Convert fma, valltrue & vanytrue to ISLE (AArch64) (#4608)

* Convert `fma`, `valltrue` & `vanytrue` to ISLE (AArch64)

Ported the existing implementations of the following opcodes to ISLE on
AArch64:
- `fma`
  - Introduced missing support for `fma` on vector values, as per the
    docs.
- `valltrue`
- `vanytrue`

Also fixed `fcmp` on scalar values in the interpreter, and enabled
interpreter tests in `simd-fma.clif`.

This introduces the `FMLA` machine instruction.

Copyright (c) 2022 Arm Limited

* Add comments for `Fmla` and `Bsl`

Copyright (c) 2022 Arm Limited

cranelift/filetests/filetests/isa/aarch64/floating-point.clif

@@ -910,3 +910,39 @@ block0(v0: f64x2):
 ; block0:
 ;   frintn v0.2d, v0.2d
 ;   ret
+
+function %f78(f32x4, f32x4, f32x4) -> f32x4 {
+block0(v0: f32x4, v1: f32x4, v2: f32x4):
+  v3 = fma v0, v1, v2
+  return v3
+}
+
+; block0:
+;   mov v17.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.4s, v17.4s, v1.4s
+;   ret
+
+function %f79(f32x2, f32x2, f32x2) -> f32x2 {
+block0(v0: f32x2, v1: f32x2, v2: f32x2):
+  v3 = fma v0, v1, v2
+  return v3
+}
+
+; block0:
+;   mov v17.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.2s, v17.2s, v1.2s
+;   ret
+
+function %f80(f64x2, f64x2, f64x2) -> f64x2 {
+block0(v0: f64x2, v1: f64x2, v2: f64x2):
+  v3 = fma v0, v1, v2
+  return v3
+}
+
+; block0:
+;   mov v17.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.2d, v17.2d, v1.2d
+;   ret

cranelift/filetests/filetests/isa/aarch64/simd-valltrue.clif

@@ -0,0 +1,94 @@
+test compile precise-output
+set unwind_info=false
+target aarch64
+
+function %fn0(b8x8) -> b1 {
+block0(v0: b8x8):
+    v1 = vall_true v0
+    return v1
+}
+
+; block0:
+;   uminv b3, v0.8b
+;   mov x5, v3.d[0]
+;   subs xzr, x5, #0
+;   cset x0, ne
+;   ret
+
+function %fn1(b8x16) -> b1 {
+block0(v0: b8x16):
+    v1 = vall_true v0
+    return v1
+}
+
+; block0:
+;   uminv b3, v0.16b
+;   mov x5, v3.d[0]
+;   subs xzr, x5, #0
+;   cset x0, ne
+;   ret
+
+function %fn2(b16x4) -> b1 {
+block0(v0: b16x4):
+    v1 = vall_true v0
+    return v1
+}
+
+; block0:
+;   uminv h3, v0.4h
+;   mov x5, v3.d[0]
+;   subs xzr, x5, #0
+;   cset x0, ne
+;   ret
+
+function %fn3(b16x8) -> b1 {
+block0(v0: b16x8):
+    v1 = vall_true v0
+    return v1
+}
+
+; block0:
+;   uminv h3, v0.8h
+;   mov x5, v3.d[0]
+;   subs xzr, x5, #0
+;   cset x0, ne
+;   ret
+
+function %fn4(b32x2) -> b1 {
+block0(v0: b32x2):
+    v1 = vall_true v0
+    return v1
+}
+
+; block0:
+;   mov x3, v0.d[0]
+;   subs xzr, xzr, x3, LSR 32
+;   ccmp w3, #0, #nZcv, ne
+;   cset x0, ne
+;   ret
+
+function %fn5(b32x4) -> b1 {
+block0(v0: b32x4):
+    v1 = vall_true v0
+    return v1
+}
+
+; block0:
+;   uminv s3, v0.4s
+;   mov x5, v3.d[0]
+;   subs xzr, x5, #0
+;   cset x0, ne
+;   ret
+
+function %fn6(b64x2) -> b1 {
+block0(v0: b64x2):
+    v1 = vall_true v0
+    return v1
+}
+
+; block0:
+;   cmeq v3.2d, v0.2d, #0
+;   addp v5.2d, v3.2d, v3.2d
+;   fcmp d5, d5
+;   cset x0, eq
+;   ret

