Added bitrev instruction for 32 and 64 bit integers (#486)

cranelift/filetests/legalizer/bitrev.clif

@@ -0,0 +1,149 @@
+test legalizer
+target x86_64
+
+function %reverse_bits_8(i8) -> i8 {
+ebb0(v0: i8):
+    v1 = bitrev.i8 v0
+    return v1
+}
+; check: v2 = band_imm v0, 170
+; check: v3 = ushr_imm v2, 1
+; check: v4 = band_imm v0, 85
+; check: v5 = ishl_imm v4, 1
+; check: v16 = uextend.i32 v3
+; check: v17 = uextend.i32 v5
+; check: v18 = bor v16, v17
+; check: v6 = ireduce.i8 v18
+; check: v7 = band_imm v6, 204
+; check: v8 = ushr_imm v7, 2
+; check: v9 = band_imm v6, 51
+; check: v10 = ushr_imm v9, 2
+; check: v19 = uextend.i32 v8
+; check: v20 = uextend.i32 v10
+; check: v21 = bor v19, v20
+; check: v11 = ireduce.i8 v21
+; check: v12 = band_imm v11, 240
+; check: v13 = ushr_imm v12, 4
+; check: v14 = band_imm v11, 15
+; check: v15 = ishl_imm v14, 4
+; check: v22 = uextend.i32 v13
+; check: v23 = uextend.i32 v15
+; check: v24 = bor v22, v23
+; check: v1 = ireduce.i8 v24
+
+function %reverse_bits_16(i16) -> i16 {
+ebb0(v0: i16):
+    v1 = bitrev.i16 v0
+    return v1
+}
+; check: v2 = band_imm v0, 0xaaaa
+; check: v3 = ushr_imm v2, 1
+; check: v4 = band_imm v0, 0x5555
+; check: v5 = ishl_imm v4, 1
+; check: v21 = uextend.i32 v3
+; check: v22 = uextend.i32 v5
+; check: v23 = bor v21, v22
+; check: v6 = ireduce.i16 v23
+; check: v7 = band_imm v6, 0xcccc
+; check: v8 = ushr_imm v7, 2
+; check: v9 = band_imm v6, 0x3333
+; check: v10 = ushr_imm v9, 2
+; check: v24 = uextend.i32 v8
+; check: v25 = uextend.i32 v10
+; check: v26 = bor v24, v25
+; check: v11 = ireduce.i16 v26
+; check: v12 = band_imm v11, 0xf0f0
+; check: v13 = ushr_imm v12, 4
+; check: v14 = band_imm v11, 3855
+; check: v15 = ishl_imm v14, 4
+; check: v27 = uextend.i32 v13
+; check: v28 = uextend.i32 v15
+; check: v29 = bor v27, v28
+; check: v16 = ireduce.i16 v29
+; check: v17 = band_imm v16, 0xff00
+; check: v18 = ushr_imm v17, 8
+; check: v19 = band_imm v16, 255
+; check: v20 = ishl_imm v19, 8
+; check: v30 = uextend.i32 v18
+; check: v31 = uextend.i32 v20
+; check: v32 = bor v30, v31
+; check: v1 = ireduce.i16 v32
+; check: return v1
+
+function %reverse_bits_32(i32) -> i32 {
+ebb0(v0: i32):
+    v1 = bitrev.i32 v0
+    return v1
+}
+; check: v24 = iconst.i32 0xaaaa_aaaa
