fix rotl.i16 with i128 shift value. (#5611)

* fix issue 5523.

* fix.

* add missing issue file.

* fix issue.

* fix duplicate shamt_128.

* issue 5523 add test target,and fix some wrong comment.

* fix output file.

* enable llvm_abi_extensions for regression test file.

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

@@ -1125,8 +1125,9 @@
     (alu_rrr (AluOPRRR.Or) part1 part3)))
 
 
-;;;; construct shift amount
+;;;; construct shift amount.rotl on i128 will use shift to implement. So can call this function.
 ;;;; this will return shift amount and (ty_bits - "shift amount")
+;;;; if ty_bits is greater than 64 like i128, then shmat will fallback to 64.because We are 64 bit platform.
 (decl gen_shamt (Type Reg) ValueRegs)
 (extern constructor gen_shamt gen_shamt)
 
@@ -1243,12 +1244,13 @@
       (high_part3 Reg (gen_select_reg (IntCC.Equal) shamt (zero_reg) (zero_reg) high_part2))
       (high Reg (alu_rrr (AluOPRRR.Or) high_part1 high_part3))
       ;;
-      (const64 Reg (load_u64_constant 64)))
+      (const64 Reg (load_u64_constant 64))
+      (shamt_128 Reg (alu_andi (value_regs_get y 0) 127)))
     ;; right now we only rotate less than 64 bits.
-    ;; if shamt is greater than 64 , we should switch low and high.
+    ;; if shamt is greater than or equal 64 , we should switch low and high.
     (value_regs
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 high low)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 high low)
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 low high)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 low high)
     )))
 
 
@@ -1272,12 +1274,13 @@
       (high Reg (alu_rrr (AluOPRRR.Or) high_part1 high_part3))
 
       ;;
-      (const64 Reg (load_u64_constant 64)))
+      (const64 Reg (load_u64_constant 64))
+      (shamt_128 Reg (alu_andi (value_regs_get y 0) 127)))
     ;; right now we only rotate less than 64 bits.
-    ;; if shamt is greater than 64 , we should switch low and high.
+    ;; if shamt is greater than or equal 64 , we should switch low and high.
     (value_regs
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 high low)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 high low)
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 low high)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 low high)
     )))
 
 
@@ -1297,10 +1300,11 @@
       (high_part3 Reg (alu_rrr (AluOPRRR.Sll) (value_regs_get x 1) shamt))
       (high Reg (alu_rrr (AluOPRRR.Or) high_part2 high_part3 ))
       ;;
-      (const64 Reg (load_u64_constant 64)))
+      (const64 Reg (load_u64_constant 64))
+      (shamt_128 Reg (alu_andi (value_regs_get y 0) 127)))
     (value_regs
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 (zero_reg) low)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 (zero_reg) low)
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 low high))))
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 low high))))
 
 (decl lower_i128_ushr (ValueRegs ValueRegs) ValueRegs)
 (rule
@@ -1320,10 +1324,11 @@
       (const64 Reg (load_u64_constant 64))
 
       ;;
-      (high Reg (alu_rrr (AluOPRRR.Srl) (value_regs_get x 1) shamt)))
+      (high Reg (alu_rrr (AluOPRRR.Srl) (value_regs_get x 1) shamt))
+      (shamt_128 Reg (alu_andi (value_regs_get y 0) 127)))
     (value_regs
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 high low)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 high low)
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 (zero_reg) high))))
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 (zero_reg) high))))
 
 
 (decl lower_i128_sshr (ValueRegs ValueRegs) ValueRegs)
@@ -1347,10 +1352,12 @@
       ;;
       (const_neg_1 Reg (load_imm12 -1))
       ;;
-      (high_replacement Reg (gen_select_reg (IntCC.SignedLessThan) (value_regs_get x 1) (zero_reg) const_neg_1 (zero_reg))))
+      (high_replacement Reg (gen_select_reg (IntCC.SignedLessThan) (value_regs_get x 1) (zero_reg) const_neg_1 (zero_reg)))
+      (const64 Reg (load_u64_constant 64))
+      (shamt_128 Reg (alu_andi (value_regs_get y 0) 127)))
     (value_regs
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 high low)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 high low)
-      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt const64 high_replacement high))))
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 high_replacement high))))
 