cranelift/filetests/filetests/runtests/simd-fma-64bit.clif

@@ -0,0 +1,47 @@
+test interpret
+test run
+target aarch64
+; x86_64 panics: `not implemented: unable to move type: f32x2`
+
+function %fma_f32x2(f32x2, f32x2, f32x2) -> f32x2 {
+block0(v0: f32x2, v1: f32x2, v2: f32x2):
+    v3 = fma v0, v1, v2
+    return v3
+}
+; run: %fma_f32x2([0x9.0 0x83.0], [0x9.0 0x2.68091p6], [0x9.0 0x9.88721p1]) == [0x1.680000p6 0x1.3b88e6p14]
+; run: %fma_f32x2([0x0.0 0x0.0], [0x0.0 0x0.0], [0x0.0 -0x0.0]) == [0x0.0 0x0.0]
+; run: %fma_f32x2([0x0.0 0x0.0], [-0x0.0 -0x0.0], [0x0.0 0x0.0]) == [0x0.0 0x0.0]
+; run: %fma_f32x2([-0x0.0 -0x0.0], [0x0.0 0x0.0], [0x0.0 0x0.0]) == [0x0.0 0x0.0]
+
+; run: %fma_f32x2([-Inf Inf], [-Inf -Inf], [0x0.0 0x0.0]) == [+Inf -Inf]
+; run: %fma_f32x2([-Inf Inf], [Inf -Inf], [0x0.0 -Inf]) == [-Inf -Inf]
+; run: %fma_f32x2([-Inf -Inf], [Inf Inf], [-Inf -Inf]) == [-Inf -Inf]
+; run: %fma_f32x2([-Inf -Inf], [Inf Inf], [-Inf -Inf]) == [-Inf -Inf]
+
+; F32 Epsilon / Max / Min Positive
+; run: %fma_f32x2([0x1.000000p-23 0x0.0], [0x1.000000p-23 0x0.0], [0x1.000000p-23 0x1.000000p-23]) == [0x1.000002p-23 0x1.000000p-23]
+; run: %fma_f32x2([0x1.fffffep127 0x0.0], [0x1.fffffep127 0x0.0], [0x1.fffffep127 0x1.fffffep127]) == [+Inf 0x1.fffffep127]
+; run: %fma_f32x2([0x1.000000p-126 0x1.000000p-126], [0x1.000000p-126 0x1.000000p-126], [0x1.000000p-126 0x1.000000p-126]) == [0x1.000000p-126 0x1.000000p-126]
+; run: %fma_f32x2([0x0.0 0x0.0], [0x0.0 0x0.0], [0x1.000000p-126 0x1.000000p-126]) == [0x1.000000p-126 0x1.000000p-126]
+
+; F32 Subnormals
+; run: %fma_f32x2([0x0.800000p-126 0x0.800000p-126], [0x0.800000p-126 0x0.800000p-126], [0x0.800000p-126 0x0.0]) == [0x0.800000p-126 0x0.0]
+; run: %fma_f32x2([0x0.0 0x0.000002p-126], [0x0.0 0x0.000002p-126], [0x0.800000p-126 0x0.000002p-126]) == [0x0.800000p-126 0x0.000002p-126]
+; run: %fma_f32x2([0x0.000002p-126 0x0.000002p-126], [0x0.000002p-126 0x0.000002p-126], [0x0.0 0x0.0]) == [0x0.0 0x0.0]
+; run: %fma_f32x2([0x0.0 0x0.0], [0x0.0 0x0.0], [0x0.000002p-126 0x0.000002p-126]) == [0x0.000002p-126 0x0.000002p-126]
+
+;; The IEEE754 Standard does not make a lot of guarantees about what
+;; comes out of NaN producing operations, we just check if its a NaN
+function %fma_is_nan_f32x2(f32x2, f32x2, f32x2) -> b1 {
+block0(v0: f32x2, v1: f32x2, v2: f32x2):
+    v3 = fma v0, v1, v2
+    v4 = fcmp ne v3, v3
+    v5 = vall_true v4
+    return v5
+}
+; run: %fma_is_nan_f32x2([Inf -Inf], [-Inf Inf], [Inf Inf]) == true
+; run: %fma_is_nan_f32x2([-Inf +NaN], [-Inf 0x0.0], [-Inf 0x0.0]) == true
+; run: %fma_is_nan_f32x2([0x0.0 0x0.0], [+NaN 0x0.0], [0x0.0 +NaN]) == true
+; run: %fma_is_nan_f32x2([-NaN 0x0.0], [0x0.0 -NaN], [0x0.0 0x0.0]) == true
+; run: %fma_is_nan_f32x2([0x0.0 NaN], [0x0.0 NaN], [-NaN NaN]) == true
+; run: %fma_is_nan_f32x2([NaN NaN], [NaN NaN], [NaN NaN]) == true