+; check: v2 = band v0, v24
+; check: v3 = ushr_imm v2, 1
+; check: v4 = band_imm v0, 0x5555_5555
+; check: v5 = ishl_imm v4, 1
+; check: v6 = bor v3, v5
+; check: v25 = iconst.i32 0xcccc_cccc
+; check: v7 = band v6, v25
+; check: v8 = ushr_imm v7, 2
+; check: v9 = band_imm v6, 0x3333_3333
+; check: v10 = ushr_imm v9, 2
+; check: v11 = bor v8, v10
+; check: v26 = iconst.i32 0xf0f0_f0f0
+; check: v12 = band v11, v26
+; check: v13 = ushr_imm v12, 4
+; check: v14 = band_imm v11, 0x0f0f_0f0f
+; check: v15 = ishl_imm v14, 4
+; check: v16 = bor v13, v15
+; check: v27 = iconst.i32 0xff00_ff00
+; check: v17 = band v16, v27
+; check: v18 = ushr_imm v17, 8
+; check: v19 = band_imm v16, 0x00ff_00ff
+; check: v20 = ishl_imm v19, 8
+; check: v21 = bor v18, v20
+; check: v22 = ushr_imm v21, 16
+; check: v23 = ishl_imm v21, 16
+; check: v1 = bor v22, v23
+
+
+function %reverse_bits_64(i64) -> i64 {
+ebb0(v0: i64):
+    v1 = bitrev.i64 v0
+    return v1
+}
+; check: v29 = iconst.i64 0xaaaa_aaaa_aaaa_aaaa
+; check: v2 = band v0, v29
+; check: v3 = ushr_imm v2, 1
+; check: v30 = iconst.i64 0x5555_5555_5555_5555
+; check: v4 = band v0, v30
+; check: v5 = ishl_imm v4, 1
+; check: v6 = bor v3, v5
+; check: v31 = iconst.i64 0xcccc_cccc_cccc_cccc
+; check: v7 = band v6, v31
+; check: v8 = ushr_imm v7, 2
+; check: v32 = iconst.i64 0x3333_3333_3333_3333
+; check: v9 = band v6, v32
+; check: v10 = ushr_imm v9, 2
+; check: v11 = bor v8, v10
+; check: v33 = iconst.i64 0xf0f0_f0f0_f0f0_f0f0
+; check: v12 = band v11, v33
+; check: v13 = ushr_imm v12, 4
+; check: v34 = iconst.i64 0x0f0f_0f0f_0f0f_0f0f
+; check: v14 = band v11, v34
+; check: v15 = ishl_imm v14, 4
+; check: v16 = bor v13, v15
+; check: v35 = iconst.i64 0xff00_ff00_ff00_ff00
+; check: v17 = band v16, v35
+; check: v18 = ushr_imm v17, 8
+; check: v36 = iconst.i64 0x00ff_00ff_00ff_00ff
+; check: v19 = band v16, v36
+; check: v20 = ishl_imm v19, 8
+; check: v21 = bor v18, v20
+; check: v37 = iconst.i64 0xffff_0000_ffff_0000
+; check: v22 = band v21, v37
+; check: v23 = ushr_imm v22, 16
+; check: v38 = iconst.i64 0xffff_0000_ffff
+; check: v24 = band v21, v38
+; check: v25 = ishl_imm v24, 16
+; check: v26 = bor v23, v25
+; check: v27 = ushr_imm v26, 32
+; check: v28 = ishl_imm v26, 32
+; check: v1 = bor v27, v28

