riscv64: Add Zba extension instructions (#6087)

* riscv64: Use `add.uw` to zero extend

* riscv64: Implement `add.uw` optimizations

* riscv64: Add `Zba` `iadd+ishl` optimizations

* riscv64: Add `shl+uextend` optimizations based on `Zba`

* riscv64: Fix some issues with `Zba` instructions

* riscv64: Restrict shnadd selection

* riscv64: Fix `extend` priorities

* riscv64: Remove redundant `addw` rule

* riscv64: Specify type for `add` extend rules

* riscv64: Use `u64_from_imm64` extractor instead of `uimm8`

* riscv64: Restrict `uextend` in `shnadd.uw` rules

* riscv64: Use concrete type in `slli.uw` rule

* riscv64: Add extra arithmetic extends tests

Co-authored-by: Jamey Sharp <jsharp@fastly.com>

* riscv64: Make `Adduw` types concrete

* riscv64: Add extra arithmetic extend tests

* riscv64: Add `sextend`+Arithmetic rules

* riscv64: Fix whitespace

* cranelift: Move arithmetic extends tests with i128 to separate file

---------

Co-authored-by: Jamey Sharp <jsharp@fastly.com>
This commit is contained in:
Afonso Bordado
2023-03-23 20:06:03 +00:00
committed by GitHub
parent 6f66abd5c7
commit 602ff71fe4
14 changed files with 782 additions and 33 deletions

View File

@@ -1157,9 +1157,15 @@
(let ((val Reg (value_regs_get val 0)))
(alu_rr_imm12 (AluOPRRI.Zexth) val (imm12_const 0))))
;; With `zba` we have a `zext.w` instruction
(rule 2 (extend val (ExtendOp.Zero) $I32 $I64)
(if-let $true (has_zba))
(let ((val Reg (value_regs_get val 0)))
(alu_rrr (AluOPRRR.Adduw) val (zero_reg))))
;;; Signed rules extending to I128
;; Extend the bottom part, and extract the sign bit from the bottom as the top
(rule 2 (extend val (ExtendOp.Signed) (fits_in_64 from_ty) $I128)
(rule 3 (extend val (ExtendOp.Signed) (fits_in_64 from_ty) $I128)
(let ((val Reg (value_regs_get val 0))
(low Reg (extend val (ExtendOp.Signed) from_ty $I64))
(high Reg (alu_rr_imm12 (AluOPRRI.Srai) low (imm12_const 63))))

View File

@@ -248,7 +248,6 @@ fn test_riscv64_binemit() {
0x28755593,
));
//
insns.push(TestUnit::new(
Inst::AluRRR {
alu_op: AluOPRRR::Adduw,
@@ -256,10 +255,21 @@ fn test_riscv64_binemit() {
rs1: a0(),
rs2: zero_reg(),
},
"add.uw a1,a0,zero",
"zext.w a1,a0",
0x80505bb,
));
insns.push(TestUnit::new(
Inst::AluRRR {
alu_op: AluOPRRR::Adduw,
rd: writable_a1(),
rs1: a0(),
rs2: a1(),
},
"add.uw a1,a0,a1",
0x08b505bb,
));
insns.push(TestUnit::new(
Inst::AluRRR {
alu_op: AluOPRRR::Andn,

View File

@@ -1220,10 +1220,17 @@ impl Inst {
rs1,
rs2,
} => {
let rs1 = format_reg(rs1, allocs);
let rs2 = format_reg(rs2, allocs);
let rd = format_reg(rd.to_reg(), allocs);
format!("{} {},{},{}", alu_op.op_name(), rd, rs1, rs2,)
let rs1_s = format_reg(rs1, allocs);
let rs2_s = format_reg(rs2, allocs);
let rd_s = format_reg(rd.to_reg(), allocs);
match alu_op {
AluOPRRR::Adduw if rs2 == zero_reg() => {
format!("zext.w {},{}", rd_s, rs1_s)
}
_ => {
format!("{} {},{},{}", alu_op.op_name(), rd_s, rs1_s, rs2_s)
}
}
}
&Inst::FpuRR {
frm,

