cranelift: Fix fdemote on the interpreter (#5158)

* cranelift: Cleanup `fdemote`/`fpromote` tests * cranelift: Fix `fdemote`/`fpromote` instruction docs The verifier fails if the input and output types are the same for these instructions * cranelift: Fix `fdemote`/`fpromote` in the interpreter * fuzzgen: Add `fdemote`/`fpromote`
2022-11-15 22:22:00 +00:00
parent a007e02bd2
commit a793648eb2
8 changed files with 249 additions and 100 deletions
--- a/cranelift/filetests/filetests/runtests/fdemote.clif
+++ b/cranelift/filetests/filetests/runtests/fdemote.clif
@@ -0,0 +1,91 @@
+test interpret
+test run
+target x86_64
+target s390x
+target aarch64
+target riscv64
+
+
+function %fdemote(f64) -> f32 {
+block0(v0: f64):
+    v1 = fdemote.f32 v0
+    return v1
+}
+; run: %fdemote(0x0.0) == 0x0.0
+; run: %fdemote(-0x0.0) == -0x0.0
+; run: %fdemote(0x0.1) == 0x0.1
+; run: %fdemote(0x0.2) == 0x0.2
+; run: %fdemote(0x0.5) == 0x0.5
+; run: %fdemote(-0x0.5) == -0x0.5
+; run: %fdemote(0x3.2) == 0x3.2
+; run: %fdemote(0x9.0) == 0x9.0
+; run: %fdemote(-0x9.0) == -0x9.0
+; run: %fdemote(0x1.1p10) == 0x1.100000p10
+; run: %fdemote(-0x1.1p10) == -0x1.100000p10
+; run: %fdemote(0x1.c555555555556p10) == 0x1.c55556p10
+; run: %fdemote(-0x1.999999999999ap-2) == -0x1.99999ap-2
+; run: %fdemote(0x1.c3c3c3c3c3c3cp-1) == 0x1.c3c3c4p-1
+; run: %fdemote(0x1.c924924924925p-1) == 0x1.c92492p-1
+; run: %fdemote(0x1.4cccccccccccdp0) == 0x1.4cccccp0
+
+
+;; Inf
+; run: %fdemote(Inf) == Inf
+; run: %fdemote(-Inf) == -Inf
+
+;; Epsilon / Max / Min Positive
+; run: %fdemote(0x1.0000000000000p-52) == 0x1.0000000000000p-52
+; run: %fdemote(-0x1.0000000000000p-52) == -0x1.0000000000000p-52
+; run: %fdemote(0x1.fffffffffffffp1023) == +Inf
+; run: %fdemote(-0x1.fffffffffffffp1023) == -Inf
+; run: %fdemote(0x1.0000000000000p-1022) == 0x0.0
+; run: %fdemote(-0x1.0000000000000p-1022) == -0x0.0
+
+;; Subnormals
+; run: %fdemote(0x0.8000000000000p-1022) == 0x0.0
+; run: %fdemote(-0x0.8000000000000p-1022) == -0x0.0
+; run: %fdemote(0x0.0000000000001p-1022) == 0x0.0
+; run: %fdemote(-0x0.0000000000001p-1022) == -0x0.0
+
+
+;; NaN's
+; For NaN's this operation is specified as producing a value that is a NaN
+function %fdemote_is_nan(f64) -> i8 {
+block0(v0: f64):
+    v1 = fdemote.f32 v0
+    v2 = fcmp ne v1, v1
+    return v2
+}
+; run: %fdemote_is_nan(+NaN) == 1
+; run: %fdemote_is_nan(-NaN) == 1
+; run: %fdemote_is_nan(+NaN:0x0) == 1
+; run: %fdemote_is_nan(+NaN:0x1) == 1
+; run: %fdemote_is_nan(+NaN:0x4000000000001) == 1
+; run: %fdemote_is_nan(-NaN:0x0) == 1
+; run: %fdemote_is_nan(-NaN:0x1) == 1
+; run: %fdemote_is_nan(-NaN:0x4000000000001) == 1
+; run: %fdemote_is_nan(+sNaN:0x1) == 1
+; run: %fdemote_is_nan(-sNaN:0x1) == 1
+; run: %fdemote_is_nan(+sNaN:0x4000000000001) == 1
+; run: %fdemote_is_nan(-sNaN:0x4000000000001) == 1
+
+
+;; Tests a fdemote+load combo which some backends may optimize
+function %fdemote_load(i64 vmctx, i64, f64) -> f32 {
+    gv0 = vmctx
+    gv1 = load.i64 notrap aligned gv0+0
+    heap0 = static gv1, min 0x10, bound 0x10, offset_guard 0x0, index_type i64
+
+block0(v0: i64, v1: i64, v2: f64):
+    v3 = heap_addr.i64 heap0, v1, 0, 8
+    store.f64 v2, v3
+    v4 = load.f64 v3
+    v5 = fdemote.f32 v4
+    return v5
+}
+; heap: static, size=0x10, ptr=vmctx+0, bound=vmctx+8
+; run: %fdemote_load(0, 0x0.0) == 0x0.0
+; run: %fdemote_load(1, 0x0.1) == 0x0.1
+; run: %fdemote_load(2, 0x0.2) == 0x0.2
+; run: %fdemote_load(3, 0x3.2) == 0x3.2
+; run: %fdemote_load(0x8, 0x3.2) == 0x3.2