lib/codegen/meta-python/base/instructions.py

@@ -1409,6 +1409,14 @@ sshr_imm = Instruction(
 x = Operand('x', iB)
 a = Operand('a', iB)
 
+bitrev = Instruction(
+        'bitrev', r"""
+        Reverse the bits of a integer.
+
+        Reverses the bits in ``x``.
+        """,
+        ins=x, outs=a)
+
 clz = Instruction(
         'clz', r"""
         Count leading zero bits.

lib/codegen/meta-python/base/legalize.py

@@ -26,6 +26,7 @@ from .instructions import rotl, rotl_imm, rotr, rotr_imm
 from .instructions import f32const, f64const
 from .instructions import store, load
 from .instructions import br_table
+from .instructions import bitrev
 from cdsl.ast import Var
 from cdsl.xform import Rtl, XFormGroup
 
@@ -91,17 +92,35 @@ y = Var('y')
 a = Var('a')
 a1 = Var('a1')
 a2 = Var('a2')
+a3 = Var('a3')
+a4 = Var('a4')
 b = Var('b')
 b1 = Var('b1')
 b2 = Var('b2')
+b3 = Var('b3')
+b4 = Var('b4')
 b_in = Var('b_in')
 b_int = Var('b_int')
 c = Var('c')
 c1 = Var('c1')
 c2 = Var('c2')
+c3 = Var('c3')
+c4 = Var('c4')
 c_in = Var('c_in')
 c_int = Var('c_int')
 d = Var('d')
+d1 = Var('d1')
+d2 = Var('d2')
+d3 = Var('d3')
+d4 = Var('d4')
+e = Var('e')
+e1 = Var('e1')
+e2 = Var('e2')
+e3 = Var('e3')
+e4 = Var('e4')
+f = Var('f')
+f1 = Var('f1')
+f2 = Var('f2')
 xl = Var('xl')
 xh = Var('xh')
 yl = Var('yl')
@@ -382,6 +401,115 @@ expand.legalize(
             a << bxor(x, y)
         ))
 
