x64: Optimize store-of-extract-lane-0 (#5924)

* x64: Optimize store-of-extract-lane-0

The `movss` and `movsd` instructions can be used to store the 0th lane
of a `t32x4` or a `t64x2` vector into memory, enabling fusing a `store`
and an `extractlane` instruction.

* Fix merge conflict with `main`

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

@@ -2653,6 +2653,24 @@
         (x64_movrm $I64 addr_lo value_lo)
         (x64_movrm $I64 addr_hi value_hi)))))
 
+;; Slightly optimize the extraction of the first lane from a vector which is
+;; stored in memory. In the case the first lane specifically is selected the
+;; standard `movss` and `movsd` instructions can be used as-if we're storing a
+;; f32 or f64 despite the source perhaps being an integer vector since the
+;; result of the instruction is the same.
+(rule 2 (lower (store flags
+                    (has_type (ty_32 _) (extractlane value (u8_from_uimm8 0)))
+                    address
+                    offset))
+      (side_effect
+       (x64_movss_store (to_amode flags address offset) value)))
+(rule 3 (lower (store flags
+                    (has_type (ty_64 _) (extractlane value (u8_from_uimm8 0)))
+                    address
+                    offset))
+      (side_effect
+       (x64_movsd_store (to_amode flags address offset) value)))
+
 ;; Rules for `load*` + ALU op + `store*` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 ;; Add mem, reg

cranelift/filetests/filetests/isa/x64/extractlane.clif

@@ -151,3 +151,107 @@ block0(v0: f64x2):
 ;   popq %rbp
 ;   retq
 
+function %extract_i32x4_lane0_to_memory(i32x4, i64) {
+block0(v0: i32x4, v1: i64):
+  v2 = extractlane v0, 0
+  store v2, v1
+  return
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movss   %xmm0, 0(%rdi)
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movss %xmm0, (%rdi) ; trap: heap_oob
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %extract_f32x4_lane0_to_memory(f32x4, i64) {
+block0(v0: f32x4, v1: i64):
+  v2 = extractlane v0, 0
+  store v2, v1
+  return
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movss   %xmm0, 0(%rdi)
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movss %xmm0, (%rdi) ; trap: heap_oob
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %extract_i64x2_lane0_to_memory(i64x2, i64) {
+block0(v0: i64x2, v1: i64):
+  v2 = extractlane v0, 0
+  store v2, v1
+  return
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movsd   %xmm0, 0(%rdi)
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movsd %xmm0, (%rdi) ; trap: heap_oob
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %extract_f64x2_lane0_to_memory(f64x2, i64) {
+block0(v0: f64x2, v1: i64):
+  v2 = extractlane v0, 0
+  store v2, v1
+  return
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movsd   %xmm0, 0(%rdi)
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movsd %xmm0, (%rdi) ; trap: heap_oob
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+