wazevo(amd64): fixes u32 i->f conv with higher bits garbage (#2136)

Signed-off-by: Takeshi Yoneda <t.y.mathetake@gmail.com>

internal/engine/wazevo/backend/isa/amd64/machine.go

@@ -2866,9 +2866,11 @@ func (m *machine) lowerFcvtFromUint(rn, rd operand, src64, dst64 bool) {
 	// >> which allows CVTSI2SS to be used after all.
 	//
 	if !src64 {
-		cvt := m.allocateInstr()
-		cvt.asGprToXmm(op, rn, rd.reg(), true)
-		m.insert(cvt)
+		// Before we convert, we have to clear the higher 32-bits of the 64-bit register
+		// to get the correct result.
+		tmp := m.c.AllocateVReg(ssa.TypeI32)
+		m.insert(m.allocateInstr().asMovzxRmR(extModeLQ, rn, tmp))
+		m.insert(m.allocateInstr().asGprToXmm(op, newOperandReg(tmp), rd.reg(), true))
 		return
 	}
 

internal/integration_test/fuzzcases/fuzzcases_test.go

@@ -1056,3 +1056,10 @@ func Test2131(t *testing.T) {
 	}
 	nodiff.RequireNoDiffT(t, getWasmBinary(t, "2131"), true, true)
 }
+
+func Test2136(t *testing.T) {
+	if !platform.CompilerSupported() {
+		return
+	}
+	nodiff.RequireNoDiffT(t, getWasmBinary(t, "2136"), true, true)
+}

internal/integration_test/fuzzcases/testdata/2136.wat

@@ -0,0 +1,11 @@
+(module
+  (type (;0;) (func (result f32)))
+  (func (;0;) (type 0) (result f32)
+    i32.const 0
+    i32.load8_s
+    f32.convert_i32_u
+  )
+  (memory (;0;) 3 8)
+  (export "" (func 0))
+  (data (;0;) (i32.const 0) "\f2")
+)