 (decl load_imm12 (i32) Reg)
 (rule

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

@@ -253,19 +253,20 @@ impl generated_code::Context for IsleContext<'_, '_, MInst, Riscv64Backend> {
 
     //
     fn gen_shamt(&mut self, ty: Type, shamt: Reg) -> ValueRegs {
+        let ty_bits = if ty.bits() > 64 { 64 } else { ty.bits() };
         let shamt = {
             let tmp = self.temp_writable_reg(I64);
             self.emit(&MInst::AluRRImm12 {
                 alu_op: AluOPRRI::Andi,
                 rd: tmp,
                 rs: shamt,
-                imm12: Imm12::from_bits((ty.bits() - 1) as i16),
+                imm12: Imm12::from_bits((ty_bits - 1) as i16),
             });
             tmp.to_reg()
         };
         let len_sub_shamt = {
             let tmp = self.temp_writable_reg(I64);
-            self.emit(&MInst::load_imm12(tmp, Imm12::from_bits(ty.bits() as i16)));
+            self.emit(&MInst::load_imm12(tmp, Imm12::from_bits(ty_bits as i16)));
             let len_sub_shamt = self.temp_writable_reg(I64);
             self.emit(&MInst::AluRRR {
                 alu_op: AluOPRRR::Sub,

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

@@ -795,19 +795,18 @@ block0(v0: i128, v1: i8):
 }
 
 ; block0:
-;   mv t2,a1
+;   andi a3,a2,63
-;   andi a1,a2,127
-;   li a3,128
-;   sub a5,a3,a1
-;   sll a7,a0,a1
-;   srl t4,a0,a5
-;   select_reg t1,zero,t4##condition=(a1 eq zero)
-;   mv a5,t2
-;   sll a0,a5,a1
-;   or a2,t1,a0
 ;   li a4,64
-;   select_reg a0,zero,a7##condition=(a1 uge a4)
+;   sub a5,a4,a3
-;   select_reg a1,a7,a2##condition=(a1 uge a4)
+;   sll a7,a0,a3
+;   srl t4,a0,a5
+;   select_reg t1,zero,t4##condition=(a3 eq zero)
+;   sll a0,a1,a3
+;   or a3,t1,a0
+;   li a4,64
+;   andi a6,a2,127
+;   select_reg a0,zero,a7##condition=(a6 uge a4)
+;   select_reg a1,a7,a3##condition=(a6 uge a4)
 ;   ret
 
 function %ishl_i128_i128(i128, i128) -> i128 {
@@ -817,17 +816,18 @@ block0(v0: i128, v1: i128):
 }
 
 ; block0:
-;   andi a2,a2,127
+;   andi a3,a2,63
-;   li a4,128
+;   li a4,64
-;   sub a6,a4,a2
+;   sub a6,a4,a3
-;   sll t3,a0,a2
+;   sll t3,a0,a3
 ;   srl t0,a0,a6
-;   select_reg t2,zero,t0##condition=(a2 eq zero)
+;   select_reg t2,zero,t0##condition=(a3 eq zero)
-;   sll a1,a1,a2
+;   sll a1,a1,a3
 ;   or a3,t2,a1
 ;   li a5,64
-;   select_reg a0,zero,t3##condition=(a2 uge a5)
+;   andi a7,a2,127
-;   select_reg a1,t3,a3##condition=(a2 uge a5)
+;   select_reg a0,zero,t3##condition=(a7 uge a5)
+;   select_reg a1,t3,a3##condition=(a7 uge a5)
 ;   ret
 
 function %ushr_i128_i8(i128, i8) -> i128 {
@@ -837,19 +837,18 @@ block0(v0: i128, v1: i8):
 }
 
