riscv64: Fix regaloc panic with bor+bnot on floats (#5857)

cranelift/codegen/src/isa/riscv64/lower.isle

@@ -186,40 +186,44 @@
     (alu_rrr (AluOPRRR.RemU) x y)))
 
 ;;;; Rules for `and` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule -1 (lower (has_type (fits_in_64 ty) (band x y)))
+(rule -1 (lower (has_type (fits_in_64 (ty_int ty)) (band x y)))
   (alu_rrr (AluOPRRR.And) x y))
 
 ;; Special cases for when one operand is an immediate that fits in 12 bits.
-(rule 2 (lower (has_type (fits_in_64 ty) (band x (imm12_from_value y))))
+(rule 2 (lower (has_type (fits_in_64 (ty_int ty)) (band x (imm12_from_value y))))
   (alu_rr_imm12 (AluOPRRI.Andi) x y))
 
-(rule 1 (lower (has_type (fits_in_64 ty) (band (imm12_from_value x) y)))
+(rule 1 (lower (has_type (fits_in_64 (ty_int ty)) (band (imm12_from_value x) y)))
   (alu_rr_imm12 (AluOPRRI.Andi) y x))
 
 (rule (lower (has_type $I128 (band x y)))
   (lower_b128_binary (AluOPRRR.And) x y))
 
 (rule (lower (has_type $F32 (band x y)))
   (lower_float_binary (AluOPRRR.And) x y $F32))
+
 (rule (lower (has_type $F64 (band x y)))
   (lower_float_binary (AluOPRRR.And) x y $F64))
 
 ;; Specialized lowerings for `(band x (bnot y))` which is additionally produced
 ;; by Cranelift's `band_not` instruction that is legalized into the simpler
 ;; forms early on.
 
-(rule 3 (lower (has_type (fits_in_64 ty) (band x (bnot y))))
+(rule 3 (lower (has_type (fits_in_64 (ty_int ty)) (band x (bnot y))))
   (if-let $true (has_zbb))
   (gen_andn x y))
-(rule 4 (lower (has_type (fits_in_64 ty) (band (bnot y) x)))
+
+(rule 4 (lower (has_type (fits_in_64 (ty_int ty)) (band (bnot y) x)))
   (if-let $true (has_zbb))
   (gen_andn x y))
+
 (rule 5 (lower (has_type $I128 (band x (bnot y))))
   (if-let $true (has_zbb))
   (let
     ((low Reg (gen_andn (value_regs_get x 0) (value_regs_get y 0)))
       (high Reg (gen_andn (value_regs_get x 1) (value_regs_get y 1))))
     (value_regs low high)))
+
 (rule 6 (lower (has_type $I128 (band (bnot y) x)))
   (if-let $true (has_zbb))
   (let
@@ -229,30 +233,34 @@
 
 
 ;;;; Rules for `or` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule -1 (lower (has_type (fits_in_64 ty) (bor x y)))
+(rule -1 (lower (has_type (fits_in_64 (ty_int ty)) (bor x y)))
   (alu_rrr (AluOPRRR.Or) x y))
 
 ;; Special cases for when one operand is an immediate that fits in 12 bits.
-(rule 2 (lower (has_type (fits_in_64 ty) (bor x (imm12_from_value y))))
+(rule 2 (lower (has_type (fits_in_64 (ty_int ty)) (bor x (imm12_from_value y))))
   (alu_rr_imm12 (AluOPRRI.Ori) x y))
 
-(rule 1 (lower (has_type (fits_in_64 ty) (bor (imm12_from_value x) y)))
+(rule 1 (lower (has_type (fits_in_64 (ty_int ty)) (bor (imm12_from_value x) y)))
   (alu_rr_imm12 (AluOPRRI.Ori) y x))
+
 (rule (lower (has_type $I128 (bor x y)))
   (lower_b128_binary (AluOPRRR.Or) x y))
+
 (rule (lower (has_type $F32 (bor x y)))
   (lower_float_binary (AluOPRRR.Or) x y $F32))
+
 (rule (lower (has_type $F64 (bor x y)))
   (lower_float_binary (AluOPRRR.Or) x y $F64))
 
 ;; Specialized lowerings for `(bor x (bnot y))` which is additionally produced
 ;; by Cranelift's `bor_not` instruction that is legalized into the simpler
 ;; forms early on.
 
-(rule 3 (lower (has_type (fits_in_64 ty) (bor x (bnot y))))
+(rule 3 (lower (has_type (fits_in_64 (ty_int ty)) (bor x (bnot y))))
   (if-let $true (has_zbb))
   (gen_orn x y))
-(rule 4 (lower (has_type (fits_in_64 ty) (bor (bnot y) x)))
+
+(rule 4 (lower (has_type (fits_in_64 (ty_int ty)) (bor (bnot y) x)))
   (if-let $true (has_zbb))
   (gen_orn x y))
 
@@ -262,6 +270,7 @@
     ((low Reg (gen_orn (value_regs_get x 0) (value_regs_get y 0)))
       (high Reg (gen_orn (value_regs_get x 1) (value_regs_get y 1))))
     (value_regs low high)))
+
 (rule 6 (lower (has_type $I128 (bor (bnot y) x)))
   (if-let $true (has_zbb))
   (let
@@ -271,40 +280,43 @@
 