cranelift/filetests/filetests/runtests/simd-fma.clif

@@ -1,5 +1,7 @@
+test interpret
 test run
 target x86_64 has_avx has_fma
+target aarch64
 
 function %fma_f32x4(f32x4, f32x4, f32x4) -> f32x4 {
 block0(v0: f32x4, v1: f32x4, v2: f32x4):

cranelift/filetests/filetests/runtests/simd-valltrue-64bit.clif

@@ -0,0 +1,58 @@
+test interpret
+test run
+target aarch64
+; s390x and x86_64 do not support 64-bit vectors.
+
+function %valltrue_b8x8_f() -> b1 {
+block0:
+    v0 = bconst.b8 false
+    v1 = splat.b8x8 v0
+    v2 = vall_true v1
+    return v2
+}
+; run: %valltrue_b8x8_f() == false
+
+function %valltrue_b8x8_t() -> b1 {
+block0:
+    v0 = bconst.b8 true
+    v1 = splat.b8x8 v0
+    v2 = vall_true v1
+    return v2
+}
+; run: %valltrue_b8x8_t() == true
+
+function %valltrue_b16x4_f() -> b1 {
+block0:
+    v0 = bconst.b16 false
+    v1 = splat.b16x4 v0
+    v2 = vall_true v1
+    return v2
+}
+; run: %valltrue_b16x4_f() == false
+
+function %valltrue_b16x4_t() -> b1 {
+block0:
+    v0 = bconst.b16 true
+    v1 = splat.b16x4 v0
+    v2 = vall_true v1
+    return v2
+}
+; run: %valltrue_b16x4_t() == true
+
+function %valltrue_b32x2_f() -> b1 {
+block0:
+    v0 = bconst.b32 false
+    v1 = splat.b32x2 v0
+    v2 = vall_true v1
+    return v2
+}
+; run: %valltrue_b32x2_f() == false
+
+function %valltrue_b32x2_t() -> b1 {
+block0:
+    v0 = bconst.b32 true
+    v1 = splat.b32x2 v0
+    v2 = vall_true v1
+    return v2
+}
+; run: %valltrue_b32x2_t() == true

cranelift/filetests/filetests/runtests/simd-vanytrue-64bit.clif

@@ -0,0 +1,58 @@
+test interpret
+test run
+target aarch64
+; s390x and x86_64 do not support 64-bit vectors.
+
+function %vanytrue_b8x8_f() -> b1 {
+block0:
+    v0 = bconst.b8 false
+    v1 = splat.b8x8 v0
+    v2 = vany_true v1
+    return v2
+}
+; run: %vanytrue_b8x8_f() == false
+
+function %vanytrue_b8x8_t() -> b1 {
+block0:
+    v0 = bconst.b8 true
+    v1 = splat.b8x8 v0
+    v2 = vany_true v1
+    return v2
+}
+; run: %vanytrue_b8x8_t() == true
+
+function %vanytrue_b16x4_f() -> b1 {
+block0:
+    v0 = bconst.b16 false
+    v1 = splat.b16x4 v0
+    v2 = vany_true v1
+    return v2
+}
+; run: %vanytrue_b16x4_f() == false
+
+function %vanytrue_b16x4_t() -> b1 {
+block0:
+    v0 = bconst.b16 true
+    v1 = splat.b16x4 v0
+    v2 = vany_true v1
+    return v2
+}
+; run: %vanytrue_b16x4_t() == true
+
+function %vanytrue_b32x2_f() -> b1 {
+block0:
+    v0 = bconst.b32 false
+    v1 = splat.b32x2 v0
+    v2 = vany_true v1
+    return v2
+}
+; run: %vanytrue_b32x2_f() == false
+
+function %vanytrue_b32x2_t() -> b1 {
+block0:
+    v0 = bconst.b32 true
+    v1 = splat.b32x2 v0
+    v2 = vany_true v1
+    return v2
+}
+; run: %vanytrue_b32x2_t() == true