrEqual) x y) y x))
(bitselect ty (icmp _ (IntCC.SignedGreaterThanOrEqual) x y) y x))
(smin ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.UnsignedGreaterThan) x y) y x))
(bitselect ty (icmp _ (IntCC.UnsignedGreaterThan) x y) y x))
(umin ty x y))
(rule (simplify
(vselect ty (icmp _ (IntCC.UnsignedGreaterThanOrEqual) x y) y x))
(bitselect ty (icmp _ (IntCC.UnsignedGreaterThanOrEqual) x y) y x))
(umin ty x y))
;; For floats convert fcmp lt into pseudo_min and gt into pseudo_max
@@ -520,13 +520,9 @@
(select ty (fcmp _ (FloatCC.GreaterThan) x y) x y))
(fmax_pseudo ty x y))
;; Do the same for vectors
(rule (simplify
(vselect ty (fcmp _ (FloatCC.LessThan) x y) x y))
(fmin_pseudo ty x y))
(rule (simplify
(vselect ty (fcmp _ (FloatCC.GreaterThan) x y) x y))
(fmax_pseudo ty x y))
;; TODO: perform this same optimization to `f{min,max}_pseudo` for vectors
;; with the `bitselect` instruction, but the pattern is a bit more complicated
;; due to most bitselects-over-floats having bitcasts.
;; If both of the multiplied arguments to an `fma` are negated then remove
;; both of them since they cancel out.