 ; block0:
-;   mv t2,a1
+;   andi a3,a2,63
-;   andi a1,a2,127
+;   li a4,64
-;   li a3,128
+;   sub a5,a4,a3
-;   sub a5,a3,a1
+;   sll a7,a1,a5
-;   mv a2,t2
+;   select_reg t4,zero,a7##condition=(a3 eq zero)
-;   sll a7,a2,a5
+;   srl t1,a0,a3
-;   select_reg t4,zero,a7##condition=(a1 eq zero)
-;   srl t1,a0,a1
 ;   or a0,t4,t1
-;   li a3,64
+;   li a4,64
-;   srl a4,a2,a1
+;   srl a5,a1,a3
-;   select_reg a0,a4,a0##condition=(a1 uge a3)
+;   andi a6,a2,127
-;   select_reg a1,zero,a4##condition=(a1 uge a3)
+;   select_reg a0,a5,a0##condition=(a6 uge a4)
+;   select_reg a1,zero,a5##condition=(a6 uge a4)
 ;   ret
 
 function %ushr_i128_i128(i128, i128) -> i128 {
@@ -859,17 +858,18 @@ block0(v0: i128, v1: i128):
 }
 
 ; block0:
-;   andi a2,a2,127
+;   andi a3,a2,63
-;   li a4,128
+;   li a4,64
-;   sub a6,a4,a2
+;   sub a6,a4,a3
 ;   sll t3,a1,a6
-;   select_reg t0,zero,t3##condition=(a2 eq zero)
+;   select_reg t0,zero,t3##condition=(a3 eq zero)
-;   srl t2,a0,a2
+;   srl t2,a0,a3
-;   or a4,t0,t2
+;   or a5,t0,t2
-;   li a3,64
+;   li a4,64
-;   srl a5,a1,a2
+;   srl a6,a1,a3
-;   select_reg a0,a5,a4##condition=(a2 uge a3)
+;   andi a7,a2,127
-;   select_reg a1,zero,a5##condition=(a2 uge a3)
+;   select_reg a0,a6,a5##condition=(a7 uge a4)
+;   select_reg a1,zero,a6##condition=(a7 uge a4)
 ;   ret
 
 function %sshr_i128_i8(i128, i8) -> i128 {
@@ -879,20 +879,21 @@ block0(v0: i128, v1: i8):
 }
 
 ; block0:
-;   mv a4,a1
+;   andi a3,a2,63
-;   andi a1,a2,127
+;   li a4,64
-;   li a3,128
+;   sub a5,a4,a3
-;   sub a5,a3,a1
+;   sll a7,a1,a5
-;   sll a7,a4,a5
+;   select_reg t4,zero,a7##condition=(a3 eq zero)
-;   select_reg t4,zero,a7##condition=(a1 eq zero)
+;   srl t1,a0,a3
-;   srl t1,a0,a1
 ;   or a0,t4,t1
-;   li a2,64
+;   li a4,64
-;   sra a5,a4,a1
+;   sra a4,a1,a3
 ;   li a6,-1
-;   select_reg t3,a6,zero##condition=(a4 slt zero)
+;   select_reg t3,a6,zero##condition=(a1 slt zero)
-;   select_reg a0,a5,a0##condition=(a1 uge a2)
+;   li t0,64
-;   select_reg a1,t3,a5##condition=(a1 uge a2)
+;   andi t2,a2,127
+;   select_reg a0,a4,a0##condition=(t2 uge t0)
+;   select_reg a1,t3,a4##condition=(t2 uge t0)
 ;   ret
 
 function %sshr_i128_i128(i128, i128) -> i128 {
@@ -902,18 +903,20 @@ block0(v0: i128, v1: i128):
 }
 
