Avoid reusing a register during constant loading (#5379)

Avoid reusing a register when loading a constant, allocating a temporary instead.

cranelift/codegen/src/isa/riscv64/inst/emit.rs

@@ -53,11 +53,16 @@ impl LoadConstant {
         }
     }
 
-    pub(crate) fn load_constant(self, rd: Writable<Reg>) -> SmallInstVec<Inst> {
+    pub(crate) fn load_constant<F: FnMut(Type) -> Writable<Reg>>(
+        self,
+        rd: Writable<Reg>,
+        alloc_tmp: &mut F,
+    ) -> SmallInstVec<Inst> {
         let mut insts = SmallInstVec::new();
         // get current pc.
+        let pc = alloc_tmp(I64);
         insts.push(Inst::Auipc {
-            rd,
+            rd: pc,
             imm: Imm20 { bits: 0 },
         });
         // load
@@ -65,7 +70,7 @@ impl LoadConstant {
             rd,
             op: self.load_op(),
             flags: MemFlags::new(),
-            from: AMode::RegOffset(rd.to_reg(), 12, self.load_ty()),
+            from: AMode::RegOffset(pc.to_reg(), 12, self.load_ty()),
         });
         let data = self.to_le_bytes();
         // jump over.
@@ -78,10 +83,10 @@ impl LoadConstant {
 
     // load and perform an extra add.
     pub(crate) fn load_constant_and_add(self, rd: Writable<Reg>, rs: Reg) -> SmallInstVec<Inst> {
-        let mut insts = self.load_constant(rd);
+        let mut insts = self.load_constant(rd, &mut |_| rd);
         insts.push(Inst::AluRRR {
             alu_op: AluOPRRR::Add,
-            rd: rd,
+            rd,
             rs1: rd.to_reg(),
             rs2: rs,
         });

cranelift/codegen/src/isa/riscv64/inst/mod.rs

@@ -235,7 +235,7 @@ impl Inst {
         alloc_tmp: &mut F,
     ) -> SmallInstVec<Inst> {
         let insts = Inst::load_const_imm(rd, value, alloc_tmp);
-        insts.unwrap_or(LoadConstant::U32(value as u32).load_constant(rd))
+        insts.unwrap_or(LoadConstant::U32(value as u32).load_constant(rd, alloc_tmp))
     }
 
     pub fn load_constant_u64<F: FnMut(Type) -> Writable<Reg>>(
@@ -244,7 +244,7 @@ impl Inst {
         alloc_tmp: &mut F,
     ) -> SmallInstVec<Inst> {
         let insts = Inst::load_const_imm(rd, value, alloc_tmp);
-        insts.unwrap_or(LoadConstant::U64(value).load_constant(rd))
+        insts.unwrap_or(LoadConstant::U64(value).load_constant(rd, alloc_tmp))
     }
 
     pub(crate) fn construct_auipc_and_jalr(

cranelift/filetests/filetests/isa/riscv64/constants.clif

@@ -50,8 +50,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xffff0000
 ;   ret
@@ -63,8 +63,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xffff00000000
 ;   ret
@@ -76,8 +76,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xffff000000000000
 ;   ret
@@ -109,8 +109,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xffffffff0000ffff
 ;   ret
@@ -122,8 +122,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xffff0000ffffffff
 ;   ret
@@ -135,8 +135,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xffffffffffff
 ;   ret
@@ -148,8 +148,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xf34bf0a31212003a
 ;   ret
@@ -161,8 +161,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0x12e900001ef40000
 ;   ret
@@ -174,8 +174,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0x12e9ffff1ef4ffff
 ;   ret
@@ -197,8 +197,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xfffffff7
 ;   ret
@@ -210,8 +210,8 @@ block0:
 }
 
 ; block0:
-;   auipc a0,0
+;   auipc t1,0
-;   ld a0,12(a0)
+;   ld a0,12(t1)
 ;   j 12
 ;   .8byte 0xfffffff7
 ;   ret
@@ -233,11 +233,11 @@ block0:
 }
 
 ; block0:
-;   auipc t1,0
+;   auipc t2,0
-;   ld t1,12(t1)
+;   ld t2,12(t2)
 ;   j 12
 ;   .8byte 0x3ff0000000000000
-;   fmv.d.x fa0,t1
+;   fmv.d.x fa0,t2
 ;   ret
 
 function %f() -> f32 {
@@ -258,11 +258,11 @@ block0:
 }
 
 ; block0:
-;   auipc t1,0
+;   auipc t2,0
-;   ld t1,12(t1)
+;   ld t2,12(t2)
 ;   j 12
 ;   .8byte 0x4049000000000000
-;   fmv.d.x fa0,t1
+;   fmv.d.x fa0,t2
 ;   ret
 
 function %f() -> f32 {
@@ -305,11 +305,11 @@ block0:
 }
 
 ; block0:
-;   auipc t1,0
+;   auipc t2,0
-;   ld t1,12(t1)
+;   ld t2,12(t2)
 ;   j 12
 ;   .8byte 0xc030000000000000
-;   fmv.d.x fa0,t1
+;   fmv.d.x fa0,t2
 ;   ret
 
 function %f() -> f32 {
@@ -319,10 +319,10 @@ block0:
 }
 
 ; block0:
-;   auipc t1,0
+;   auipc t2,0
-;   lwu t1,12(t1)
+;   lwu t2,12(t2)
 ;   j 8
 ;   .4byte 0xc1800000
-;   fmv.w.x fa0,t1
+;   fmv.w.x fa0,t2
 ;   ret