 
 ;;;; Rules for `xor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule -1 (lower (has_type (fits_in_64 ty) (bxor x y)))
+(rule -1 (lower (has_type (fits_in_64 (ty_int ty)) (bxor x y)))
   (alu_rrr (AluOPRRR.Xor) x y))
 
 ;; Special cases for when one operand is an immediate that fits in 12 bits.
-(rule 2 (lower (has_type (fits_in_64 ty) (bxor x (imm12_from_value y))))
+(rule 2 (lower (has_type (fits_in_64 (ty_int ty)) (bxor x (imm12_from_value y))))
   (alu_rr_imm12 (AluOPRRI.Xori) x y))
 
-(rule 1 (lower (has_type (fits_in_64 ty) (bxor (imm12_from_value x) y)))
+(rule 1 (lower (has_type (fits_in_64 (ty_int ty)) (bxor (imm12_from_value x) y)))
   (alu_rr_imm12 (AluOPRRI.Xori) y x))
+
 (rule (lower (has_type $I128 (bxor x y)))
   (lower_b128_binary (AluOPRRR.Xor) x y))
+
 (rule (lower (has_type $F32 (bxor x y)))
   (lower_float_binary (AluOPRRR.Xor) x y $F32))
+
 (rule (lower (has_type $F64 (bxor x y)))
   (lower_float_binary (AluOPRRR.Xor) x y $F64))
 
 
 ;;;; Rules for `bnot` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule -1 (lower (has_type fits_in_64 (bnot x)))
+(rule -1 (lower (has_type (fits_in_64 (ty_int ty)) (bnot x)))
   (alu_rr_imm12 (AluOPRRI.Xori) x (imm_from_neg_bits -1)))
 
 (rule (lower (has_type $I128 (bnot x)))
   (bnot_128 x))
+
 (rule
   (lower (has_type $F32 (bnot x)))
-  (lower_float_bnot x $F32)
-)
+  (lower_float_bnot x $F32))
+
 (rule
   (lower (has_type $F64 (bnot x)))
-  (lower_float_bnot x $F64)
-)
+  (lower_float_bnot x $F64))
 
 ;;;; Rules for `bit_reverse` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule (lower (has_type ty (bitrev x)))
+(rule (lower (has_type (fits_in_64 (ty_int ty)) (bitrev x)))
   (lower_bit_reverse x ty))
 
 (rule 1 (lower (has_type $I128 (bitrev x)))

cranelift/filetests/filetests/isa/riscv64/bitops-float.clif

@@ -0,0 +1,79 @@
+test compile precise-output
+set unwind_info=false
+target riscv64 has_zbb
+
+;; This is a regression test for a bug in the RISC-V backend where
+;; When enabling `Zbb` the backend would try to use one of the
+;; integer instructions (`orn`) to implement a float operation
+;; causing a regalloc panic.
+
+function %or_not_optimization_float() -> i32 system_v {
+block0:
+    v0 = iconst.i32 0
+    v1 = f32const 0.0
+    v2 = bnot v1 
+    v3 = bor v2, v2
+    br_table v0, block1(v3), [block1(v1)]
+
+block1(v4: f32):
+    return v0
+}
+
+; VCode:
+; block0:
+;   li a1,0
+;   fmv.w.x ft9,a1
+;   li t1,0
+;   fmv.w.x fa6,t1
+;   fmv.x.w a1,fa6
+;   not a3,a1
+;   fmv.w.x ft1,a3
+;   fmv.x.w t1,ft1
+;   fmv.x.w a0,ft1
+;   or a2,t1,a0
+;   fmv.w.x fa2,a2
+;   li t2,0
+;   br_table t2,[MachLabel(1),MachLabel(2)]##tmp1=a1,tmp2=a2
+; block1:
+;   j label3
+; block2:
+;   fmv.d fa2,ft9
+;   j label3
+; block3:
+;   li a0,0
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   mv a1, zero
+;   fmv.w.x ft9, a1
+;   mv t1, zero
+;   fmv.w.x fa6, t1
+;   fmv.x.w a1, fa6
+;   not a3, a1
+;   fmv.w.x ft1, a3
+;   fmv.x.w t1, ft1
+;   fmv.x.w a0, ft1
+;   or a2, t1, a0
+;   fmv.w.x fa2, a2
+;   mv t2, zero
+;   slli t6, t2, 0x20
+;   srli t6, t6, 0x20
+;   addi a2, zero, 1
+;   bltu t6, a2, 0xc
+;   auipc a2, 0
+;   jalr zero, a2, 0x28
+;   auipc a1, 0
+;   slli a2, t6, 3
+;   add a1, a1, a2
+;   jalr zero, a1, 0x10
+;   auipc a2, 0
+;   jalr zero, a2, 0xc
+; block1: ; offset 0x60
+;   j 8
+; block2: ; offset 0x64
+;   fmv.d fa2, ft9
+; block3: ; offset 0x68
+;   mv a0, zero
+;   ret
+