 ; block0:
-;   andi a2,a2,127
+;   andi a3,a2,63
-;   li a4,128
+;   li a4,64
-;   sub a6,a4,a2
+;   sub a6,a4,a3
 ;   sll t3,a1,a6
-;   select_reg t0,zero,t3##condition=(a2 eq zero)
+;   select_reg t0,zero,t3##condition=(a3 eq zero)
-;   srl t2,a0,a2
+;   srl t2,a0,a3
 ;   or a4,t0,t2
-;   li a3,64
+;   li a5,64
-;   sra a5,a1,a2
+;   sra a5,a1,a3
 ;   li a7,-1
 ;   select_reg t4,a7,zero##condition=(a1 slt zero)
-;   select_reg a0,a5,a4##condition=(a2 uge a3)
+;   li t1,64
-;   select_reg a1,t4,a5##condition=(a2 uge a3)
+;   andi a1,a2,127
+;   select_reg a0,a5,a4##condition=(a1 uge t1)
+;   select_reg a1,t4,a5##condition=(a1 uge t1)
 ;   ret
 

cranelift/filetests/filetests/isa/riscv64/shift-rotate.clif

@@ -13,20 +13,22 @@ block0(v0: i128, v1: i128):
 }
 
 ; block0:
-;   andi a2,a2,127
+;   andi a3,a2,63
-;   li a4,128
+;   li a4,64
-;   sub a6,a4,a2
+;   sub a6,a4,a3
-;   srl t3,a0,a2
+;   srl t3,a0,a3
 ;   sll t0,a1,a6
-;   select_reg t2,zero,t0##condition=(a2 eq zero)
+;   mv t1,a1
-;   or a3,t3,t2
+;   select_reg t2,zero,t0##condition=(a3 eq zero)
-;   srl a4,a1,a2
+;   or a1,t3,t2
+;   srl a4,t1,a3
 ;   sll a5,a0,a6
-;   select_reg a7,zero,a5##condition=(a2 eq zero)
+;   select_reg a7,zero,a5##condition=(a3 eq zero)
 ;   or t4,a4,a7
 ;   li t1,64
-;   select_reg a0,t4,a3##condition=(a2 uge t1)
+;   andi a2,a2,127
-;   select_reg a1,a3,t4##condition=(a2 uge t1)
+;   select_reg a0,t4,a1##condition=(a2 uge t1)
+;   select_reg a1,a1,t4##condition=(a2 uge t1)
 ;   ret
 
 function %f0(i64, i64) -> i64 {
@@ -105,20 +107,22 @@ block0(v0: i128, v1: i128):
 }
 
 ; block0:
-;   andi a2,a2,127
+;   andi a3,a2,63
-;   li a4,128
+;   li a4,64
-;   sub a6,a4,a2
+;   sub a6,a4,a3
-;   sll t3,a0,a2
+;   sll t3,a0,a3
 ;   srl t0,a1,a6
-;   select_reg t2,zero,t0##condition=(a2 eq zero)
+;   mv t1,a1
-;   or a3,t3,t2
+;   select_reg t2,zero,t0##condition=(a3 eq zero)
-;   sll a4,a1,a2
+;   or a1,t3,t2
+;   sll a4,t1,a3
 ;   srl a5,a0,a6
-;   select_reg a7,zero,a5##condition=(a2 eq zero)
+;   select_reg a7,zero,a5##condition=(a3 eq zero)
 ;   or t4,a4,a7
 ;   li t1,64
-;   select_reg a0,t4,a3##condition=(a2 uge t1)
+;   andi a2,a2,127
-;   select_reg a1,a3,t4##condition=(a2 uge t1)
+;   select_reg a0,t4,a1##condition=(a2 uge t1)
+;   select_reg a1,a1,t4##condition=(a2 uge t1)
 ;   ret
 
 function %f4(i64, i64) -> i64 {

cranelift/filetests/filetests/runtests/issue5523.clif

@@ -0,0 +1,15 @@
+test interpret
+test run
+set enable_llvm_abi_extensions=true
+target riscv64
+target aarch64
+target s390x
+target x86_64
+
+function %a(i16, i128) -> i128 system_v {
+block0(v0: i16, v1: i128):
+    v2 = rotl v1, v0
+    return v2
+}
+
+; run: %a(64, 1095219937288) == 20203241887575960770402119057408