View File

@@ -26,11 +26,9 @@
;;;; Rules for `iadd` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule -1 (lower (has_type (fits_in_32 ty) (iadd x y)))
(alu_rrr (AluOPRRR.Addw) x y))
;; Base case, simply adding things in registers.
(rule -2 (lower (has_type (fits_in_64 ty) (iadd x y)))
(rule 0 (lower (has_type (fits_in_64 ty) (iadd x y)))
(alu_add x y))
;; Special cases for when one operand is an immediate that fits in 12 bits.
@@ -40,17 +38,63 @@
(rule 2 (lower (has_type (fits_in_64 ty) (iadd (imm12_from_value x) y)))
(alu_rr_imm12 (select_addi ty) y x))
(rule
(lower (has_type $I128 (iadd x y)))
(let
( ;; low part.
(low Reg (alu_add (value_regs_get x 0) (value_regs_get y 0)))
;; compute carry.
(carry Reg (alu_rrr (AluOPRRR.SltU) low (value_regs_get y 0)))
;;
(high_tmp Reg (alu_add (value_regs_get x 1) (value_regs_get y 1)))
;; add carry.
(high Reg (alu_add high_tmp carry)))
;; Special case when one of the operands is uextended
;; Needs `Zba`
(rule 3 (lower (has_type $I64 (iadd x (uextend y @ (value_type $I32)))))
(if-let $true (has_zba))
(alu_rrr (AluOPRRR.Adduw) y x))
(rule 4 (lower (has_type $I64 (iadd (uextend x @ (value_type $I32)) y)))
(if-let $true (has_zba))
(alu_rrr (AluOPRRR.Adduw) x y))
;; Add with const shift. We have a few of these instructions with `Zba`.
(decl pure partial match_shnadd (Imm64) AluOPRRR)
(rule (match_shnadd (u64_from_imm64 1)) (AluOPRRR.Sh1add))
(rule (match_shnadd (u64_from_imm64 2)) (AluOPRRR.Sh2add))
(rule (match_shnadd (u64_from_imm64 3)) (AluOPRRR.Sh3add))
(rule 3 (lower (has_type $I64 (iadd x (ishl y (maybe_uextend (iconst n))))))
(if-let $true (has_zba))
(if-let shnadd (match_shnadd n))
(alu_rrr shnadd y x))
(rule 4 (lower (has_type $I64 (iadd (ishl x (maybe_uextend (iconst n))) y)))
(if-let $true (has_zba))
(if-let shnadd (match_shnadd n))
(alu_rrr shnadd x y))
;; Add with uextended const shift. We have a few of these instructions with `Zba`.
;;
;; !!! Important !!!
;; These rules only work for (ishl (uextend _) _) and not for (uextend (ishl _ _))!
;; Getting this wrong means a potential misscalculation of the shift amount.
;; Additionaly we can only ensure that this is correct if the uextend is 32 to 64 bits.
(decl pure partial match_shnadd_uw (Imm64) AluOPRRR)
(rule (match_shnadd_uw (u64_from_imm64 1)) (AluOPRRR.Sh1adduw))
(rule (match_shnadd_uw (u64_from_imm64 2)) (AluOPRRR.Sh2adduw))
(rule (match_shnadd_uw (u64_from_imm64 3)) (AluOPRRR.Sh3adduw))
(rule 5 (lower (has_type $I64 (iadd x (ishl (uextend y @ (value_type $I32)) (maybe_uextend (iconst n))))))
(if-let $true (has_zba))
(if-let shnadd_uw (match_shnadd_uw n))
(alu_rrr shnadd_uw y x))
(rule 6 (lower (has_type $I64 (iadd (ishl (uextend x @ (value_type $I32)) (maybe_uextend (iconst n))) y)))
(if-let $true (has_zba))
(if-let shnadd_uw (match_shnadd_uw n))
(alu_rrr shnadd_uw x y))
;; I128 cases
(rule 7 (lower (has_type $I128 (iadd x y)))
(let ((low Reg (alu_add (value_regs_get x 0) (value_regs_get y 0)))
;; compute carry.
(carry Reg (alu_rrr (AluOPRRR.SltU) low (value_regs_get y 0)))
;;
(high_tmp Reg (alu_add (value_regs_get x 1) (value_regs_get y 1)))
;; add carry.
(high Reg (alu_add high_tmp carry)))
(value_regs low high)))
;;; Rules for `uadd_overflow_trap` ;;;;;;;;;;;;;
@@ -355,6 +399,38 @@
(rule (lower (has_type out_ty (sextend val @ (value_type in_ty))))
(sext val in_ty out_ty))
;; The instructions below are present in RV64I and sign-extend the result to 64 bits.
(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (iadd x y)))))
(alu_rrr (AluOPRRR.Addw) x y))
(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (isub x y)))))
(alu_rrr (AluOPRRR.Subw) x y))
(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (ishl x y)))))
(alu_rrr (AluOPRRR.Sllw) x (value_regs_get y 0)))
(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (ushr x y)))))
(alu_rrr (AluOPRRR.Srlw) x (value_regs_get y 0)))
(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (sshr x y)))))
(alu_rrr (AluOPRRR.Sraw) x (value_regs_get y 0)))
(rule 2 (lower (has_type $I64 (sextend (has_type $I32 (iadd x (imm12_from_value y))))))
(alu_rr_imm12 (AluOPRRI.Addiw) x y))
(rule 3 (lower (has_type $I64 (sextend (has_type $I32 (iadd (imm12_from_value x) y)))))
(alu_rr_imm12 (AluOPRRI.Addiw) y x))
(rule 2 (lower (has_type $I64 (sextend (has_type $I32 (ishl x (imm12_from_value y))))))
(alu_rr_imm12 (AluOPRRI.Slliw) x y))
(rule 2 (lower (has_type $I64 (sextend (has_type $I32 (ushr x (imm12_from_value y))))))
(alu_rr_imm12 (AluOPRRI.SrliW) x y))
(rule 2 (lower (has_type $I64 (sextend (has_type $I32 (sshr x (imm12_from_value y))))))
(alu_rr_imm12 (AluOPRRI.Sraiw) x y))
;;;; Rules for `popcnt` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type (fits_in_64 ty) (popcnt x)))
@@ -385,6 +461,12 @@
(rule 1 (lower (has_type $I64 (ishl x y)))
(alu_rrr (AluOPRRR.Sll) x (value_regs_get y 0)))
;; With `Zba` we have a shift that zero extends the LHS argument.
(rule 3 (lower (has_type $I64 (ishl (uextend x @ (value_type $I32)) (maybe_uextend (imm12_from_value y)))))
(if-let $true (has_zba))
(alu_rr_imm12 (AluOPRRI.SlliUw) x y))
;; I128 cases
(rule 0 (lower (has_type $I128 (ishl x y)))
(lower_i128_ishl x y))