+# Expand bitrev
+# Adapted from Stack Overflow.
+# https://stackoverflow.com/questions/746171/most-efficient-algorithm-for-bit-reversal-from-msb-lsb-to-lsb-msb-in-c
+widen.legalize(
+        a << bitrev.i8(x),
+        Rtl(
+            a1 << band_imm(x, imm64(0xaa)),
+            a2 << ushr_imm(a1, imm64(1)),
+            a3 << band_imm(x, imm64(0x55)),
+            a4 << ishl_imm(a3, imm64(1)),
+            b << bor(a2, a4),
+            b1 << band_imm(b, imm64(0xcc)),
+            b2 << ushr_imm(b1, imm64(2)),
+            b3 << band_imm(b, imm64(0x33)),
+            b4 << ushr_imm(b3, imm64(2)),
+            c << bor(b2, b4),
+            c1 << band_imm(c, imm64(0xf0)),
+            c2 << ushr_imm(c1, imm64(4)),
+            c3 << band_imm(c, imm64(0x0f)),
+            c4 << ishl_imm(c3, imm64(4)),
+            a << bor(c2, c4),
+        ))
+
+widen.legalize(
+        a << bitrev.i16(x),
+        Rtl(
+            a1 << band_imm(x, imm64(0xaaaa)),
+            a2 << ushr_imm(a1, imm64(1)),
+            a3 << band_imm(x, imm64(0x5555)),
+            a4 << ishl_imm(a3, imm64(1)),
+            b << bor(a2, a4),
+            b1 << band_imm(b, imm64(0xcccc)),
+            b2 << ushr_imm(b1, imm64(2)),
+            b3 << band_imm(b, imm64(0x3333)),
+            b4 << ushr_imm(b3, imm64(2)),
+            c << bor(b2, b4),
+            c1 << band_imm(c, imm64(0xf0f0)),
+            c2 << ushr_imm(c1, imm64(4)),
+            c3 << band_imm(c, imm64(0x0f0f)),
+            c4 << ishl_imm(c3, imm64(4)),
+            d << bor(c2, c4),
+            d1 << band_imm(d, imm64(0xff00)),
+            d2 << ushr_imm(d1, imm64(8)),
+            d3 << band_imm(d, imm64(0x00ff)),
+            d4 << ishl_imm(d3, imm64(8)),
+            a << bor(d2, d4),
+        ))
+
+expand.legalize(
+        a << bitrev.i32(x),
+        Rtl(
+            a1 << band_imm(x, imm64(0xaaaaaaaa)),
+            a2 << ushr_imm(a1, imm64(1)),
+            a3 << band_imm(x, imm64(0x55555555)),
+            a4 << ishl_imm(a3, imm64(1)),
+            b << bor(a2, a4),
+            b1 << band_imm(b, imm64(0xcccccccc)),
+            b2 << ushr_imm(b1, imm64(2)),
+            b3 << band_imm(b, imm64(0x33333333)),
+            b4 << ushr_imm(b3, imm64(2)),
+            c << bor(b2, b4),
+            c1 << band_imm(c, imm64(0xf0f0f0f0)),
+            c2 << ushr_imm(c1, imm64(4)),
+            c3 << band_imm(c, imm64(0x0f0f0f0f)),
+            c4 << ishl_imm(c3, imm64(4)),
+            d << bor(c2, c4),
+            d1 << band_imm(d, imm64(0xff00ff00)),
+            d2 << ushr_imm(d1, imm64(8)),
+            d3 << band_imm(d, imm64(0x00ff00ff)),
+            d4 << ishl_imm(d3, imm64(8)),
+            e << bor(d2, d4),
+            e1 << ushr_imm(e, imm64(16)),
+            e2 << ishl_imm(e, imm64(16)),
+            a << bor(e1, e2),
+        ))
+
+expand.legalize(
+        a << bitrev.i64(x),
+        Rtl(
+            a1 << band_imm(x, imm64(0xaaaaaaaaaaaaaaaa)),
+            a2 << ushr_imm(a1, imm64(1)),
+            a3 << band_imm(x, imm64(0x5555555555555555)),
+            a4 << ishl_imm(a3, imm64(1)),
+            b << bor(a2, a4),
+            b1 << band_imm(b, imm64(0xcccccccccccccccc)),
+            b2 << ushr_imm(b1, imm64(2)),
+            b3 << band_imm(b, imm64(0x3333333333333333)),
+            b4 << ushr_imm(b3, imm64(2)),
+            c << bor(b2, b4),
+            c1 << band_imm(c, imm64(0xf0f0f0f0f0f0f0f0)),
+            c2 << ushr_imm(c1, imm64(4)),
+            c3 << band_imm(c, imm64(0x0f0f0f0f0f0f0f0f)),
+            c4 << ishl_imm(c3, imm64(4)),
+            d << bor(c2, c4),
+            d1 << band_imm(d, imm64(0xff00ff00ff00ff00)),
+            d2 << ushr_imm(d1, imm64(8)),
+            d3 << band_imm(d, imm64(0x00ff00ff00ff00ff)),
+            d4 << ishl_imm(d3, imm64(8)),
+            e << bor(d2, d4),
+            e1 << band_imm(e, imm64(0xffff0000ffff0000)),
+            e2 << ushr_imm(e1, imm64(16)),
+            e3 << band_imm(e, imm64(0x0000ffff0000ffff)),
+            e4 << ishl_imm(e3, imm64(16)),
+            f << bor(e2, e4),
+            f1 << ushr_imm(f, imm64(32)),
+            f2 << ishl_imm(f, imm64(32)),
+            a << bor(f1, f2),
+        ))
+
 # Floating-point sign manipulations.
 for ty,             minus_zero in [
         (types.f32, f32const(ieee32.bits(0x80000000))),

lib/codegen/meta-python/cdsl/ast.py

@@ -523,10 +523,14 @@ class ConstantInt(Literal):
 
     def __str__(self):
         # type: () -> str
-        """
+        # If the value is in the signed imm64 range, print it as-is.
-        Get the Rust expression form of this constant.
+        if self.value >= -(2**63) and self.value < (2**63):
-        """
             return str(self.value)
+        # Otherwise if the value is in the unsigned imm64 range, print its
+        # bitwise counterpart in the signed imm64 range.
+        if self.value >= (2**63) and self.value < (2**64):
+            return str(self.value - (2**64))
+        assert False, "immediate value not in signed or unsigned imm64 range"
 
 
 class ConstantBits(Literal):