x64: fix scalar_to_vector with i16x8 types

cranelift/codegen/src/isa/x64/inst.isle

@@ -5058,7 +5058,7 @@
 
 (decl bitcast_gpr_to_xmm (u8 Gpr) Xmm)
 (rule (bitcast_gpr_to_xmm 16 src)
-      (x64_pinsrw (xmm_uninit_value) src 0))
+      (x64_pinsrw (xmm_zero $I16X8) src 0))
 (rule (bitcast_gpr_to_xmm 32 src)
       (x64_movd_to_xmm src))
 (rule (bitcast_gpr_to_xmm 64 src)

cranelift/filetests/filetests/isa/x64/bitcast.clif

@@ -38,6 +38,7 @@ block0(v0: i16):
 ;   movq    %rsp, %rbp
 ; block0:
 ;   uninit  %xmm0
+;   pxor %xmm0, %xmm0
 ;   pinsrw $0x0, %edi, %xmm0
 ;   movq    %rbp, %rsp
 ;   popq    %rbp
@@ -48,6 +49,7 @@ block0(v0: i16):
 ;   pushq %rbp
 ;   movq %rsp, %rbp
 ; block1: ; offset 0x4
+;   pxor %xmm0, %xmm0
 ;   pinsrw $0, %edi, %xmm0
 ;   movq %rbp, %rsp
 ;   popq %rbp
@@ -405,6 +407,7 @@ block0(v0: i16):
 ;   movq    %rsp, %rbp
 ; block0:
 ;   uninit  %xmm0
+;   pxor %xmm0, %xmm0
 ;   pinsrw $0x0, %edi, %xmm0
 ;   movq    %rbp, %rsp
 ;   popq    %rbp
@@ -415,6 +418,7 @@ block0(v0: i16):
 ;   pushq %rbp
 ;   movq %rsp, %rbp
 ; block1: ; offset 0x4
+;   pxor %xmm0, %xmm0
 ;   pinsrw $0, %edi, %xmm0
 ;   movq %rbp, %rsp
 ;   popq %rbp

cranelift/filetests/filetests/isa/x64/f16const.clif

@@ -39,6 +39,7 @@ block0():
 ; block0:
 ;   movl    $15360, %esi
 ;   uninit  %xmm0
+;   pxor %xmm0, %xmm0
 ;   pinsrw $0x0, %esi, %xmm0
 ;   movq    %rbp, %rsp
 ;   popq    %rbp
@@ -50,6 +51,7 @@ block0():
 ;   movq %rsp, %rbp
 ; block1: ; offset 0x4
 ;   movl $0x3c00, %esi
+;   pxor %xmm0, %xmm0
 ;   pinsrw $0, %esi, %xmm0
 ;   movq %rbp, %rsp
 ;   popq %rbp

cranelift/filetests/filetests/isa/x64/issue-10906.clif

@@ -0,0 +1,72 @@
+;; This has an associated runtest; this test explains that one by disassembling
+;; the test case here.
+
+test compile precise-output
+set opt_level=none
+set preserve_frame_pointers=true
+set enable_multi_ret_implicit_sret=true
+target x86_64
+
+function %munge_xmm0() -> i16x8 fast {
+block0:
+    v0 = vconst.i16x8 [0xffff 0xffff 0xffff 0xffff 0xffff 0xffff 0xffff 0xffff]
+    return v0
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   uninit  %xmm0
+;   pcmpeqd %xmm0, %xmm0
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   pcmpeqd %xmm0, %xmm0
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %scalar_to_vector() -> i16x8 fast {
+    fn0 = %munge_xmm0() -> i16x8 fast
+block0:
+    v0 = call fn0()
+    v1 = iconst.i16 42
+    v2 = scalar_to_vector.i16x8 v1
+    return v2
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   load_ext_name %munge_xmm0+0, %r8
+;   call    *%r8
+;   movl    $42, %r8d
+;   uninit  %xmm0
+;   pxor %xmm0, %xmm0
+;   pinsrw $0x0, %r8d, %xmm0
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movabsq $0, %r8 ; reloc_external Abs8 %munge_xmm0 0
+;   callq *%r8
+;   movl $0x2a, %r8d
+;   pxor %xmm0, %xmm0
+;   pinsrw $0, %r8d, %xmm0
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+

cranelift/filetests/filetests/runtests/issue-10906.clif

@@ -0,0 +1,31 @@
+;; When CLIF's `scalar_to_vector` operates on an `i16x8` vector, we need to be
+;; sure to zero out the upper bits before inserting the `i16` into the lowest
+;; lane. This test uses some ABI knowledge (both functions return their result
+;; in `xmm0`) to fill in all of the bits (`%munge_xmm0`) and see if they "show
+;; through" in the actual test, `%scalar_to_vector`. See
+;; https://github.com/bytecodealliance/wasmtime/issues/10906.
+
+test interpret
+test run
+set opt_level=none
+set preserve_frame_pointers=true
+set enable_multi_ret_implicit_sret=true
+target x86_64
+
+function %munge_xmm0() -> i16x8 fast {
+block0:
+    v0 = vconst.i16x8 [0xffff 0xffff 0xffff 0xffff 0xffff 0xffff 0xffff 0xffff]
+    return v0
+}
+
+function %scalar_to_vector() -> i16x8 fast {
+    fn0 = %munge_xmm0() -> i16x8 fast
+block0:
+    v0 = call fn0()
+    v1 = iconst.i16 42
+    v2 = scalar_to_vector.i16x8 v1
+    return v2
+}
+
+; run: %scalar_to_vector() == [42 0 0 0 0 0 0 0]
+